{ "bench": "rta-bench-1", "entries": 67, "generated_utc": "2026-11-14", "models": 6, "periods": { "2020": { "anchors": [ { "avg_exposure": 0.759225, "avg_holding_days": 46.364911, "excess_vs_equal_weight": 0.273739, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.135881, "name": "swing_trend", "period": "2020", "round_trips": 23.0, "sharpe": 2.181326, "total_return": 0.674441, "turnover": 2.217335, "volatility": 0.249226 }, { "avg_exposure": 0.35554, "avg_holding_days": 3.512317, "excess_vs_equal_weight": -0.337886, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.198434, "name": "swing_rsi2", "period": "2020", "round_trips": 257.0, "sharpe": 0.388742, "total_return": 0.062817, "turnover": 26.432631, "volatility": 0.201147 }, { "avg_exposure": 0.817167, "avg_holding_days": 28.865865, "excess_vs_equal_weight": -0.125137, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.134181, "name": "swing_breakout", "period": "2020", "round_trips": 69.0, "sharpe": 1.232287, "total_return": 0.275566, "turnover": 7.093022, "volatility": 0.213156 }, { "avg_exposure": 0.231373, "avg_holding_days": 9.306144, "excess_vs_equal_weight": -0.37836, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.119089, "name": "swing_pullback", "period": "2020", "round_trips": 61.0, "sharpe": 0.247831, "total_return": 0.022342, "turnover": 6.335951, "volatility": 0.091454 }, { "avg_exposure": 0.795237, "avg_holding_days": 9.420352, "excess_vs_equal_weight": -0.166868, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.176497, "name": "swing_news", "period": "2020", "round_trips": 212.0, "sharpe": 1.178298, "total_return": 0.233835, "turnover": 21.637562, "volatility": 0.190898 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.337172, "name": "SPY", "period": "2020", "round_trips": null, "sharpe": 0.63404, "total_return": 0.172352, "turnover": null, "volatility": 0.333872 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.325267, "name": "Equal weight", "period": "2020", "round_trips": null, "sharpe": 1.125525, "total_return": 0.400702, "turnover": null, "volatility": 0.351896 } ], "clean": [], "may_remember": [ { "blind": { "after_cutoff": null, "alpha_annual": 0.164349, "analysis": { "by_sector": [ { "contribution": 0.107035, "group": "Technology", "runs": 1 }, { "contribution": 0.076665, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": 0.028367, "group": "Health Care", "runs": 1 }, { "contribution": 0.007708, "group": "Communication Services", "runs": 1 }, { "contribution": 0.002533, "group": "Real Estate", "runs": 1 }, { "contribution": 0.000832, "group": "Financials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.136908, "group": "other", "runs": 1 }, { "contribution": 0.115904, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.28% on average (58 buys, 55% of them ahead of SPY; t = 0.4, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.51%, so selling gave that up (94 sells; t = -0.9, within noise).", "Decision IC: on the 50 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.001 with the next five days' returns (IC IR -0.00, t = -0.0, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.038 with the next five days' returns (t = 1.2, within noise).", "Factor exposures: market 0.34 (t 6.3), size -0.00 (t -0.0), growth 0.48 (t 4.0), momentum 0.09 (t 1.8); together they explain 67% of the daily variation (R²).", "Of the +23.8% a year it earned over T-bills (arithmetic mean): growth +11.0%, market +7.2%, momentum +2.0%, size -0.0%; the alpha left is +3.6% a year (t = 0.4, within noise).", "Verdict: most of the return is factor exposure (+20.2% a year, growth above all); the alpha left is not significant.", "P&L: +$25,433 in all, +$25,281 from 49 positions and +$152 of interest on idle cash.", "By holding period: 61+ days made the most (+$18,417, 10 positions); 6-10 days the least (-$1,134, 4 positions).", "By sector: Technology made the most (+$10,703, 19 positions); Financials the least (+$83, 2 positions).", "By stock: NVDA made the most (+$5,431, 2 positions); INTU the least (-$802, 2 positions).", "By market regime of the previous close: risk-on 26 days +2.3% (SPY -4.3%); uncertain 142 days +14.4% (SPY +14.2%); risk-off 85 days +7.3% (SPY +7.2%).", "News: 25 positions opened on news made +$11,590, the other 24 +$13,691; 12 positions closed on news made +$3,331; over the five days from the fill, news-triggered buys beat SPY by 0.64% (t 0.7).", "Concentration: the profit is spread out; the best two positions made 32% of the positions' net P&L, the best being NVDA bought 2020-04-15 and still open at the end, +$5,464. Had the best position made nothing, the run would have returned +20.0% instead of +25.4%." ], "runs": 1, "summary": { "alpha_annual": 0.035786, "alpha_t": 0.360177, "best_trade_share": 0.21614, "beta_growth": 0.482102, "beta_market": 0.341267, "beta_momentum": 0.090367, "beta_size": -0.002257, "buy_excess_5d": 0.002756, "buy_excess_5d_t": 0.371266, "capm_alpha_annual": 0.164349, "capm_alpha_t": 1.359572, "concentrated": false, "decision_excess_5d": -0.002085, "decision_excess_5d_t": -0.483459, "decision_hit_rate_5d": 0.460526, "decision_ic": -0.000713, "decision_ic_t": -0.03438, "decisions": 152.0, "exposure_return_annual": 0.201759, "holdings_ic": 0.037936, "holdings_ic_t": 1.150064, "news_decision_share": 0.381579, "news_entry_pnl": 11590.399234, "r2": 0.671693, "return_without_best_trade": 0.199691, "sell_avoided_5d": -0.005072, "sell_avoided_5d_t": -0.850803, "spy_return": 0.172352, "total_return": 0.253922 } }, "avg_exposure": 0.622788, "avg_holding_days": 32.421832, "beta": 0.345774, "excess_vs_equal_weight": -0.14678, "excess_vs_spy": 0.08157, "files": [ "claude-sonnet-4.6-2020-blind-s0.json" ], "max_drawdown": -0.140136, "n": 1, "round_trips": 94.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 1.383689, "sharpe_range": null, "steps": 51, "total_return": 0.253922, "total_return_range": null, "turnover": 4.028402, "usage": { "calls": 51.0, "input_tokens": 637730.0, "list_price_usd": 2.462, "output_tokens": 36585.0 } }, "cutoff": "2026-01", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": "2026", "missing_answers": { "blind": 200, "named": 200, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ 0.34676864380749145, 0.7964736570321175 ], "recall_direction_accuracy": 0.78, "recall_majority_baseline": 0.68, "recall_score": 0.6036494597839136, "recall_verdict": "recalls", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-sonnet-4.6", "open_weights": false, "provider": "Anthropic", "slug": "claude-sonnet-4.6", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.069552, "analysis": { "by_sector": [ { "contribution": 0.115826, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": 0.063931, "group": "Technology", "runs": 1 }, { "contribution": 0.022897, "group": "Health Care", "runs": 1 }, { "contribution": -0.006669, "group": "Communication Services", "runs": 1 }, { "contribution": -0.016948, "group": "Utilities", "runs": 1 }, { "contribution": -0.023898, "group": "Financials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.115169, "group": "news", "runs": 1 }, { "contribution": 0.053242, "group": "other", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.61% on average (79 buys, 56% of them ahead of SPY; t = 0.9, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.24%, so selling gave that up (125 sells; t = -0.6, within noise).", "Decision IC: on the 49 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.003 with the next five days' returns (IC IR -0.02, t = -0.1, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.015 with the next five days' returns (t = 0.5, within noise).", "Factor exposures: market 0.50 (t 13.7), size -0.12 (t -2.6), growth 0.42 (t 3.5), momentum -0.00 (t -0.1); together they explain 81% of the daily variation (R²).", "Of its +17.2% a year over T-bills (arithmetic mean): market +10.5%, growth +9.6%, size -0.6%, momentum -0.1%; the alpha left is -2.3% a year (t = -0.2, within noise).", "Verdict: most of the return is factor exposure (+19.4% a year, market above all); the alpha left is not significant.", "P&L: +$16,985 in all, +$16,841 from 59 positions and +$144 of interest on idle cash.", "By holding period: 61+ days made the most (+$18,630, 11 positions); 6-10 days lost the most (-$2,331, 4 positions).", "By sector: Consumer Discretionary made the most (+$11,583, 12 positions); Financials lost the most (-$2,390, 3 positions).", "By stock: TSLA made the most (+$7,036, 6 positions); GOOGL lost the most (-$1,982, 1 position).", "By market regime of the previous close: risk-on 26 days +0.4% (SPY -4.3%); uncertain 142 days +10.1% (SPY +14.2%); risk-off 85 days +5.9% (SPY +7.2%).", "News: 42 positions opened on news made +$11,517, the other 17 +$5,324; 17 positions closed on news made +$3,551; over the five days from the fill, news-triggered buys beat SPY by 0.55% (t 0.7).", "Concentration: the profit is spread out; the best two positions made 43% of the positions' net P&L, the best being TSLA bought 2020-07-02 and sold 2020-09-04, +$4,066. Had the best position made nothing, the run would have returned +12.9% instead of +17.0%." ], "runs": 1, "summary": { "alpha_annual": -0.022899, "alpha_t": -0.226091, "best_trade_share": 0.241451, "beta_growth": 0.422978, "beta_market": 0.495998, "beta_momentum": -0.003453, "beta_size": -0.117402, "buy_excess_5d": 0.006085, "buy_excess_5d_t": 0.872287, "capm_alpha_annual": 0.069552, "capm_alpha_t": 0.592272, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.000884, "decision_excess_5d_t": 0.230933, "decision_hit_rate_5d": 0.509804, "decision_ic": -0.003261, "decision_ic_t": -0.148222, "decisions": 204.0, "exposure_return_annual": 0.194438, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.014801, "holdings_ic_t": 0.489878, "long_contribution": 0.16841, "news_decision_share": 0.458537, "news_entry_pnl": 11516.856461, "r2": 0.811907, "return_without_best_trade": 0.129191, "sell_avoided_5d": -0.002402, "sell_avoided_5d_t": -0.551319, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.16947 } }, "avg_exposure": 0.594437, "avg_holding_days": 24.727721, "beta": 0.481778, "excess_vs_equal_weight": -0.231232, "excess_vs_spy": -0.002882, "files": [ "claude-haiku-4.5-2020-blind-s0.json" ], "max_drawdown": -0.202916, "n": 1, "round_trips": 126.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.891048, "sharpe_range": null, "steps": 51, "total_return": 0.16947, "total_return_range": null, "turnover": 5.244807, "usage": { "calls": 51.0, "input_tokens": 639555.0, "list_price_usd": 0.8205, "output_tokens": 36184.0 } }, "cutoff": "2025-07", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": "2026", "missing_answers": { "blind": 200, "named": 200, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ 0.025646044594756782, 0.6609236705305581 ], "recall_direction_accuracy": 0.78, "recall_majority_baseline": 0.68, "recall_score": 0.3847779111644658, "recall_verdict": "recalls", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-haiku-4.5", "named": { "after_cutoff": null, "alpha_annual": 0.026049, "analysis": { "by_sector": [ { "contribution": 0.107536, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": 0.070851, "group": "Technology", "runs": 1 }, { "contribution": 0.008316, "group": "Materials", "runs": 1 }, { "contribution": -0.008367, "group": "Financials", "runs": 1 }, { "contribution": -0.0134, "group": "Health Care", "runs": 1 }, { "contribution": -0.044125, "group": "Consumer Staples", "runs": 1 } ], "by_trigger": [ { "contribution": 0.083526, "group": "news", "runs": 1 }, { "contribution": 0.043579, "group": "other", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.85% on average (121 buys, 59% of them ahead of SPY; t = 1.6, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.57%, so selling gave that up (132 sells; t = -1.3, within noise).", "Decision IC: on the 49 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of 0.004 with the next five days' returns (IC IR 0.02, t = 0.2, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.030 with the next five days' returns (t = 1.1, within noise).", "Factor exposures: market 0.55 (t 15.4), size -0.20 (t -3.7), growth 0.28 (t 2.4), momentum 0.02 (t 0.4); together they explain 85% of the daily variation (R²).", "Of its +13.8% a year over T-bills (arithmetic mean): market +11.6%, growth +6.4%, size -1.0%, momentum +0.4%; the alpha left is -3.7% a year (t = -0.4, within noise).", "Verdict: most of the return is factor exposure (+17.4% a year, market above all); the alpha left is not significant.", "P&L: +$12,864 in all, +$12,710 from 69 positions and +$153 of interest on idle cash.", "By holding period: 61+ days made the most (+$18,956, 12 positions); 11-20 days lost the most (-$3,666, 9 positions).", "By sector: Consumer Discretionary made the most (+$10,754, 9 positions); Consumer Staples lost the most (-$4,412, 11 positions).", "By stock: AMZN made the most (+$5,910, 1 position); UNH lost the most (-$3,126, 4 positions).", "By market regime of the previous close: risk-on 26 days +0.1% (SPY -4.3%); uncertain 142 days +8.5% (SPY +14.2%); risk-off 85 days +4.0% (SPY +7.2%).", "News: 44 positions opened on news made +$8,353, the other 25 +$4,358; 18 positions closed on news made +$1,784; over the five days from the fill, news-triggered buys beat SPY by 0.81% (t 1.3).", "Concentration: the gains are spread out (the best two positions made 31% of them) but winners and losers nearly cancel, so the best two made 70% of the net P&L; the best was AMZN bought 2020-03-17 and sold 2020-10-19, +$5,910. Had the best position made nothing, the run would have returned +7.0% instead of +12.9%." ], "runs": 1, "summary": { "alpha_annual": -0.03656, "alpha_t": -0.407357, "best_trade_share": 0.464977, "beta_growth": 0.282042, "beta_market": 0.547242, "beta_momentum": 0.018192, "beta_size": -0.19698, "buy_excess_5d": 0.008527, "buy_excess_5d_t": 1.552594, "capm_alpha_annual": 0.026049, "capm_alpha_t": 0.256222, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.001082, "decision_excess_5d_t": 0.311437, "decision_hit_rate_5d": 0.509881, "decision_ic": 0.003601, "decision_ic_t": 0.156141, "decisions": 253.0, "exposure_return_annual": 0.17412, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.030431, "holdings_ic_t": 1.123381, "long_contribution": 0.127105, "news_decision_share": 0.384314, "news_entry_pnl": 8352.560031, "r2": 0.853604, "return_without_best_trade": 0.069535, "sell_avoided_5d": -0.005742, "sell_avoided_5d_t": -1.330806, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.128266 } }, "avg_exposure": 0.617873, "avg_holding_days": 29.144955, "beta": 0.526773, "excess_vs_equal_weight": -0.272436, "excess_vs_spy": -0.044086, "files": [ "claude-haiku-4.5-2020-named-s0.json" ], "max_drawdown": -0.17863, "n": 1, "round_trips": 133.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.680373, "sharpe_range": null, "steps": 51, "total_return": 0.128266, "total_return_range": null, "turnover": 5.753521, "usage": { "calls": 51.0, "input_tokens": 638506.0, "list_price_usd": 0.8296, "output_tokens": 38217.0 } }, "named_minus_blind": -0.041204, "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-4.5", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.023616, "analysis": { "by_sector": [ { "contribution": 0.071301, "group": "Technology", "runs": 1 }, { "contribution": 0.014128, "group": "Health Care", "runs": 1 }, { "contribution": -0.003868, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.010768, "group": "Financials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.084711, "group": "news", "runs": 1 }, { "contribution": -0.013918, "group": "other", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.95% on average (22 buys, 32% of them ahead of SPY; t = -1.7, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 1.08%, so selling gave that up (22 sells; t = -0.5, within noise).", "Decision IC: on the 25 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.015 with the next five days' returns (IC IR -0.08, t = -0.4, within noise).", "Holdings IC: the weights it held after each of 47 decision days had a mean rank IC of -0.008 with the next five days' returns (t = -0.3, within noise).", "Factor exposures: market 0.23 (t 13.5), size -0.02 (t -0.7), growth 0.19 (t 4.6), momentum 0.03 (t 1.5); together they explain 78% of the daily variation (R²).", "Of its +7.2% a year over T-bills (arithmetic mean): market +4.8%, growth +4.4%, momentum +0.7%, size -0.1%; the alpha left is -2.6% a year (t = -0.6, within noise).", "Verdict: most of the return is factor exposure (+9.8% a year, market above all); the alpha left is not significant.", "P&L: +$7,386 in all, +$7,079 from 17 positions and +$306 of interest on idle cash.", "By holding period: 61+ days made the most (+$9,230, 4 positions); 3-5 days lost the most (-$1,682, 5 positions).", "By sector: Technology made the most (+$7,130, 5 positions); Financials lost the most (-$1,077, 8 positions).", "By stock: CRM made the most (+$3,340, 1 position); NOW lost the most (-$1,721, 2 positions).", "By market regime of the previous close: risk-on 26 days -1.7% (SPY -4.3%); uncertain 142 days +3.1% (SPY +14.2%); risk-off 85 days +5.9% (SPY +7.2%).", "News: 11 positions opened on news made +$8,471, the other 6 -$1,392; 4 positions closed on news made -$424; over the five days from the fill, news-triggered buys trailed SPY by 0.66% (t -1.6).", "Concentration: the profit rests on the best two positions, which made 94% of the positions' net P&L and 57% of all their gains; the best was CRM bought 2020-05-13 and still open at the end, +$3,340. Had the best position made nothing, the run would have returned +4.0% instead of +7.4%." ], "runs": 1, "summary": { "alpha_annual": -0.026025, "alpha_t": -0.553177, "best_trade_share": 0.471841, "beta_growth": 0.193074, "beta_market": 0.226964, "beta_momentum": 0.029207, "beta_size": -0.015284, "buy_excess_5d": -0.009503, "buy_excess_5d_t": -1.721728, "capm_alpha_annual": 0.023616, "capm_alpha_t": 0.420876, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.010142, "decision_excess_5d_t": -0.918319, "decision_hit_rate_5d": 0.477273, "decision_ic": -0.015112, "decision_ic_t": -0.406581, "decisions": 44.0, "exposure_return_annual": 0.097642, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.007958, "holdings_ic_t": -0.298872, "long_contribution": 0.070793, "news_decision_share": 0.545455, "news_entry_pnl": 8471.128248, "r2": 0.780507, "return_without_best_trade": 0.040453, "sell_avoided_5d": -0.010781, "sell_avoided_5d_t": -0.479034, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.073504 } }, "avg_exposure": 0.219766, "avg_holding_days": 26.823558, "beta": 0.226747, "excess_vs_equal_weight": -0.327198, "excess_vs_spy": -0.098848, "files": [ "claude-haiku-5.5-2020-blind-s0.json" ], "max_drawdown": -0.089644, "n": 1, "round_trips": 22.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.770211, "sharpe_range": null, "steps": 51, "total_return": 0.073504, "total_return_range": null, "turnover": 1.31116, "usage": { "calls": 51.0, "input_tokens": 825287.0, "list_price_usd": 0.088, "output_tokens": 11000.0 } }, "cutoff": "2026-06", "memory": null, "model": "claude-haiku-5.5", "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-5.5", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.073697, "analysis": { "by_sector": [ { "contribution": 0.151158, "group": "Technology", "runs": 2 }, { "contribution": 0.056215, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": 0.043518, "group": "Health Care", "runs": 2 }, { "contribution": 0.009096, "group": "Industrials", "runs": 3 }, { "contribution": -0.00504, "group": "Financials", "runs": 1 }, { "contribution": -0.006613, "group": "Communication Services", "runs": 2 }, { "contribution": -0.021908, "group": "Consumer Staples", "runs": 3 }, { "contribution": -0.022053, "group": "Utilities", "runs": 2 } ], "by_trigger": [ { "contribution": 0.175253, "group": "news", "runs": 3 }, { "contribution": -0.011955, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.084503, "alpha_t": -0.684, "best_trade_share": 0.316485, "beta_growth": 0.581485, "beta_market": 0.478371, "beta_momentum": 0.12753, "beta_size": -0.028284, "buy_excess_5d": 0.012146, "buy_excess_5d_t": 2.723813, "capm_alpha_annual": 0.073697, "capm_alpha_t": 0.489754, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.003361, "decision_excess_5d_t": 0.916351, "decision_hit_rate_5d": 0.524904, "decision_ic": 0.0224, "decision_ic_t": 1.01627, "decisions": 118.666667, "exposure_return_annual": 0.260333, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.035083, "holdings_ic_t": 1.113882, "long_contribution": 0.163298, "news_decision_share": 0.480844, "news_entry_pnl": 17525.346174, "r2": 0.646213, "return_without_best_trade": 0.113914, "sell_avoided_5d": -0.006319, "sell_avoided_5d_t": -1.109452, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.163562 } }, "avg_exposure": 0.806221, "avg_holding_days": 27.985675, "beta": 0.482472, "excess_vs_equal_weight": -0.23714, "excess_vs_spy": -0.00879, "files": [ "gemini-3.5-flash-2020-blind-s0.json", "gemini-3.5-flash-2020-blind-s1.json", "gemini-3.5-flash-2020-blind-s2.json" ], "max_drawdown": -0.196101, "n": 3, "round_trips": 57.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.749567, "sharpe_range": [ 0.491401, 0.953794 ], "steps": 51, "total_return": 0.163562, "total_return_range": [ 0.101138, 0.208257 ], "turnover": 5.392319, "usage": { "calls": 153.0, "input_tokens": 2126873.0, "list_price_usd": 3.4937, "output_tokens": 33712.0 } }, "cutoff": "2025-01", "memory": { "beyond_control": 0.06408548996521679, "beyond_control_ci95": [ -0.07959760831144111, 0.21650274511978695 ], "ci95": [ -0.08596144262714035, 0.11176663521831402 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": 0.014205682272909081, "verdict": "inconclusive", "why": "too few pairs to tell: AUC gap +0.014 before the cutoff and -0.050 after it (the control), difference +0.064, 95% CI [-0.080, +0.217]" }, "model": "gemini-3.5-flash", "named": { "after_cutoff": null, "alpha_annual": 0.02721, "analysis": { "by_sector": [ { "contribution": 0.060016, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": 0.05229, "group": "Technology", "runs": 3 }, { "contribution": 0.023783, "group": "Health Care", "runs": 3 }, { "contribution": 0.002638, "group": "Consumer Staples", "runs": 3 }, { "contribution": -0.010518, "group": "Financials", "runs": 3 }, { "contribution": -0.018082, "group": "Utilities", "runs": 3 } ], "by_trigger": [ { "contribution": 0.077389, "group": "news", "runs": 3 }, { "contribution": 0.044484, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.12104, "alpha_t": -1.084319, "best_trade_share": 0.393431, "beta_growth": 0.597183, "beta_market": 0.54414, "beta_momentum": 0.075966, "beta_size": -0.130872, "buy_excess_5d": 0.007475, "buy_excess_5d_t": 1.544435, "capm_alpha_annual": 0.02721, "capm_alpha_t": 0.188169, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 8e-06, "decision_excess_5d_t": -0.020034, "decision_hit_rate_5d": 0.48905, "decision_ic": 0.000415, "decision_ic_t": 0.021904, "decisions": 89.666667, "exposure_return_annual": 0.261213, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.010944, "holdings_ic_t": 0.369309, "long_contribution": 0.121873, "news_decision_share": 0.380305, "news_entry_pnl": 7738.909435, "r2": 0.741752, "return_without_best_trade": 0.07802, "sell_avoided_5d": -0.009281, "sell_avoided_5d_t": -1.600521, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.121959 } }, "avg_exposure": 0.868446, "avg_holding_days": 26.933413, "beta": 0.53363, "excess_vs_equal_weight": -0.278743, "excess_vs_spy": -0.050393, "files": [ "gemini-3.5-flash-2020-named-s0.json", "gemini-3.5-flash-2020-named-s1.json", "gemini-3.5-flash-2020-named-s2.json" ], "max_drawdown": -0.198985, "n": 3, "round_trips": 40.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.581903, "sharpe_range": [ 0.468495, 0.724677 ], "steps": 51, "total_return": 0.121959, "total_return_range": [ 0.092899, 0.166826 ], "turnover": 4.035691, "usage": { "calls": 153.0, "input_tokens": 2097441.0, "list_price_usd": 3.3883, "output_tokens": 26898.0 } }, "named_minus_blind": -0.041603, "open_weights": false, "provider": "Google", "slug": "gemini-3.5-flash", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.065425, "analysis": { "by_sector": [ { "contribution": 0.065453, "group": "Financials", "runs": 3 }, { "contribution": 0.05703, "group": "Technology", "runs": 2 }, { "contribution": 0.008002, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.000878, "group": "Health Care", "runs": 1 }, { "contribution": -0.002931, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.005438, "group": "Industrials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.132765, "group": "no reason", "runs": 2 }, { "contribution": 0.021822, "group": "other", "runs": 2 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.076872, "alpha_t": 0.394252, "best_trade_share": 1.054094, "beta_growth": 0.012151, "beta_market": 0.187723, "beta_momentum": -0.071952, "beta_size": 0.025222, "buy_excess_5d": 0.013261, "buy_excess_5d_t": 0.726623, "capm_alpha_annual": 0.065425, "capm_alpha_t": 0.45293, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.005211, "decision_excess_5d_t": -1.110544, "decision_hit_rate_5d": 0.633333, "decision_ic": 0.056333, "decision_ic_t": -0.426407, "decisions": 8.666667, "exposure_return_annual": 0.027733, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.006504, "holdings_ic_t": 0.352465, "long_contribution": 0.103058, "news_decision_share": 0.016667, "news_entry_pnl": 0.0, "r2": 0.552732, "return_without_best_trade": 0.004055, "sell_avoided_5d": -0.028301, "sell_avoided_5d_t": -1.216004, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.106041 } }, "avg_exposure": 0.255436, "avg_holding_days": 78.072866, "beta": 0.185078, "excess_vs_equal_weight": -0.294661, "excess_vs_spy": -0.066311, "files": [ "llama3.2-2020-blind-s0.json", "llama3.2-2020-blind-s1.json", "llama3.2-2020-blind-s2.json" ], "max_drawdown": -0.089168, "n": 3, "round_trips": 3.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.726222, "sharpe_range": [ 0.462698, 0.979459 ], "steps": 51, "total_return": 0.106041, "total_return_range": [ 0.023473, 0.214319 ], "turnover": 0.496121, "usage": { "calls": 153.0, "input_tokens": 1654346.0, "list_price_usd": 0.0, "output_tokens": 15386.0 } }, "cutoff": "2023-12", "memory": { "beyond_control": 0.14511742196878757, "beyond_control_ci95": [ -0.06396049829004259, 0.3427582462219258 ], "ci95": [ -0.057204526154891286, 0.2559020000276531 ], "control_period": "2026", "missing_answers": { "blind": 0, "named": 0, "recall": 0 }, "p_value": 0.11794102948525736, "paired": "run", "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "recall_ci95": [ -0.3678638150533243, 0.25645363429085083 ], "recall_direction_accuracy": 0.62, "recall_majority_baseline": 0.68, "recall_score": -0.05018007202881152, "recall_verdict": "no sign", "recall_windows": 50, "score": 0.09563825530212083, "scored_points": "before", "verdict": "inconclusive", "why": "too few pairs to tell: AUC gap +0.096 before the cutoff and -0.049 after it (the control), difference +0.145, 95% CI [-0.064, +0.343]" }, "model": "llama3.2", "named": { "after_cutoff": null, "alpha_annual": 0.096208, "analysis": { "by_sector": [ { "contribution": 0.150649, "group": "Technology", "runs": 3 } ], "by_trigger": [ { "contribution": 0.150649, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.031968, "alpha_t": 0.641915, "best_trade_share": 1.0, "beta_growth": 0.256481, "beta_market": 0.228548, "beta_momentum": 0.031573, "beta_size": -0.027799, "buy_excess_5d": 0.033874, "buy_excess_5d_t": 4.290835, "capm_alpha_annual": 0.096208, "capm_alpha_t": 1.49704, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.032265, "decision_excess_5d_t": 3.902066, "decision_hit_rate_5d": 1.0, "decision_ic": 0.155029, "decision_ic_t": 7.711397, "decisions": 3.0, "exposure_return_annual": 0.112281, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.023526, "holdings_ic_t": 1.275654, "long_contribution": 0.150649, "news_decision_share": 0.333333, "news_entry_pnl": 0.0, "r2": 0.743435, "return_without_best_trade": 0.003263, "sell_avoided_5d": 0.028793, "sell_avoided_5d_t": null, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.153534 } }, "avg_exposure": 0.225138, "avg_holding_days": 182.5, "beta": 0.226945, "excess_vs_equal_weight": -0.247168, "excess_vs_spy": -0.018818, "files": [ "llama3.2-2020-named-s0.json", "llama3.2-2020-named-s1.json", "llama3.2-2020-named-s2.json" ], "max_drawdown": -0.064983, "n": 3, "round_trips": 0.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 1.436702, "sharpe_range": [ 1.408132, 1.459285 ], "steps": 51, "total_return": 0.153534, "total_return_range": [ 0.148241, 0.162334 ], "turnover": 0.121686, "usage": { "calls": 153.0, "input_tokens": 1643693.0, "list_price_usd": 0.0, "output_tokens": 9109.0 } }, "named_minus_blind": 0.047493, "open_weights": true, "provider": "Meta", "slug": "llama3.2", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": -0.063129, "analysis": { "by_sector": [ { "contribution": 0.039477, "group": "Technology", "runs": 2 }, { "contribution": 0.012813, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": 0.001582, "group": "Communication Services", "runs": 1 }, { "contribution": -0.00107, "group": "Consumer Staples", "runs": 3 }, { "contribution": -0.005379, "group": "Financials", "runs": 2 }, { "contribution": -0.006081, "group": "Industrials", "runs": 3 }, { "contribution": -0.012886, "group": "Utilities", "runs": 1 }, { "contribution": -0.016085, "group": "Materials", "runs": 1 }, { "contribution": -0.019604, "group": "Health Care", "runs": 2 } ], "by_trigger": [ { "contribution": 0.021229, "group": "news", "runs": 3 }, { "contribution": -0.026947, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.100699, "alpha_t": -1.016226, "best_trade_share": 0.316723, "beta_growth": 0.184646, "beta_market": 0.309182, "beta_momentum": -0.008142, "beta_size": -0.102294, "buy_excess_5d": 4.7e-05, "buy_excess_5d_t": 0.061623, "capm_alpha_annual": -0.063129, "capm_alpha_t": -0.650148, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.002236, "decision_excess_5d_t": -0.388334, "decision_hit_rate_5d": 0.513643, "decision_ic": 0.00531, "decision_ic_t": 0.168763, "decisions": 121.666667, "exposure_return_annual": 0.100468, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.008728, "holdings_ic_t": -0.396199, "long_contribution": -0.005719, "news_decision_share": 0.182389, "news_entry_pnl": 2122.850872, "r2": 0.587869, "return_without_best_trade": -0.025079, "sell_avoided_5d": -0.003434, "sell_avoided_5d_t": -0.499337, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": -0.004827 } }, "avg_exposure": 0.398104, "avg_holding_days": 28.145642, "beta": 0.297123, "excess_vs_equal_weight": -0.405529, "excess_vs_spy": -0.177179, "files": [ "gpt-5.6-luna-2020-blind-s0.json", "gpt-5.6-luna-2020-blind-s1.json", "gpt-5.6-luna-2020-blind-s2.json" ], "max_drawdown": -0.153662, "n": 3, "round_trips": 82.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.090087, "sharpe_range": [ -0.361806, 0.597034 ], "steps": 51, "total_return": -0.004827, "total_return_range": [ -0.076126, 0.061234 ], "turnover": 3.329345, "usage": { "calls": 153.0, "input_tokens": 1694465.0, "list_price_usd": 0.3649, "output_tokens": 21751.0 } }, "cutoff": "2026-02-16", "memory": { "beyond_control": -0.06380984393757505, "beyond_control_ci95": [ -0.20238300571495493, 0.06719636480966 ], "ci95": [ -0.1189769059065934, 0.06516242205168742 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": -0.024609843937575038, "verdict": "inconclusive", "why": "too few pairs to tell: AUC gap -0.025 before the cutoff and +0.039 after it (the control), difference -0.064, 95% CI [-0.202, +0.067]" }, "model": "gpt-5.6-luna", "named": { "after_cutoff": null, "alpha_annual": -0.008384, "analysis": { "by_sector": [ { "contribution": 0.052981, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": 0.03551, "group": "Technology", "runs": 3 }, { "contribution": 0.02398, "group": "Consumer Staples", "runs": 1 }, { "contribution": 0.021133, "group": "Utilities", "runs": 2 }, { "contribution": 0.021027, "group": "Health Care", "runs": 3 }, { "contribution": -0.005202, "group": "Financials", "runs": 3 }, { "contribution": -0.045631, "group": "Industrials", "runs": 3 } ], "by_trigger": [ { "contribution": 0.117201, "group": "other", "runs": 3 }, { "contribution": 0.001348, "group": "news", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.051255, "alpha_t": -0.652499, "best_trade_share": 0.578391, "beta_growth": 0.201291, "beta_market": 0.712402, "beta_momentum": 0.008511, "beta_size": -0.172554, "buy_excess_5d": 0.005756, "buy_excess_5d_t": 1.705652, "capm_alpha_annual": -0.008384, "capm_alpha_t": -0.056022, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.000432, "decision_excess_5d_t": 0.119296, "decision_hit_rate_5d": 0.589447, "decision_ic": 0.007854, "decision_ic_t": 0.146981, "decisions": 56.666667, "exposure_return_annual": 0.189805, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.0054, "holdings_ic_t": 0.246991, "long_contribution": 0.118549, "news_decision_share": 0.094247, "news_entry_pnl": 134.83192, "r2": 0.927529, "return_without_best_trade": 0.066113, "sell_avoided_5d": -0.003477, "sell_avoided_5d_t": -0.285715, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.172352, "total_return": 0.118726 } }, "avg_exposure": 0.65876, "avg_holding_days": 87.64972, "beta": 0.694105, "excess_vs_equal_weight": -0.281976, "excess_vs_spy": -0.053626, "files": [ "gpt-5.6-luna-2020-named-s0.json", "gpt-5.6-luna-2020-named-s1.json", "gpt-5.6-luna-2020-named-s2.json" ], "max_drawdown": -0.245474, "n": 3, "round_trips": 33.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.577819, "sharpe_range": [ 0.322094, 0.720604 ], "steps": 51, "total_return": 0.118726, "total_return_range": [ 0.0554, 0.155297 ], "turnover": 1.008265, "usage": { "calls": 153.0, "input_tokens": 1691347.0, "list_price_usd": 0.3533, "output_tokens": 12521.0 } }, "named_minus_blind": 0.123553, "open_weights": false, "provider": "OpenAI", "slug": "gpt-5.6-luna", "status": "may_remember" } ], "title": "2020: the COVID crash and rebound", "window": { "end": "2020-12-31", "start": "2020-01-02" } }, "2022": { "anchors": [ { "avg_exposure": 0.729439, "avg_holding_days": 24.891321, "excess_vs_equal_weight": 0.064376, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.177636, "name": "swing_trend", "period": "2022", "round_trips": 56.0, "sharpe": -0.474017, "total_return": -0.059125, "turnover": 5.852223, "volatility": 0.139019 }, { "avg_exposure": 0.320629, "avg_holding_days": 4.226861, "excess_vs_equal_weight": 0.037708, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.165751, "name": "swing_rsi2", "period": "2022", "round_trips": 192.0, "sharpe": -0.975663, "total_return": -0.085793, "turnover": 19.47609, "volatility": 0.101873 }, { "avg_exposure": 0.687894, "avg_holding_days": 19.193698, "excess_vs_equal_weight": -0.02719, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.22038, "name": "swing_breakout", "period": "2022", "round_trips": 86.0, "sharpe": -1.351775, "total_return": -0.150691, "turnover": 8.600809, "volatility": 0.126306 }, { "avg_exposure": 0.188515, "avg_holding_days": 9.229792, "excess_vs_equal_weight": 0.081268, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.062646, "name": "swing_pullback", "period": "2022", "round_trips": 47.0, "sharpe": -1.034249, "total_return": -0.042232, "turnover": 4.921742, "volatility": 0.054324 }, { "avg_exposure": 0.560332, "avg_holding_days": 9.155957, "excess_vs_equal_weight": -0.039013, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.209793, "name": "swing_news", "period": "2022", "round_trips": 152.0, "sharpe": -1.370111, "total_return": -0.162513, "turnover": 15.572197, "volatility": 0.134188 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.244964, "name": "SPY", "period": "2022", "round_trips": null, "sharpe": -0.79525, "total_return": -0.186464, "turnover": null, "volatility": 0.242342 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.230743, "name": "Equal weight", "period": "2022", "round_trips": null, "sharpe": -0.528791, "total_return": -0.123501, "turnover": null, "volatility": 0.228571 } ], "clean": [], "may_remember": [ { "blind": { "after_cutoff": null, "alpha_annual": 0.062863, "analysis": { "by_sector": [ { "contribution": 0.075234, "group": "Energy", "runs": 1 }, { "contribution": 0.047731, "group": "Health Care", "runs": 1 }, { "contribution": 0.000445, "group": "Real Estate", "runs": 1 }, { "contribution": -0.019578, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.031621, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.03335, "group": "Technology", "runs": 1 } ], "by_trigger": [ { "contribution": 0.042026, "group": "other", "runs": 1 }, { "contribution": -0.043375, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.33% on average (66 buys, 50% of them ahead of SPY; t = 0.8, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.38%, so selling gave that up (90 sells; t = -1.1, within noise).", "Decision IC: on the 49 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.013 with the next five days' returns (IC IR -0.10, t = -0.7, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.049 with the next five days' returns (t = 1.5, within noise).", "Factor exposures: market 0.55 (t 17.1), size -0.06 (t -1.1), growth -0.56 (t -6.7), momentum 0.09 (t 2.7); together they explain 68% of the daily variation (R²).", "Of the -0.3% a year it earned over T-bills (arithmetic mean): market -10.5%, growth +9.9%, momentum +0.6%, size +0.1%; the alpha left is -0.5% a year (t = -0.1, within noise).", "Verdict: the exposures account for +0.2% a year of it, and the -0.5% a year left over is not significant: one period of daily returns cannot tell it from luck.", "P&L: +$355 in all, -$135 from 55 positions and +$490 of interest on idle cash.", "By holding period: 61+ days made the most (+$11,789, 10 positions); 6-10 days the least (-$5,822, 11 positions).", "By sector: Energy made the most (+$7,523, 2 positions); Technology the least (-$3,335, 5 positions).", "By stock: XOM made the most (+$4,024, 1 position); DE the least (-$2,033, 3 positions).", "By market regime of the previous close: risk-on 11 days -2.3% (SPY -5.1%); neutral 6 days -0.7% (SPY -3.3%); uncertain 162 days -7.2% (SPY -19.0%); risk-off 72 days +11.4% (SPY +9.5%).", "News: 15 positions opened on news made -$4,337, the other 40 +$4,203; 8 positions closed on news made -$2,157; over the five days from the fill, news-triggered buys trailed SPY by 0.04% (t -0.1).", "Concentration: the positions lost -$135 in all; the best was XOM bought 2022-01-11 and still open at the end, +$4,024. Had the best position made nothing, the run would have returned -3.7% instead of +0.4%." ], "runs": 1, "summary": { "alpha_annual": -0.004983, "alpha_t": -0.081379, "best_trade_share": null, "beta_growth": -0.56319, "beta_market": 0.545618, "beta_momentum": 0.085772, "beta_size": -0.055884, "buy_excess_5d": 0.003313, "buy_excess_5d_t": 0.790061, "capm_alpha_annual": 0.062863, "capm_alpha_t": 0.760007, "concentrated": false, "decision_excess_5d": -0.000782, "decision_excess_5d_t": -0.272504, "decision_hit_rate_5d": 0.461538, "decision_ic": -0.012965, "decision_ic_t": -0.68472, "decisions": 156.0, "exposure_return_annual": 0.001802, "holdings_ic": 0.048817, "holdings_ic_t": 1.495186, "news_decision_share": 0.211538, "news_entry_pnl": -4337.450349, "r2": 0.678519, "return_without_best_trade": -0.036688, "sell_avoided_5d": -0.003785, "sell_avoided_5d_t": -1.136415, "spy_return": -0.186464, "total_return": 0.003664 } }, "avg_exposure": 0.697733, "avg_holding_days": 35.912692, "beta": 0.342692, "excess_vs_equal_weight": 0.127165, "excess_vs_spy": 0.190128, "files": [ "claude-sonnet-4.6-2022-blind-s0.json" ], "max_drawdown": -0.13514, "n": 1, "round_trips": 90.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": -0.026332, "sharpe_range": null, "steps": 50, "total_return": 0.003664, "total_return_range": null, "turnover": 4.897863, "usage": { "calls": 50.0, "input_tokens": 676717.0, "list_price_usd": 2.5493, "output_tokens": 34613.0 } }, "cutoff": "2026-01", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": "2026", "missing_answers": { "blind": 200, "named": 200, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ 0.25737000177481123, 0.7248425981412457 ], "recall_direction_accuracy": 0.66, "recall_majority_baseline": 0.54, "recall_score": 0.5227851140456182, "recall_verdict": "recalls", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-sonnet-4.6", "open_weights": false, "provider": "Anthropic", "slug": "claude-sonnet-4.6", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.000906, "analysis": { "by_sector": [ { "contribution": 0.002022, "group": "Health Care", "runs": 1 }, { "contribution": -0.001418, "group": "Industrials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.004746, "group": "other", "runs": 1 }, { "contribution": -0.004142, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.27% on average (3 buys, 67% of them ahead of SPY; t = 0.4, within noise).", "Sells: over the five trading days from the fill, the stocks it sold lagged SPY by 1.82%, so selling avoided that (1 sells; t = n/a, ).", "Decision IC: on the 3 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of 0.052 with the next five days' returns (IC IR 0.32, t = 0.5, within noise).", "Holdings IC: the weights it held after each of 5 decision days had a mean rank IC of 0.018 with the next five days' returns (t = 0.2, within noise).", "Factor exposures: market 0.01 (t 1.9), size 0.00 (t 1.0), growth -0.01 (t -1.6), momentum 0.00 (t 1.3); together they explain 3% of the daily variation (R²).", "Of its -0.0% a year over T-bills (arithmetic mean): market -0.2%, growth +0.2%, momentum +0.0%, size -0.0%; the alpha left is -0.0% a year (t = -0.0, within noise).", "Verdict: most of the loss is factor exposure (-0.0% a year, market above all); the alpha left is not significant.", "P&L: +$1,365 in all, +$60 from 3 positions and +$1,304 of interest on idle cash.", "By holding period: 21-60 days made the most (+$333, 2 positions); 11-20 days lost the most (-$272, 1 position).", "By sector: Health Care made the most (+$202, 2 positions); Industrials lost the most (-$142, 1 position).", "By stock: MRK made the most (+$475, 1 position); ABBV lost the most (-$272, 1 position).", "By market regime of the previous close: risk-on 11 days -0.0% (SPY -5.1%); neutral 6 days -0.1% (SPY -3.3%); uncertain 162 days +1.2% (SPY -19.0%); risk-off 72 days +0.3% (SPY +9.5%).", "News: 2 positions opened on news made -$414, the other 1 +$475; over the five days from the fill, news-triggered buys trailed SPY by 0.42% (t -1.0).", "Concentration: the profit rests on the best two positions, which made 786% of the positions' net P&L and 100% of all their gains; the best was MRK bought 2022-11-25 and still open at the end, +$475. Had the best position made nothing, the run would have returned +0.9% instead of +1.4%." ], "runs": 1, "summary": { "alpha_annual": -0.000123, "alpha_t": -0.017709, "best_trade_share": 7.85576, "beta_growth": -0.009202, "beta_market": 0.010239, "beta_momentum": 0.003186, "beta_size": 0.003684, "buy_excess_5d": 0.002713, "buy_excess_5d_t": 0.418988, "capm_alpha_annual": 0.000906, "capm_alpha_t": 0.129895, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.006589, "decision_excess_5d_t": 1.178733, "decision_hit_rate_5d": 0.75, "decision_ic": 0.051621, "decision_ic_t": 0.546589, "decisions": 4.0, "exposure_return_annual": -0.000218, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.017575, "holdings_ic_t": 0.170136, "long_contribution": 0.000604, "news_decision_share": 0.5, "news_entry_pnl": -414.225439, "r2": 0.033313, "return_without_best_trade": 0.008901, "sell_avoided_5d": 0.018217, "sell_avoided_5d_t": null, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": 0.013758 } }, "avg_exposure": 0.024867, "avg_holding_days": 10.0, "beta": 0.00647, "excess_vs_equal_weight": 0.137259, "excess_vs_spy": 0.200222, "files": [ "claude-haiku-5.5-2022-blind-s0.json" ], "max_drawdown": -0.006987, "n": 1, "round_trips": 1.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": -0.03157, "sharpe_range": null, "steps": 50, "total_return": 0.013758, "total_return_range": null, "turnover": 0.177028, "usage": { "calls": 50.0, "input_tokens": 874228.0, "list_price_usd": 0.0908, "output_tokens": 6759.0 } }, "cutoff": "2026-06", "memory": null, "model": "claude-haiku-5.5", "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-5.5", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.02924, "analysis": { "by_sector": [ { "contribution": 0.063169, "group": "Energy", "runs": 1 }, { "contribution": 0.030206, "group": "Health Care", "runs": 1 }, { "contribution": 0.020041, "group": "Communication Services", "runs": 1 }, { "contribution": -0.028534, "group": "Financials", "runs": 1 }, { "contribution": -0.047184, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.069594, "group": "Technology", "runs": 1 } ], "by_trigger": [ { "contribution": -0.009984, "group": "other", "runs": 1 }, { "contribution": -0.046672, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.15% on average (108 buys, 56% of them ahead of SPY; t = 0.4, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.14%, so selling gave that up (132 sells; t = -0.5, within noise).", "Decision IC: on the 47 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.006 with the next five days' returns (IC IR -0.04, t = -0.3, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.032 with the next five days' returns (t = 1.2, within noise).", "Factor exposures: market 0.58 (t 18.4), size -0.05 (t -0.9), growth -0.33 (t -3.2), momentum 0.05 (t 1.1); together they explain 74% of the daily variation (R²).", "Of its -6.0% a year over T-bills (arithmetic mean): market -11.3%, growth +5.8%, momentum +0.4%, size +0.1%; the alpha left is -1.0% a year (t = -0.2, within noise).", "Verdict: most of the loss is factor exposure (-5.0% a year, market above all); the alpha left is not significant.", "P&L: -$5,319 in all, -$5,666 from 70 positions and +$346 of interest on idle cash.", "By holding period: 21-60 days made the most (+$6,452, 18 positions); 3-5 days lost the most (-$7,413, 19 positions).", "By sector: Energy made the most (+$6,317, 9 positions); Technology lost the most (-$6,959, 9 positions).", "By stock: CVX made the most (+$3,436, 6 positions); AMD lost the most (-$2,299, 2 positions).", "By market regime of the previous close: risk-on 11 days -3.7% (SPY -5.1%); neutral 6 days -1.5% (SPY -3.3%); uncertain 162 days -8.0% (SPY -19.0%); risk-off 72 days +8.4% (SPY +9.5%).", "News: 32 positions opened on news made -$4,667, the other 38 -$998; 19 positions closed on news made -$6,853; over the five days from the fill, news-triggered buys trailed SPY by 0.53% (t -0.8).", "Concentration: the positions lost $5,666 in all; the best was CVX bought 2022-01-11 and sold 2022-03-10, +$4,005. Had the best position made nothing, the run would have returned -9.3% instead of -5.3%." ], "runs": 1, "summary": { "alpha_annual": -0.010378, "alpha_t": -0.152792, "best_trade_share": null, "beta_growth": -0.32817, "beta_market": 0.583983, "beta_momentum": 0.054402, "beta_size": -0.048018, "buy_excess_5d": 0.001525, "buy_excess_5d_t": 0.360637, "capm_alpha_annual": 0.02924, "capm_alpha_t": 0.374767, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -8.2e-05, "decision_excess_5d_t": -0.036616, "decision_hit_rate_5d": 0.504167, "decision_ic": -0.005842, "decision_ic_t": -0.285309, "decisions": 240.0, "exposure_return_annual": -0.049508, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.032177, "holdings_ic_t": 1.171045, "long_contribution": -0.056655, "news_decision_share": 0.3875, "news_entry_pnl": -4667.163822, "r2": 0.739891, "return_without_best_trade": -0.093244, "sell_avoided_5d": -0.001398, "sell_avoided_5d_t": -0.506709, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.053088 } }, "avg_exposure": 0.717214, "avg_holding_days": 27.189908, "beta": 0.462461, "excess_vs_equal_weight": 0.070413, "excess_vs_spy": 0.133376, "files": [ "claude-haiku-4.5-2022-blind-s0.json" ], "max_drawdown": -0.149295, "n": 1, "round_trips": 132.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": -0.440788, "sharpe_range": null, "steps": 50, "total_return": -0.053088, "total_return_range": null, "turnover": 6.360059, "usage": { "calls": 50.0, "input_tokens": 677720.0, "list_price_usd": 0.8674, "output_tokens": 37945.0 } }, "cutoff": "2025-07", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": "2026", "missing_answers": { "blind": 200, "named": 200, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ -0.08508802945648607, 0.5407086179596146 ], "recall_direction_accuracy": 0.6, "recall_majority_baseline": 0.54, "recall_score": 0.24177671068427373, "recall_verdict": "no sign", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-haiku-4.5", "named": { "after_cutoff": null, "alpha_annual": -0.038387, "analysis": { "by_sector": [ { "contribution": 0.038699, "group": "Health Care", "runs": 1 }, { "contribution": 0.016754, "group": "Energy", "runs": 1 }, { "contribution": 0.0128, "group": "Communication Services", "runs": 1 }, { "contribution": -0.030821, "group": "Financials", "runs": 1 }, { "contribution": -0.046007, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.047108, "group": "Technology", "runs": 1 } ], "by_trigger": [ { "contribution": -0.029843, "group": "other", "runs": 1 }, { "contribution": -0.068285, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.05% on average (142 buys, 52% of them ahead of SPY; t = -0.1, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.58%, so selling gave that up (136 sells; t = -1.4, within noise).", "Decision IC: on the 50 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.022 with the next five days' returns (IC IR -0.16, t = -1.1, within noise).", "Holdings IC: the weights it held after each of 50 decision days had a mean rank IC of 0.014 with the next five days' returns (t = 0.5, within noise).", "Factor exposures: market 0.48 (t 16.9), size -0.11 (t -1.9), growth -0.31 (t -3.4), momentum 0.06 (t 1.5); together they explain 62% of the daily variation (R²).", "Of its -10.7% a year over T-bills (arithmetic mean): market -9.3%, growth +5.5%, momentum +0.4%, size +0.3%; the alpha left is -7.6% a year (t = -1.0, within noise).", "Verdict: the exposures account for -3.1% a year of it, and the -7.6% a year left over is not significant: one period of daily returns cannot tell it from luck.", "P&L: -$9,487 in all, -$9,813 from 85 positions and +$326 of interest on idle cash.", "By holding period: 61+ days made the most (+$3,517, 10 positions); 3-5 days lost the most (-$7,259, 37 positions).", "By sector: Health Care made the most (+$3,870, 20 positions); Technology lost the most (-$4,711, 9 positions).", "By stock: CVX made the most (+$1,677, 4 positions); TSLA lost the most (-$3,236, 4 positions).", "By market regime of the previous close: risk-on 11 days -2.7% (SPY -5.1%); neutral 6 days -1.8% (SPY -3.3%); uncertain 162 days -11.6% (SPY -19.0%); risk-off 72 days +7.1% (SPY +9.5%).", "News: 45 positions opened on news made -$6,828, the other 40 -$2,984; 17 positions closed on news made -$5,521; over the five days from the fill, news-triggered buys beat SPY by 0.17% (t 0.4).", "Concentration: the positions lost $9,813 in all; the best was CVX bought 2022-01-11 and sold 2022-03-10, +$2,591. Had the best position made nothing, the run would have returned -12.1% instead of -9.5%." ], "runs": 1, "summary": { "alpha_annual": -0.076172, "alpha_t": -0.987738, "best_trade_share": null, "beta_growth": -0.309157, "beta_market": 0.482774, "beta_momentum": 0.06198, "beta_size": -0.107307, "buy_excess_5d": -0.000503, "buy_excess_5d_t": -0.139004, "capm_alpha_annual": -0.038387, "capm_alpha_t": -0.427844, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.003099, "decision_excess_5d_t": -1.208969, "decision_hit_rate_5d": 0.464029, "decision_ic": -0.022363, "decision_ic_t": -1.097698, "decisions": 278.0, "exposure_return_annual": -0.031281, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.013944, "holdings_ic_t": 0.451933, "long_contribution": -0.098128, "news_decision_share": 0.323741, "news_entry_pnl": -6828.466641, "r2": 0.620813, "return_without_best_trade": -0.12078, "sell_avoided_5d": -0.00581, "sell_avoided_5d_t": -1.440309, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.094767 } }, "avg_exposure": 0.662429, "avg_holding_days": 21.401295, "beta": 0.35837, "excess_vs_equal_weight": 0.028734, "excess_vs_spy": 0.091697, "files": [ "claude-haiku-4.5-2022-named-s0.json" ], "max_drawdown": -0.185292, "n": 1, "round_trips": 136.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": -0.904849, "sharpe_range": null, "steps": 50, "total_return": -0.094767, "total_return_range": null, "turnover": 7.772601, "usage": { "calls": 50.0, "input_tokens": 677988.0, "list_price_usd": 0.8684, "output_tokens": 38090.0 } }, "named_minus_blind": -0.041679, "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-4.5", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": 0.028732, "analysis": { "by_sector": [ { "contribution": 0.013213, "group": "Industrials", "runs": 1 }, { "contribution": 0.006053, "group": "Consumer Discretionary", "runs": 2 }, { "contribution": 0.002087, "group": "Health Care", "runs": 1 }, { "contribution": -0.004273, "group": "Financials", "runs": 3 }, { "contribution": -0.043589, "group": "Technology", "runs": 2 } ], "by_trigger": [ { "contribution": 0.048944, "group": "no reason", "runs": 1 }, { "contribution": -0.011154, "group": "news", "runs": 1 }, { "contribution": -0.036794, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.027089, "alpha_t": -0.050425, "best_trade_share": null, "beta_growth": -0.008731, "beta_market": 0.252236, "beta_momentum": -0.040955, "beta_size": 0.013178, "buy_excess_5d": -0.006626, "buy_excess_5d_t": -0.203234, "capm_alpha_annual": 0.028732, "capm_alpha_t": -0.034975, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.002272, "decision_excess_5d_t": 0.36771, "decision_hit_rate_5d": 0.366667, "decision_ic": -0.008203, "decision_ic_t": 0.109088, "decisions": 4.666667, "exposure_return_annual": -0.050335, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.006797, "holdings_ic_t": -0.38996, "long_contribution": -0.024197, "news_decision_share": 0.133333, "news_entry_pnl": -371.79743, "r2": 0.573685, "return_without_best_trade": -0.023385, "sell_avoided_5d": 0.049652, "sell_avoided_5d_t": 4.137159, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.015511 } }, "avg_exposure": 0.25823, "avg_holding_days": 37.541652, "beta": 0.269701, "excess_vs_equal_weight": 0.10799, "excess_vs_spy": 0.170953, "files": [ "llama3.2-2022-blind-s0.json", "llama3.2-2022-blind-s1.json", "llama3.2-2022-blind-s2.json" ], "max_drawdown": -0.085675, "n": 3, "round_trips": 0.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.615363, "sharpe_range": [ -1.191246, -0.129505 ], "steps": 50, "total_return": -0.015511, "total_return_range": [ -0.030259, -0.00237 ], "turnover": 0.275148, "usage": { "calls": 150.0, "input_tokens": 1740207.0, "list_price_usd": 0.0, "output_tokens": 13583.0 } }, "cutoff": "2023-12", "memory": { "beyond_control": 0.12406250000000008, "beyond_control_ci95": [ -0.09981181168302933, 0.31991059966273083 ], "ci95": [ -0.083958877845794, 0.21685592336407544 ], "control_period": "2026", "missing_answers": { "blind": 0, "named": 0, "recall": 0 }, "p_value": 0.18190904547726136, "paired": "run", "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "recall_ci95": [ -0.21643455879237317, 0.3287651931288638 ], "recall_direction_accuracy": 0.46, "recall_majority_baseline": 0.54, "recall_score": 0.06122743378756368, "recall_verdict": "no sign", "recall_windows": 50, "score": 0.07458333333333333, "scored_points": "before", "verdict": "inconclusive", "why": "too few pairs to tell: AUC gap +0.075 before the cutoff and -0.049 after it (the control), difference +0.124, 95% CI [-0.100, +0.320]" }, "model": "llama3.2", "named": { "after_cutoff": null, "alpha_annual": -0.027517, "analysis": { "by_sector": [ { "contribution": 0.003356, "group": "Health Care", "runs": 1 }, { "contribution": -0.008227, "group": "Financials", "runs": 1 }, { "contribution": -0.022693, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.088187, "group": "Technology", "runs": 3 } ], "by_trigger": [ { "contribution": -0.057683, "group": "other", "runs": 3 }, { "contribution": -0.059538, "group": "no reason", "runs": 2 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.003759, "alpha_t": 0.047565, "best_trade_share": null, "beta_growth": 0.203154, "beta_market": 0.330833, "beta_momentum": -0.012193, "beta_size": -0.079615, "buy_excess_5d": -0.003647, "buy_excess_5d_t": -0.527211, "capm_alpha_annual": -0.027517, "capm_alpha_t": -0.57915, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.002609, "decision_excess_5d_t": -0.47491, "decision_hit_rate_5d": 0.428205, "decision_ic": -0.031527, "decision_ic_t": -0.972528, "decisions": 14.0, "exposure_return_annual": -0.098416, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.021799, "holdings_ic_t": -1.322065, "long_contribution": -0.097375, "news_decision_share": 0.359951, "news_entry_pnl": 0.0, "r2": 0.856543, "return_without_best_trade": -0.089411, "sell_avoided_5d": 0.001367, "sell_avoided_5d_t": 0.167432, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.088192 } }, "avg_exposure": 0.294751, "avg_holding_days": 119.905395, "beta": 0.38739, "excess_vs_equal_weight": 0.035309, "excess_vs_spy": 0.098272, "files": [ "llama3.2-2022-named-s0.json", "llama3.2-2022-named-s1.json", "llama3.2-2022-named-s2.json" ], "max_drawdown": -0.1087, "n": 3, "round_trips": 2.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.968585, "sharpe_range": [ -1.098714, -0.797703 ], "steps": 50, "total_return": -0.088192, "total_return_range": [ -0.144194, -0.043892 ], "turnover": 0.258319, "usage": { "calls": 150.0, "input_tokens": 1733462.0, "list_price_usd": 0.0, "output_tokens": 10738.0 } }, "named_minus_blind": -0.072681, "open_weights": true, "provider": "Meta", "slug": "llama3.2", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": -0.025727, "analysis": { "by_sector": [ { "contribution": 0.024486, "group": "Health Care", "runs": 3 }, { "contribution": 0.009015, "group": "Energy", "runs": 3 }, { "contribution": 0.000899, "group": "Communication Services", "runs": 1 }, { "contribution": -0.010675, "group": "Consumer Staples", "runs": 2 }, { "contribution": -0.010977, "group": "Industrials", "runs": 3 }, { "contribution": -0.016622, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": -0.029023, "group": "Technology", "runs": 3 } ], "by_trigger": [ { "contribution": 0.002761, "group": "news", "runs": 3 }, { "contribution": -0.061429, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.042145, "alpha_t": -1.11445, "best_trade_share": null, "beta_growth": -0.14154, "beta_market": 0.268907, "beta_momentum": 0.066045, "beta_size": -0.027761, "buy_excess_5d": 0.004674, "buy_excess_5d_t": 1.085243, "capm_alpha_annual": -0.025727, "capm_alpha_t": -0.614716, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.001723, "decision_excess_5d_t": -0.558028, "decision_hit_rate_5d": 0.491737, "decision_ic": -0.015515, "decision_ic_t": -0.595651, "decisions": 164.0, "exposure_return_annual": -0.021415, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.027423, "holdings_ic_t": 0.965511, "long_contribution": -0.058668, "news_decision_share": 0.174341, "news_entry_pnl": 276.073993, "r2": 0.620816, "return_without_best_trade": -0.061534, "sell_avoided_5d": -0.005018, "sell_avoided_5d_t": -1.244007, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.049695 } }, "avg_exposure": 0.370506, "avg_holding_days": 25.783974, "beta": 0.196308, "excess_vs_equal_weight": 0.073806, "excess_vs_spy": 0.136769, "files": [ "gpt-5.6-luna-2022-blind-s0.json", "gpt-5.6-luna-2022-blind-s1.json", "gpt-5.6-luna-2022-blind-s2.json" ], "max_drawdown": -0.077288, "n": 3, "round_trips": 108.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.959283, "sharpe_range": [ -1.346211, -0.56567 ], "steps": 50, "total_return": -0.049695, "total_return_range": [ -0.070379, -0.023285 ], "turnover": 3.84358, "usage": { "calls": 150.0, "input_tokens": 1785053.0, "list_price_usd": 0.3873, "output_tokens": 25182.0 } }, "cutoff": "2026-02-16", "memory": { "beyond_control": -0.05253333333333343, "beyond_control_ci95": [ -0.22930225587023276, 0.12641406209681835 ], "ci95": [ -0.15411434262961507, 0.1311767792882845 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": -0.01333333333333342, "verdict": "inconclusive", "why": "too few pairs to tell: AUC gap -0.013 before the cutoff and +0.039 after it (the control), difference -0.053, 95% CI [-0.229, +0.126]" }, "model": "gpt-5.6-luna", "named": { "after_cutoff": null, "alpha_annual": -0.010612, "analysis": { "by_sector": [ { "contribution": 0.016561, "group": "Energy", "runs": 3 }, { "contribution": 0.003979, "group": "Financials", "runs": 3 }, { "contribution": 0.001057, "group": "Utilities", "runs": 1 }, { "contribution": -0.004118, "group": "Health Care", "runs": 2 }, { "contribution": -0.011841, "group": "Industrials", "runs": 1 }, { "contribution": -0.017988, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.027172, "group": "Communication Services", "runs": 3 }, { "contribution": -0.03798, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.044782, "group": "Technology", "runs": 3 } ], "by_trigger": [ { "contribution": -0.004887, "group": "news", "runs": 3 }, { "contribution": -0.101316, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.001575, "alpha_t": 0.028276, "best_trade_share": null, "beta_growth": 0.065727, "beta_market": 0.572627, "beta_momentum": 0.126724, "beta_size": -0.114614, "buy_excess_5d": 0.002184, "buy_excess_5d_t": 0.768957, "capm_alpha_annual": -0.010612, "capm_alpha_t": -0.134515, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.000298, "decision_excess_5d_t": -0.028952, "decision_hit_rate_5d": 0.461303, "decision_ic": -0.006886, "decision_ic_t": -0.106783, "decisions": 69.0, "exposure_return_annual": -0.109945, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.013672, "holdings_ic_t": -0.79278, "long_contribution": -0.106203, "news_decision_share": 0.208298, "news_entry_pnl": -488.675265, "r2": 0.877282, "return_without_best_trade": -0.119207, "sell_avoided_5d": -0.00236, "sell_avoided_5d_t": -0.403203, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.100315 } }, "avg_exposure": 0.63409, "avg_holding_days": 95.534754, "beta": 0.523588, "excess_vs_equal_weight": 0.023186, "excess_vs_spy": 0.086149, "files": [ "gpt-5.6-luna-2022-named-s0.json", "gpt-5.6-luna-2022-named-s1.json", "gpt-5.6-luna-2022-named-s2.json" ], "max_drawdown": -0.163449, "n": 3, "round_trips": 41.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.795087, "sharpe_range": [ -1.036546, -0.617438 ], "steps": 50, "total_return": -0.100315, "total_return_range": [ -0.140737, -0.077766 ], "turnover": 1.124191, "usage": { "calls": 150.0, "input_tokens": 1784768.0, "list_price_usd": 0.3742, "output_tokens": 14349.0 } }, "named_minus_blind": -0.05062, "open_weights": false, "provider": "OpenAI", "slug": "gpt-5.6-luna", "status": "may_remember" }, { "blind": { "after_cutoff": null, "alpha_annual": -0.05571, "analysis": { "by_sector": [ { "contribution": 0.049705, "group": "Health Care", "runs": 3 }, { "contribution": 0.035309, "group": "Energy", "runs": 3 }, { "contribution": -0.006524, "group": "Communication Services", "runs": 2 }, { "contribution": -0.026619, "group": "Financials", "runs": 3 }, { "contribution": -0.031899, "group": "Consumer Staples", "runs": 2 }, { "contribution": -0.043213, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.043816, "group": "Technology", "runs": 2 }, { "contribution": -0.053353, "group": "Industrials", "runs": 2 } ], "by_trigger": [ { "contribution": -0.053864, "group": "other", "runs": 3 }, { "contribution": -0.057057, "group": "news", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.11097, "alpha_t": -1.50129, "best_trade_share": null, "beta_growth": -0.453485, "beta_market": 0.503988, "beta_momentum": 0.062031, "beta_size": -0.093588, "buy_excess_5d": 0.006735, "buy_excess_5d_t": 1.735393, "capm_alpha_annual": -0.05571, "capm_alpha_t": -0.668793, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.004763, "decision_excess_5d_t": 1.378194, "decision_hit_rate_5d": 0.551022, "decision_ic": 0.022687, "decision_ic_t": 1.039809, "decisions": 134.333333, "exposure_return_annual": -0.010252, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.042422, "holdings_ic_t": 1.522314, "long_contribution": -0.110921, "news_decision_share": 0.357181, "news_entry_pnl": -5705.692581, "r2": 0.575144, "return_without_best_trade": -0.13699, "sell_avoided_5d": 0.002712, "sell_avoided_5d_t": 0.486661, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.107477 } }, "avg_exposure": 0.691637, "avg_holding_days": 22.586625, "beta": 0.339934, "excess_vs_equal_weight": 0.016024, "excess_vs_spy": 0.078987, "files": [ "gemini-3.5-flash-2022-blind-s0.json", "gemini-3.5-flash-2022-blind-s1.json", "gemini-3.5-flash-2022-blind-s2.json" ], "max_drawdown": -0.158627, "n": 3, "round_trips": 66.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.972961, "sharpe_range": [ -1.034873, -0.850557 ], "steps": 50, "total_return": -0.107477, "total_return_range": [ -0.12531, -0.08294 ], "turnover": 6.072651, "usage": { "calls": 150.0, "input_tokens": 2218170.0, "list_price_usd": 3.6581, "output_tokens": 36762.0 } }, "cutoff": "2025-01", "memory": { "beyond_control": 0.26217147435897437, "beyond_control_ci95": [ 0.10692519251739968, 0.41460789037310525 ], "ci95": [ 0.10234197443181825, 0.31672936965397164 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": 0.21229166666666666, "verdict": "contaminated", "why": "named records help more before the cutoff than after it: AUC gap +0.212 before the cutoff and -0.050 after it (the control), difference +0.262, 95% CI [+0.107, +0.415]" }, "model": "gemini-3.5-flash", "named": { "after_cutoff": null, "alpha_annual": 0.034468, "analysis": { "by_sector": [ { "contribution": 0.06734, "group": "Health Care", "runs": 3 }, { "contribution": 0.054457, "group": "Energy", "runs": 3 }, { "contribution": 0.011427, "group": "Communication Services", "runs": 3 }, { "contribution": -0.034156, "group": "Technology", "runs": 2 }, { "contribution": -0.036638, "group": "Financials", "runs": 3 }, { "contribution": -0.039452, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.065466, "group": "Consumer Staples", "runs": 3 } ], "by_trigger": [ { "contribution": 0.008913, "group": "news", "runs": 3 }, { "contribution": -0.067107, "group": "other", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.026046, "alpha_t": -0.340059, "best_trade_share": null, "beta_growth": -0.502537, "beta_market": 0.726389, "beta_momentum": 0.162586, "beta_size": -0.19097, "buy_excess_5d": 0.001345, "buy_excess_5d_t": 0.321745, "capm_alpha_annual": 0.034468, "capm_alpha_t": 0.346684, "concentrated": false, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.001015, "decision_excess_5d_t": -0.186471, "decision_hit_rate_5d": 0.488783, "decision_ic": -0.000682, "decision_ic_t": 0.00244, "decisions": 92.0, "exposure_return_annual": -0.034759, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.032893, "holdings_ic_t": 0.965934, "long_contribution": -0.058193, "news_decision_share": 0.3013, "news_entry_pnl": 891.312479, "r2": 0.742064, "return_without_best_trade": -0.088156, "sell_avoided_5d": -0.003661, "sell_avoided_5d_t": -0.50406, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": -0.186464, "total_return": -0.05701 } }, "avg_exposure": 0.905814, "avg_holding_days": 37.572545, "beta": 0.494348, "excess_vs_equal_weight": 0.066491, "excess_vs_spy": 0.129454, "files": [ "gemini-3.5-flash-2022-named-s0.json", "gemini-3.5-flash-2022-named-s1.json", "gemini-3.5-flash-2022-named-s2.json" ], "max_drawdown": -0.157997, "n": 3, "round_trips": 43.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.383553, "sharpe_range": [ -0.446177, -0.346037 ], "steps": 50, "total_return": -0.05701, "total_return_range": [ -0.068302, -0.050836 ], "turnover": 4.294541, "usage": { "calls": 150.0, "input_tokens": 2197363.0, "list_price_usd": 3.5426, "output_tokens": 27399.0 } }, "named_minus_blind": 0.050467, "open_weights": false, "provider": "Google", "slug": "gemini-3.5-flash", "status": "may_remember" } ], "title": "2022: the rate-hike bear market", "window": { "end": "2022-12-30", "start": "2022-01-03" } }, "2026": { "anchors": [ { "avg_exposure": 0.853892, "avg_holding_days": 36.548022, "excess_vs_equal_weight": 0.0754, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.049534, "name": "swing_trend", "period": "2026", "round_trips": 38.0, "sharpe": 1.270018, "total_return": 0.189646, "turnover": 4.525364, "volatility": 0.139155 }, { "avg_exposure": 0.507315, "avg_holding_days": 3.715188, "excess_vs_equal_weight": -0.064751, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.077081, "name": "swing_rsi2", "period": "2026", "round_trips": 293.0, "sharpe": 0.231632, "total_return": 0.049495, "turnover": 33.967142, "volatility": 0.121751 }, { "avg_exposure": 0.897368, "avg_holding_days": 18.646184, "excess_vs_equal_weight": -0.179699, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.168003, "name": "swing_breakout", "period": "2026", "round_trips": 97.0, "sharpe": -0.604915, "total_return": -0.065453, "turnover": 11.161807, "volatility": 0.166392 }, { "avg_exposure": 0.243079, "avg_holding_days": 10.884779, "excess_vs_equal_weight": -0.103709, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.045152, "name": "swing_pullback", "period": "2026", "round_trips": 49.0, "sharpe": -0.388048, "total_return": 0.010536, "turnover": 5.749205, "volatility": 0.056162 }, { "avg_exposure": 0.606111, "avg_holding_days": 9.42834, "excess_vs_equal_weight": -0.094541, "kind": "rule", "leak_check": "pass", "max_drawdown": -0.067143, "name": "swing_news", "period": "2026", "round_trips": 140.0, "sharpe": -0.051136, "total_return": 0.019704, "turnover": 16.246393, "volatility": 0.11709 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.08885, "name": "SPY", "period": "2026", "round_trips": null, "sharpe": 0.76515, "total_return": 0.12087, "turnover": null, "volatility": 0.140422 }, { "avg_exposure": null, "avg_holding_days": null, "kind": "buy and hold", "leak_check": null, "max_drawdown": -0.071956, "name": "Equal weight", "period": "2026", "round_trips": null, "sharpe": 0.848303, "total_return": 0.114246, "turnover": null, "volatility": 0.114603 } ], "clean": [ { "blind": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.042593, "excess_vs_spy": -0.049217, "from": "2026-01-02", "max_drawdown": -0.074214, "sharpe": 0.416132, "sharpe_range": null, "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": 0.071653, "total_return_range": null, "volatility": 0.128224 }, "alpha_annual": -0.005151, "analysis": { "by_sector": [ { "contribution": 0.043143, "group": "Energy", "runs": 1 }, { "contribution": 0.038572, "group": "Industrials", "runs": 1 }, { "contribution": 0.018063, "group": "Financials", "runs": 1 }, { "contribution": -0.003036, "group": "Real Estate", "runs": 1 }, { "contribution": -0.017267, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.022707, "group": "Technology", "runs": 1 } ], "by_trigger": [ { "contribution": 0.073687, "group": "other", "runs": 1 }, { "contribution": -0.010504, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.73% on average (106 buys, 44% of them ahead of SPY; t = -1.3, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.67%, so selling gave that up (112 sells; t = -1.1, within noise).", "Decision IC: on the 42 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.027 with the next five days' returns (IC IR -0.16, t = -1.1, within noise).", "Holdings IC: the weights it held after each of 43 decision days had a mean rank IC of -0.018 with the next five days' returns (t = -1.0, within noise).", "Factor exposures: market 0.59 (t 8.5), size 0.20 (t 2.7), growth -0.15 (t -1.4), momentum 0.01 (t 0.1); together they explain 39% of the daily variation (R²).", "Of its +5.3% a year over T-bills (arithmetic mean): market +6.4%, size +0.7%, growth -0.4%, momentum -0.1%; the alpha left is -1.2% a year (t = -0.1, within noise).", "Verdict: most of the return is factor exposure (+6.5% a year, market above all); the alpha left is not significant.", "P&L: +$7,212 in all, +$6,318 from 77 positions and +$894 of interest on idle cash.", "By holding period: 61+ days made the most (+$19,047, 11 positions); 3-5 days lost the most (-$12,279, 31 positions).", "By sector: Energy made the most (+$4,314, 3 positions); Technology lost the most (-$2,271, 27 positions).", "By stock: CAT made the most (+$3,457, 1 position); ORCL lost the most (-$2,550, 3 positions).", "By market regime of the previous close: risk-on 96 days +1.1% (SPY +2.0%); neutral 18 days +0.2% (SPY +4.5%); uncertain 97 days +5.7% (SPY +4.8%); risk-off 6 days +0.1% (SPY +0.3%).", "News: 41 positions opened on news made -$1,050, the other 36 +$7,369; 19 positions closed on news made -$921; over the five days from the fill, news-triggered buys trailed SPY by 0.81% (t -1.2).", "Concentration: the gains are spread out (the best two positions made 37% of them) but winners and losers nearly cancel, so the best two made 164% of the net P&L; the best was AMD bought 2026-04-09 and sold 2026-07-28, +$6,894. Had the best position made nothing, the run would have returned +0.3% instead of +7.2%." ], "runs": 1, "summary": { "alpha_annual": -0.011681, "alpha_t": -0.125095, "best_trade_share": 1.091151, "beta_growth": -0.150741, "beta_market": 0.591836, "beta_momentum": 0.005958, "beta_size": 0.195565, "buy_excess_5d": -0.007254, "buy_excess_5d_t": -1.329226, "capm_alpha_annual": -0.005151, "capm_alpha_t": -0.054525, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.006979, "decision_excess_5d_t": -1.434463, "decision_hit_rate_5d": 0.472477, "decision_ic": -0.027452, "decision_ic_t": -1.05061, "decisions": 218.0, "exposure_return_annual": 0.065039, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.0177, "holdings_ic_t": -0.972096, "long_contribution": 0.063183, "news_decision_share": 0.346667, "news_entry_pnl": -1050.402644, "r2": 0.392107, "return_without_best_trade": 0.00318, "sell_avoided_5d": -0.00672, "sell_avoided_5d_t": -1.091354, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.071653 } }, "avg_exposure": 0.737002, "avg_holding_days": 21.38943, "beta": 0.54456, "excess_vs_equal_weight": -0.042593, "excess_vs_spy": -0.049217, "files": [ "claude-haiku-4.5-2026-blind-s0.json" ], "max_drawdown": -0.074214, "n": 1, "round_trips": 116.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.416132, "sharpe_range": null, "steps": 44, "total_return": 0.071653, "total_return_range": null, "turnover": 7.856045, "usage": { "calls": 44.0, "input_tokens": 539675.0, "list_price_usd": 0.7011, "output_tokens": 32278.0 } }, "cutoff": "2025-07", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": null, "missing_answers": { "blind": 100, "named": 100, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ -0.09538192223476778, 0.46016522783197206 ], "recall_direction_accuracy": 0.44, "recall_majority_baseline": 0.56, "recall_score": 0.19462184873949578, "recall_verdict": "no sign", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-haiku-4.5", "named": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.025972, "excess_vs_spy": -0.032596, "from": "2026-01-02", "max_drawdown": -0.088627, "sharpe": 0.519476, "sharpe_range": null, "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": 0.088274, "total_return_range": null, "volatility": 0.1405 }, "alpha_annual": 0.004006, "analysis": { "by_sector": [ { "contribution": 0.022906, "group": "Technology", "runs": 1 }, { "contribution": 0.022683, "group": "Energy", "runs": 1 }, { "contribution": 0.02235, "group": "Industrials", "runs": 1 }, { "contribution": -0.004003, "group": "Materials", "runs": 1 }, { "contribution": -0.004412, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.012799, "group": "Communication Services", "runs": 1 } ], "by_trigger": [ { "contribution": 0.044824, "group": "news", "runs": 1 }, { "contribution": 0.033402, "group": "other", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.81% on average (121 buys, 40% of them ahead of SPY; t = -1.6, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.41%, so selling gave that up (140 sells; t = -0.8, within noise).", "Decision IC: on the 42 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.025 with the next five days' returns (IC IR -0.15, t = -1.0, within noise).", "Holdings IC: the weights it held after each of 43 decision days had a mean rank IC of -0.009 with the next five days' returns (t = -0.3, within noise).", "Factor exposures: market 0.57 (t 8.8), size 0.21 (t 3.1), growth 0.16 (t 1.0), momentum -0.07 (t -1.6); together they explain 44% of the daily variation (R²).", "Of its +7.3% a year over T-bills (arithmetic mean): market +6.1%, momentum +1.6%, size +0.7%, growth +0.4%; the alpha left is -1.5% a year (t = -0.1, within noise).", "Verdict: most of the return is factor exposure (+8.8% a year, market above all); the alpha left is not significant.", "P&L: +$8,875 in all, +$7,823 from 85 positions and +$1,052 of interest on idle cash.", "By holding period: 61+ days made the most (+$13,548, 13 positions); 3-5 days lost the most (-$10,691, 34 positions).", "By sector: Technology made the most (+$2,291, 33 positions); Communication Services lost the most (-$1,280, 10 positions).", "By stock: AMD made the most (+$7,116, 5 positions); ORCL lost the most (-$3,568, 5 positions).", "By market regime of the previous close: risk-on 96 days +4.2% (SPY +2.0%); neutral 18 days +2.9% (SPY +4.5%); uncertain 97 days +2.0% (SPY +4.8%); risk-off 6 days -0.4% (SPY +0.3%).", "News: 42 positions opened on news made +$4,482, the other 43 +$3,340; 22 positions closed on news made +$5,862; over the five days from the fill, news-triggered buys trailed SPY by 1.46% (t -2.2).", "Concentration: the profit rests on the best two positions, which made 173% of the positions' net P&L and 43% of all their gains; the best was AMD bought 2026-04-09 and sold 2026-06-22, +$9,462. Had the best position made nothing, the run would have returned -0.6% instead of +8.9%." ], "runs": 1, "summary": { "alpha_annual": -0.015113, "alpha_t": -0.138149, "best_trade_share": 1.209517, "beta_growth": 0.155078, "beta_market": 0.567272, "beta_momentum": -0.070839, "beta_size": 0.207047, "buy_excess_5d": -0.008055, "buy_excess_5d_t": -1.605659, "capm_alpha_annual": 0.004006, "capm_alpha_t": 0.035747, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.005952, "decision_excess_5d_t": -1.494801, "decision_hit_rate_5d": 0.482759, "decision_ic": -0.024516, "decision_ic_t": -0.9708, "decisions": 261.0, "exposure_return_annual": 0.088099, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.008622, "holdings_ic_t": -0.347963, "long_contribution": 0.078226, "news_decision_share": 0.335878, "news_entry_pnl": 4482.356775, "r2": 0.444719, "return_without_best_trade": -0.005865, "sell_avoided_5d": -0.004135, "sell_avoided_5d_t": -0.792523, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.088274 } }, "avg_exposure": 0.684796, "avg_holding_days": 20.591824, "beta": 0.642014, "excess_vs_equal_weight": -0.025972, "excess_vs_spy": -0.032596, "files": [ "claude-haiku-4.5-2026-named-s0.json" ], "max_drawdown": -0.088627, "n": 1, "round_trips": 141.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.519476, "sharpe_range": null, "steps": 44, "total_return": 0.088274, "total_return_range": null, "turnover": 8.294019, "usage": { "calls": 44.0, "input_tokens": 539325.0, "list_price_usd": 0.7255, "output_tokens": 37239.0 } }, "named_minus_blind": 0.016621, "named_minus_blind_after_cutoff": 0.016621, "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-4.5", "status": "clean" }, { "blind": { "after_cutoff": { "days": 197, "equal_weight_sharpe": 0.862105, "equal_weight_total_return": 0.106305, "excess_vs_equal_weight": -0.0447, "excess_vs_spy": -0.04501, "from": "2026-01-30", "max_drawdown": -0.090248, "sharpe": 0.343889, "sharpe_range": null, "spy_sharpe": 0.725438, "spy_total_return": 0.106615, "to": "2026-11-11", "total_return": 0.061605, "total_return_range": null, "volatility": 0.151784 }, "alpha_annual": -0.004197, "analysis": { "by_sector": [ { "contribution": 0.047322, "group": "Health Care", "runs": 1 }, { "contribution": 0.03487, "group": "Technology", "runs": 1 }, { "contribution": 0.016797, "group": "Industrials", "runs": 1 }, { "contribution": -0.001772, "group": "Real Estate", "runs": 1 }, { "contribution": -0.018701, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.020545, "group": "Communication Services", "runs": 1 } ], "by_trigger": [ { "contribution": 0.13883, "group": "other", "runs": 1 }, { "contribution": -0.062307, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought beat SPY by 0.25% on average (72 buys, 58% of them ahead of SPY; t = 0.5, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 0.06%, so selling gave that up (98 sells; t = -0.1, within noise).", "Decision IC: on the 43 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of 0.023 with the next five days' returns (IC IR 0.18, t = 1.2, within noise).", "Holdings IC: the weights it held after each of 43 decision days had a mean rank IC of -0.023 with the next five days' returns (t = -0.8, within noise).", "Factor exposures: market 0.59 (t 8.4), size 0.34 (t 4.7), growth -0.02 (t -0.1), momentum 0.11 (t 2.7); together they explain 47% of the daily variation (R²).", "Of the +6.7% a year it earned over T-bills (arithmetic mean): market +6.4%, momentum -2.6%, size +1.1%, growth -0.0%; the alpha left is +1.8% a year (t = 0.2, within noise).", "Verdict: most of the return is factor exposure (+4.8% a year, market above all); the alpha left is not significant.", "P&L: +$8,145 in all, +$7,652 from 67 positions and +$493 of interest on idle cash.", "By holding period: 61+ days made the most (+$16,519, 11 positions); 3-5 days the least (-$6,474, 12 positions).", "By sector: Health Care made the most (+$4,732, 10 positions); Communication Services the least (-$2,054, 5 positions).", "By stock: CAT made the most (+$4,994, 1 position); ORCL the least (-$3,072, 2 positions).", "By market regime of the previous close: risk-on 96 days +4.0% (SPY +2.0%); neutral 18 days -3.5% (SPY +4.5%); uncertain 97 days +7.6% (SPY +4.8%); risk-off 6 days +0.2% (SPY +0.3%).", "News: 26 positions opened on news made -$6,231, the other 41 +$13,883; 13 positions closed on news made -$4,777; over the five days from the fill, news-triggered buys trailed SPY by 0.95% (t -1.0).", "Concentration: the profit rests on the best two positions, which made 152% of the positions' net P&L and 42% of all their gains; the best was AMD bought 2026-04-09 and sold 2026-08-25, +$6,628. Had the best position made nothing, the run would have returned +1.5% instead of +8.1%." ], "runs": 1, "summary": { "alpha_annual": 0.018403, "alpha_t": 0.155763, "best_trade_share": 0.866167, "beta_growth": -0.017005, "beta_market": 0.593185, "beta_momentum": 0.113267, "beta_size": 0.337245, "buy_excess_5d": 0.002484, "buy_excess_5d_t": 0.526614, "capm_alpha_annual": -0.004197, "capm_alpha_t": -0.032286, "concentrated": true, "decision_excess_5d": 0.000714, "decision_excess_5d_t": 0.224041, "decision_hit_rate_5d": 0.558824, "decision_ic": 0.02252, "decision_ic_t": 1.167828, "decisions": 170.0, "exposure_return_annual": 0.048151, "holdings_ic": -0.022828, "holdings_ic_t": -0.832104, "news_decision_share": 0.290698, "news_entry_pnl": -6230.674462, "r2": 0.473616, "return_without_best_trade": 0.01517, "sell_avoided_5d": -0.000586, "sell_avoided_5d_t": -0.134679, "spy_return": 0.12087, "total_return": 0.080978 } }, "avg_exposure": 0.858106, "avg_holding_days": 28.776131, "beta": 0.658498, "excess_vs_equal_weight": -0.033268, "excess_vs_spy": -0.039892, "files": [ "claude-sonnet-4.6-2026-blind-s0.json" ], "max_drawdown": -0.090248, "n": 1, "round_trips": 100.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.443112, "sharpe_range": null, "steps": 44, "total_return": 0.080978, "total_return_range": null, "turnover": 7.123251, "usage": { "calls": 44.0, "input_tokens": 538998.0, "list_price_usd": 2.1823, "output_tokens": 37690.0 } }, "cutoff": "2026-01", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": null, "control_period": null, "missing_answers": { "blind": 100, "named": 100, "recall": 0 }, "p_value": null, "paired": "not run (Claude row cap)", "probe": "riskaware.research.memory: direct recall only (50 windows per period, temperature 0), with the model's cutoff; the paired named/blind probe was not run (Claude row cap); recall_score = rank correlation of the guessed and actual moves, each scaled by the stock's usual 20-day move", "recall_ci95": [ -0.37569943068195544, 0.222223318922643 ], "recall_direction_accuracy": 0.42, "recall_majority_baseline": 0.56, "recall_score": -0.06746698679471788, "recall_verdict": "no sign", "recall_windows": 50, "score": null, "scored_points": "all", "verdict": null, "why": "paired probe not run (Claude row cap); direct recall only" }, "model": "claude-sonnet-4.6", "named": { "after_cutoff": { "days": 197, "equal_weight_sharpe": 0.862105, "equal_weight_total_return": 0.106305, "excess_vs_equal_weight": -0.031026, "excess_vs_spy": -0.031336, "from": "2026-01-30", "max_drawdown": -0.072699, "sharpe": 0.513104, "sharpe_range": null, "spy_sharpe": 0.725438, "spy_total_return": 0.106615, "to": "2026-11-11", "total_return": 0.075279, "total_return_range": null, "volatility": 0.126829 }, "alpha_annual": -0.005549, "analysis": { "by_sector": [ { "contribution": 0.044222, "group": "Technology", "runs": 1 }, { "contribution": 0.020711, "group": "Health Care", "runs": 1 }, { "contribution": 0.015746, "group": "Financials", "runs": 1 }, { "contribution": 0.004236, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.002626, "group": "Consumer Discretionary", "runs": 1 }, { "contribution": -0.023954, "group": "Communication Services", "runs": 1 } ], "by_trigger": [ { "contribution": 0.055618, "group": "other", "runs": 1 }, { "contribution": 0.021839, "group": "news", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.48% on average (68 buys, 46% of them ahead of SPY; t = -0.8, within noise).", "Sells: over the five trading days from the fill, the stocks it sold lagged SPY by 0.11%, so selling avoided that (97 sells; t = 0.3, within noise).", "Decision IC: on the 43 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of 0.007 with the next five days' returns (IC IR 0.05, t = 0.3, within noise).", "Holdings IC: the weights it held after each of 43 decision days had a mean rank IC of -0.018 with the next five days' returns (t = -0.7, within noise).", "Factor exposures: market 0.70 (t 10.9), size 0.04 (t 0.7), growth -0.11 (t -1.3), momentum 0.09 (t 3.4); together they explain 62% of the daily variation (R²).", "Of the +6.8% a year it earned over T-bills (arithmetic mean): market +7.5%, momentum -2.1%, growth -0.3%, size +0.1%; the alpha left is +1.5% a year (t = 0.2, within noise).", "Verdict: most of the return is factor exposure (+5.3% a year, market above all); the alpha left is not significant.", "P&L: +$8,579 in all, +$7,746 from 39 positions and +$834 of interest on idle cash.", "By holding period: 61+ days made the most (+$17,424, 15 positions); 21-60 days the least (-$3,574, 10 positions).", "By sector: Technology made the most (+$4,422, 10 positions); Communication Services the least (-$2,395, 4 positions).", "By stock: CSCO made the most (+$4,163, 1 position); NVDA the least (-$1,672, 2 positions).", "By market regime of the previous close: risk-on 96 days -1.0% (SPY +2.0%); neutral 18 days +1.9% (SPY +4.5%); uncertain 97 days +7.3% (SPY +4.8%); risk-off 6 days +0.3% (SPY +0.3%).", "News: 13 positions opened on news made +$2,184, the other 26 +$5,562; 1 positions closed on news made -$1,077; over the five days from the fill, news-triggered buys trailed SPY by 0.53% (t -0.4).", "Concentration: the gains are spread out (the best two positions made 32% of them) but winners and losers nearly cancel, so the best two made 91% of the net P&L; the best was CSCO bought 2026-01-27 and sold 2026-07-21, +$4,163. Had the best position made nothing, the run would have returned +4.4% instead of +8.6%." ], "runs": 1, "summary": { "alpha_annual": 0.014971, "alpha_t": 0.172936, "best_trade_share": 0.537513, "beta_growth": -0.10798, "beta_market": 0.700496, "beta_momentum": 0.090757, "beta_size": 0.037292, "buy_excess_5d": -0.004805, "buy_excess_5d_t": -0.80491, "capm_alpha_annual": -0.005549, "capm_alpha_t": -0.062013, "concentrated": true, "decision_excess_5d": -0.001321, "decision_excess_5d_t": -0.37099, "decision_hit_rate_5d": 0.50303, "decision_ic": 0.006649, "decision_ic_t": 0.31338, "decisions": 165.0, "exposure_return_annual": 0.052742, "holdings_ic": -0.018027, "holdings_ic_t": -0.692922, "news_decision_share": 0.220238, "news_entry_pnl": 2183.903714, "r2": 0.619227, "return_without_best_trade": 0.044161, "sell_avoided_5d": 0.001121, "sell_avoided_5d_t": 0.273711, "spy_return": 0.12087, "total_return": 0.08532 } }, "avg_exposure": 0.750009, "avg_holding_days": 42.719808, "beta": 0.681861, "excess_vs_equal_weight": -0.028926, "excess_vs_spy": -0.03555, "files": [ "claude-sonnet-4.6-2026-named-s0.json" ], "max_drawdown": -0.072699, "n": 1, "round_trips": 100.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": 0.542748, "sharpe_range": null, "steps": 44, "total_return": 0.08532, "total_return_range": null, "turnover": 4.311977, "usage": { "calls": 44.0, "input_tokens": 539491.0, "list_price_usd": 2.1861, "output_tokens": 37839.0 } }, "named_minus_blind": 0.004342, "named_minus_blind_after_cutoff": 0.013674, "open_weights": false, "provider": "Anthropic", "slug": "claude-sonnet-4.6", "status": "partly" }, { "blind": { "after_cutoff": { "days": 187, "equal_weight_sharpe": 0.935723, "equal_weight_total_return": 0.107209, "excess_vs_equal_weight": -0.061779, "excess_vs_spy": -0.077775, "from": "2026-02-13", "max_drawdown": -0.036582, "sharpe": 0.335632, "sharpe_range": [ -0.379239, 1.729767 ], "spy_sharpe": 0.916899, "spy_total_return": 0.123205, "to": "2026-11-11", "total_return": 0.04543, "total_return_range": [ 0.008466, 0.117679 ], "volatility": 0.06211 }, "alpha_annual": 0.057998, "analysis": { "by_sector": [ { "contribution": 0.0311, "group": "Health Care", "runs": 2 }, { "contribution": 0.031027, "group": "Industrials", "runs": 3 }, { "contribution": 0.029025, "group": "Energy", "runs": 3 }, { "contribution": 0.015524, "group": "Consumer Staples", "runs": 1 }, { "contribution": 0.0086, "group": "Real Estate", "runs": 1 }, { "contribution": -0.003547, "group": "Consumer Discretionary", "runs": 2 }, { "contribution": -0.006686, "group": "Utilities", "runs": 2 }, { "contribution": -0.008942, "group": "Communication Services", "runs": 1 }, { "contribution": -0.00965, "group": "Financials", "runs": 1 }, { "contribution": -0.019932, "group": "Technology", "runs": 2 } ], "by_trigger": [ { "contribution": 0.052144, "group": "other", "runs": 3 }, { "contribution": 0.033266, "group": "news", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.058696, "alpha_t": 0.917292, "best_trade_share": 0.334317, "beta_growth": -0.355781, "beta_market": 0.347626, "beta_momentum": 0.046387, "beta_size": 0.105722, "buy_excess_5d": -0.000717, "buy_excess_5d_t": -0.199453, "capm_alpha_annual": 0.057998, "capm_alpha_t": 0.802843, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.002786, "decision_excess_5d_t": -0.904485, "decision_hit_rate_5d": 0.482998, "decision_ic": -0.014671, "decision_ic_t": -0.613145, "decisions": 177.666667, "exposure_return_annual": 0.021415, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.003888, "holdings_ic_t": 0.171815, "long_contribution": 0.085409, "news_decision_share": 0.174377, "news_entry_pnl": 3326.568745, "r2": 0.392262, "return_without_best_trade": 0.081107, "sell_avoided_5d": -0.003789, "sell_avoided_5d_t": -0.857939, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.10261 } }, "avg_exposure": 0.470381, "avg_holding_days": 25.53365, "beta": 0.20581, "excess_vs_equal_weight": -0.011636, "excess_vs_spy": -0.01826, "files": [ "gpt-5.6-luna-2026-blind-s0.json", "gpt-5.6-luna-2026-blind-s1.json", "gpt-5.6-luna-2026-blind-s2.json" ], "max_drawdown": -0.036582, "n": 3, "round_trips": 126.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 1.099159, "sharpe_range": [ 0.723859, 1.845646 ], "steps": 44, "total_return": 0.10261, "total_return_range": [ 0.073108, 0.160954 ], "turnover": 5.022479, "usage": { "calls": 132.0, "input_tokens": 1439207.0, "list_price_usd": 0.3235, "output_tokens": 29708.0 } }, "cutoff": "2026-02-16", "memory": { "beyond_control": -0.4726608187134502, "beyond_control_ci95": [ -0.9402735226034858, -0.08676659891598945 ], "ci95": [ -0.8571428571428572, -0.0714285714285714 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": -0.42499999999999993, "verdict": "inconclusive", "why": "only 9 pairs; 30 or more are needed for a verdict (no control window after the cutoff to compare with)" }, "model": "gpt-5.6-luna", "named": { "after_cutoff": { "days": 187, "equal_weight_sharpe": 0.935723, "equal_weight_total_return": 0.107209, "excess_vs_equal_weight": 0.022626, "excess_vs_spy": 0.00663, "from": "2026-02-13", "max_drawdown": -0.079036, "sharpe": 1.334337, "sharpe_range": [ 0.905854, 1.76578 ], "spy_sharpe": 0.916899, "spy_total_return": 0.123205, "to": "2026-11-11", "total_return": 0.129835, "total_return_range": [ 0.083022, 0.180588 ], "volatility": 0.096544 }, "alpha_annual": 0.022904, "analysis": { "by_sector": [ { "contribution": 0.023353, "group": "Financials", "runs": 2 }, { "contribution": 0.022087, "group": "Industrials", "runs": 2 }, { "contribution": 0.019686, "group": "Technology", "runs": 3 }, { "contribution": 0.016658, "group": "Consumer Discretionary", "runs": 2 }, { "contribution": 0.015575, "group": "Health Care", "runs": 3 }, { "contribution": 0.003918, "group": "Real Estate", "runs": 3 }, { "contribution": 0.003072, "group": "Consumer Staples", "runs": 1 }, { "contribution": 0.002569, "group": "Communication Services", "runs": 2 } ], "by_trigger": [ { "contribution": 0.091156, "group": "other", "runs": 3 }, { "contribution": 0.002021, "group": "news", "runs": 1 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.010204, "alpha_t": 0.153699, "best_trade_share": 0.299585, "beta_growth": -0.516443, "beta_market": 0.779636, "beta_momentum": 0.001264, "beta_size": -0.030606, "buy_excess_5d": 0.000126, "buy_excess_5d_t": 0.044549, "capm_alpha_annual": 0.022904, "capm_alpha_t": 0.293271, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.001351, "decision_excess_5d_t": -0.289787, "decision_hit_rate_5d": 0.510329, "decision_ic": -0.019801, "decision_ic_t": -0.537039, "decisions": 70.666667, "exposure_return_annual": 0.069905, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.005783, "holdings_ic_t": 0.346796, "long_contribution": 0.091829, "news_decision_share": 0.11549, "news_entry_pnl": 67.353652, "r2": 0.750927, "return_without_best_trade": 0.074364, "sell_avoided_5d": -0.002741, "sell_avoided_5d_t": -0.172794, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.099913 } }, "avg_exposure": 0.748447, "avg_holding_days": 104.449676, "beta": 0.532423, "excess_vs_equal_weight": -0.014333, "excess_vs_spy": -0.020957, "files": [ "gpt-5.6-luna-2026-named-s0.json", "gpt-5.6-luna-2026-named-s1.json", "gpt-5.6-luna-2026-named-s2.json" ], "max_drawdown": -0.09676, "n": 3, "round_trips": 43.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.786383, "sharpe_range": [ 0.533455, 1.128592 ], "steps": 44, "total_return": 0.099913, "total_return_range": [ 0.074604, 0.142568 ], "turnover": 1.170806, "usage": { "calls": 132.0, "input_tokens": 1431790.0, "list_price_usd": 0.3044, "output_tokens": 15014.0 } }, "named_minus_blind": -0.002697, "named_minus_blind_after_cutoff": 0.084405, "open_weights": false, "provider": "OpenAI", "slug": "gpt-5.6-luna", "status": "partly" }, { "blind": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.048863, "excess_vs_spy": -0.055487, "from": "2026-01-02", "max_drawdown": -0.100508, "sharpe": 0.325717, "sharpe_range": [ 0.149895, 0.543446 ], "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": 0.065383, "total_return_range": [ 0.040861, 0.095147 ], "volatility": 0.157953 }, "alpha_annual": -0.029174, "analysis": { "by_sector": [ { "contribution": 0.088283, "group": "Health Care", "runs": 3 }, { "contribution": 0.020642, "group": "Technology", "runs": 3 }, { "contribution": 0.015697, "group": "Energy", "runs": 3 }, { "contribution": 0.005, "group": "Industrials", "runs": 2 }, { "contribution": -0.017121, "group": "Consumer Staples", "runs": 1 }, { "contribution": -0.024098, "group": "Consumer Discretionary", "runs": 3 }, { "contribution": -0.024648, "group": "Communication Services", "runs": 2 }, { "contribution": -0.028545, "group": "Financials", "runs": 1 } ], "by_trigger": [ { "contribution": 0.052767, "group": "other", "runs": 3 }, { "contribution": 0.011894, "group": "news", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.013333, "alpha_t": -0.093869, "best_trade_share": 1.125042, "beta_growth": -0.040903, "beta_market": 0.703753, "beta_momentum": 0.085224, "beta_size": 0.27359, "buy_excess_5d": -0.005308, "buy_excess_5d_t": -1.016205, "capm_alpha_annual": -0.029174, "capm_alpha_t": -0.213733, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.002736, "decision_excess_5d_t": -0.662786, "decision_hit_rate_5d": 0.530457, "decision_ic": -0.002676, "decision_ic_t": -0.107227, "decisions": 139.0, "exposure_return_annual": 0.063866, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.022328, "holdings_ic_t": -0.986034, "long_contribution": 0.064661, "news_decision_share": 0.385081, "news_entry_pnl": 1189.426262, "r2": 0.491357, "return_without_best_trade": 0.000315, "sell_avoided_5d": 0.000194, "sell_avoided_5d_t": 0.050604, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.065383 } }, "avg_exposure": 0.972167, "avg_holding_days": 30.795886, "beta": 0.741852, "excess_vs_equal_weight": -0.048863, "excess_vs_spy": -0.055487, "files": [ "gemini-3.5-flash-2026-blind-s0.json", "gemini-3.5-flash-2026-blind-s1.json", "gemini-3.5-flash-2026-blind-s2.json" ], "max_drawdown": -0.100508, "n": 3, "round_trips": 66.666667, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.325717, "sharpe_range": [ 0.149895, 0.543446 ], "steps": 44, "total_return": 0.065383, "total_return_range": [ 0.040861, 0.095147 ], "turnover": 7.183559, "usage": { "calls": 132.0, "input_tokens": 1811529.0, "list_price_usd": 3.0528, "output_tokens": 37284.0 } }, "cutoff": "2025-01", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": [ -0.15927828394953178, 0.05865783004899907 ], "measured_on": null, "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "score": -0.04987980769230771, "verdict": "clean", "why": "every point starts after the training cutoff, so the model cannot have read how it ended" }, "model": "gemini-3.5-flash", "named": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.107915, "excess_vs_spy": -0.114539, "from": "2026-01-02", "max_drawdown": -0.10973, "sharpe": -0.111383, "sharpe_range": [ -0.199219, -0.031619 ], "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": 0.006331, "total_return_range": [ -0.006049, 0.01677 ], "volatility": 0.150747 }, "alpha_annual": -0.100986, "analysis": { "by_sector": [ { "contribution": 0.067868, "group": "Health Care", "runs": 3 }, { "contribution": 0.024899, "group": "Energy", "runs": 3 }, { "contribution": 0.015036, "group": "Industrials", "runs": 3 }, { "contribution": -0.028273, "group": "Financials", "runs": 3 }, { "contribution": -0.031242, "group": "Communication Services", "runs": 3 }, { "contribution": -0.042161, "group": "Consumer Discretionary", "runs": 3 } ], "by_trigger": [ { "contribution": 0.02671, "group": "other", "runs": 3 }, { "contribution": -0.021012, "group": "news", "runs": 3 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.070074, "alpha_t": -0.701455, "best_trade_share": 5.633532, "beta_growth": -0.148573, "beta_market": 0.78502, "beta_momentum": 0.14651, "beta_size": 0.197526, "buy_excess_5d": -0.004686, "buy_excess_5d_t": -0.920089, "capm_alpha_annual": -0.100986, "capm_alpha_t": -0.916185, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.000983, "decision_excess_5d_t": 0.255378, "decision_hit_rate_5d": 0.510432, "decision_ic": 0.001661, "decision_ic_t": 0.080118, "decisions": 113.666667, "exposure_return_annual": 0.053179, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.033531, "holdings_ic_t": -1.30612, "long_contribution": 0.005698, "news_decision_share": 0.354201, "news_entry_pnl": -2101.226807, "r2": 0.611315, "return_without_best_trade": -0.048811, "sell_avoided_5d": 0.008036, "sell_avoided_5d_t": 1.434698, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.006331 } }, "avg_exposure": 0.97502, "avg_holding_days": 37.087564, "beta": 0.782647, "excess_vs_equal_weight": -0.107915, "excess_vs_spy": -0.114539, "files": [ "gemini-3.5-flash-2026-named-s0.json", "gemini-3.5-flash-2026-named-s1.json", "gemini-3.5-flash-2026-named-s2.json" ], "max_drawdown": -0.10973, "n": 3, "round_trips": 52.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.111383, "sharpe_range": [ -0.199219, -0.031619 ], "steps": 44, "total_return": 0.006331, "total_return_range": [ -0.006049, 0.01677 ], "turnover": 5.683501, "usage": { "calls": 132.0, "input_tokens": 1785598.0, "list_price_usd": 2.9497, "output_tokens": 30149.0 } }, "named_minus_blind": -0.059052, "named_minus_blind_after_cutoff": -0.059052, "open_weights": false, "provider": "Google", "slug": "gemini-3.5-flash", "status": "clean" }, { "blind": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.122158, "excess_vs_spy": -0.128782, "from": "2026-01-02", "max_drawdown": -0.086142, "sharpe": -0.315606, "sharpe_range": [ -0.802985, 0.011518 ], "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": -0.007912, "total_return_range": [ -0.080364, 0.029882 ], "volatility": 0.080908 }, "alpha_annual": -0.070344, "analysis": { "by_sector": [ { "contribution": -0.001251, "group": "Financials", "runs": 1 }, { "contribution": -0.050165, "group": "Technology", "runs": 2 } ], "by_trigger": [ { "contribution": 0.002013, "group": "no reason", "runs": 1 }, { "contribution": -0.051797, "group": "other", "runs": 2 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": -0.088933, "alpha_t": -0.964939, "best_trade_share": 1.0, "beta_growth": 0.206957, "beta_market": 0.190979, "beta_momentum": -0.087766, "beta_size": 0.033299, "buy_excess_5d": 0.005147, "buy_excess_5d_t": null, "capm_alpha_annual": -0.070344, "capm_alpha_t": -0.598168, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": 0.005147, "decision_excess_5d_t": null, "decision_hit_rate_5d": 0.666667, "decision_ic": 0.015294, "decision_ic_t": null, "decisions": 1.0, "exposure_return_annual": 0.047096, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": -0.014737, "holdings_ic_t": -0.479348, "long_contribution": -0.03386, "news_decision_share": 0.0, "news_entry_pnl": 0.0, "r2": 0.340202, "return_without_best_trade": -0.008148, "sell_avoided_5d": null, "sell_avoided_5d_t": null, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": -0.007912 } }, "avg_exposure": 0.13998, "avg_holding_days": null, "beta": 0.265323, "excess_vs_equal_weight": -0.122158, "excess_vs_spy": -0.128782, "files": [ "llama3.2-2026-blind-s0.json", "llama3.2-2026-blind-s1.json", "llama3.2-2026-blind-s2.json" ], "max_drawdown": -0.086142, "n": 3, "round_trips": 0.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": -0.315606, "sharpe_range": [ -0.802985, 0.011518 ], "steps": 44, "total_return": -0.007912, "total_return_range": [ -0.080364, 0.029882 ], "turnover": 0.098804, "usage": { "calls": 132.0, "input_tokens": 1390800.0, "list_price_usd": 0.0, "output_tokens": 7436.0 } }, "cutoff": "2023-12", "memory": { "beyond_control": null, "beyond_control_ci95": null, "ci95": [ -0.17036720399366648, 0.08513631714574144 ], "control_period": null, "missing_answers": { "blind": 0, "named": 0, "recall": 0 }, "p_value": 0.7461269365317341, "paired": "run", "probe": "riskaware.research.memory.memory_probe: 100 named/blind pairs and 50 recall questions per period, seed 0, temperature 0, with the model's cutoff; score = AUC gap named minus blind on the points before the cutoff", "recall_ci95": [ -0.30084930201742666, 0.27386066618393234 ], "recall_direction_accuracy": 0.46, "recall_majority_baseline": 0.56, "recall_score": 0.005496005867545086, "recall_verdict": "no sign", "recall_windows": 50, "score": -0.04947916666666674, "scored_points": "after", "verdict": "clean", "why": "every point starts after the training cutoff, so the model cannot have read how it ended" }, "model": "llama3.2", "named": { "after_cutoff": { "days": 216, "equal_weight_sharpe": 0.848303, "equal_weight_total_return": 0.114246, "excess_vs_equal_weight": -0.041527, "excess_vs_spy": -0.048151, "from": "2026-01-02", "max_drawdown": -0.062236, "sharpe": 0.560731, "sharpe_range": [ 0.284338, 1.069882 ], "spy_sharpe": 0.76515, "spy_total_return": 0.12087, "to": "2026-11-11", "total_return": 0.072719, "total_return_range": [ 0.049143, 0.105992 ], "volatility": 0.100575 }, "alpha_annual": 0.019453, "analysis": { "by_sector": [ { "contribution": 0.067106, "group": "Technology", "runs": 3 }, { "contribution": -0.010441, "group": "Financials", "runs": 2 }, { "contribution": -0.011226, "group": "Consumer Discretionary", "runs": 2 } ], "by_trigger": [ { "contribution": 0.070026, "group": "other", "runs": 3 }, { "contribution": -0.026047, "group": "no reason", "runs": 2 } ], "findings": null, "runs": 3, "summary": { "alpha_annual": 0.015481, "alpha_t": 0.189495, "best_trade_share": 1.54129, "beta_growth": -0.192278, "beta_market": 0.394895, "beta_momentum": -0.000147, "beta_size": -0.055073, "buy_excess_5d": -0.025023, "buy_excess_5d_t": -1.084856, "capm_alpha_annual": 0.019453, "capm_alpha_t": 0.25041, "concentrated": true, "cover_excess_5d": null, "cover_excess_5d_t": null, "decision_excess_5d": -0.011496, "decision_excess_5d_t": -0.65487, "decision_hit_rate_5d": 0.443723, "decision_ic": -0.065946, "decision_ic_t": -0.816845, "decisions": 6.666667, "exposure_return_annual": 0.03596, "fees_contribution": 0.0, "forced_fills": 0.0, "holdings_ic": 0.006944, "holdings_ic_t": 0.386163, "long_contribution": 0.052661, "news_decision_share": 0.214286, "news_entry_pnl": 0.0, "r2": 0.225861, "return_without_best_trade": 0.003162, "sell_avoided_5d": -0.000377, "sell_avoided_5d_t": 3.712462, "short_avoided_5d": null, "short_avoided_5d_t": null, "short_contribution": 0.0, "spy_return": 0.12087, "total_return": 0.072719 } }, "avg_exposure": 0.356629, "avg_holding_days": 137.139895, "beta": 0.297719, "excess_vs_equal_weight": -0.041527, "excess_vs_spy": -0.048151, "files": [ "llama3.2-2026-named-s0.json", "llama3.2-2026-named-s1.json", "llama3.2-2026-named-s2.json" ], "max_drawdown": -0.062236, "n": 3, "round_trips": 2.333333, "run_dates": [ "2026-11-14" ], "seeds": [ 0, 1, 2 ], "sharpe": 0.560731, "sharpe_range": [ 0.284338, 1.069882 ], "steps": 44, "total_return": 0.072719, "total_return_range": [ 0.049143, 0.105992 ], "turnover": 0.592731, "usage": { "calls": 132.0, "input_tokens": 1385878.0, "list_price_usd": 0.0, "output_tokens": 11692.0 } }, "named_minus_blind": 0.080631, "named_minus_blind_after_cutoff": 0.080631, "open_weights": true, "provider": "Meta", "slug": "llama3.2", "status": "clean" }, { "blind": { "after_cutoff": { "days": 94, "equal_weight_sharpe": 1.014392, "equal_weight_total_return": 0.057091, "excess_vs_equal_weight": -0.063181, "excess_vs_spy": -0.026078, "from": "2026-06-30", "max_drawdown": -0.030665, "sharpe": -0.994655, "sharpe_range": null, "spy_sharpe": 0.188135, "spy_total_return": 0.019988, "to": "2026-11-11", "total_return": -0.00609, "total_return_range": null, "volatility": 0.051981 }, "alpha_annual": -0.033815, "analysis": { "by_sector": [ { "contribution": 0.031674, "group": "Financials", "runs": 1 }, { "contribution": 0.021134, "group": "Industrials", "runs": 1 }, { "contribution": -0.00456, "group": "Health Care", "runs": 1 }, { "contribution": -0.006292, "group": "Communication Services", "runs": 1 }, { "contribution": -0.039111, "group": "Technology", "runs": 1 } ], "by_trigger": [ { "contribution": 0.020345, "group": "news", "runs": 1 }, { "contribution": -0.017502, "group": "other", "runs": 1 } ], "findings": [ "Buys: over the five trading days from the fill, the stocks it bought trailed SPY by 0.30% on average (33 buys, 52% of them ahead of SPY; t = -0.2, within noise).", "Sells: over the five trading days from the fill, the stocks it sold beat SPY by 1.96%, so selling gave that up (33 sells; t = -1.8, within noise).", "Decision IC: on the 30 days it traded, what it bought ranked against what it sold and left alone by a mean rank IC of -0.036 with the next five days' returns (IC IR -0.20, t = -1.1, within noise).", "Holdings IC: the weights it held after each of 39 decision days had a mean rank IC of -0.007 with the next five days' returns (t = -0.4, within noise).", "Factor exposures: market 0.23 (t 5.3), size 0.14 (t 3.5), growth 0.04 (t 0.5), momentum 0.01 (t 0.6); together they explain 29% of the daily variation (R²).", "Of the -0.5% a year it earned over T-bills (arithmetic mean): market +2.5%, size +0.5%, momentum -0.3%, growth +0.1%; the alpha left is -3.2% a year (t = -0.5, within noise).", "Verdict: the exposures account for +2.7% a year of it, and the -3.2% a year left over is not significant: one period of daily returns cannot tell it from luck.", "P&L: +$2,482 in all, +$284 from 23 positions and +$2,197 of interest on idle cash.", "By holding period: 61+ days made the most (+$6,879, 4 positions); 6-10 days the least (-$5,914, 5 positions).", "By sector: Financials made the most (+$3,167, 1 position); Technology the least (-$3,911, 13 positions).", "By stock: AMD made the most (+$3,839, 2 positions); QCOM the least (-$3,413, 2 positions).", "By market regime of the previous close: risk-on 96 days +1.8% (SPY +2.0%); neutral 18 days -1.5% (SPY +4.5%); uncertain 97 days +1.7% (SPY +4.8%); risk-off 6 days +0.4% (SPY +0.3%).", "News: 15 positions opened on news made +$2,035, the other 8 -$1,750; 6 positions closed on news made -$2,429; over the five days from the fill, news-triggered buys trailed SPY by 0.99% (t -0.7).", "Concentration: the profit rests on the best two positions, which made 2721% of the positions' net P&L and 66% of all their gains; the best was AMD bought 2026-04-23 and sold 2026-05-29, +$4,572. Had the best position made nothing, the run would have returned -2.1% instead of +2.5%." ], "runs": 1, "summary": { "alpha_annual": -0.03171, "alpha_t": -0.492925, "best_trade_share": 16.075251, "beta_growth": 0.040444, "beta_market": 0.232713, "beta_momentum": 0.014859, "beta_size": 0.138652, "buy_excess_5d": -0.002977, "buy_excess_5d_t": -0.249545, "capm_alpha_annual": -0.033815, "capm_alpha_t": -0.519571, "concentrated": true, "decision_excess_5d": -0.011278, "decision_excess_5d_t": -1.394412, "decision_hit_rate_5d": 0.424242, "decision_ic": -0.03604, "decision_ic_t": -1.092237, "decisions": 66.0, "exposure_return_annual": 0.027135, "holdings_ic": -0.007222, "holdings_ic_t": -0.384303, "news_decision_share": 0.432836, "news_entry_pnl": 2034.538739, "r2": 0.289337, "return_without_best_trade": -0.020897, "sell_avoided_5d": -0.019579, "sell_avoided_5d_t": -1.844518, "spy_return": 0.12087, "total_return": 0.02437 } }, "avg_exposure": 0.306503, "avg_holding_days": 22.552645, "beta": 0.27214, "excess_vs_equal_weight": -0.089876, "excess_vs_spy": -0.0965, "files": [ "claude-haiku-5.5-2026-blind-s0.json" ], "max_drawdown": -0.056366, "n": 1, "round_trips": 34.0, "run_dates": [ "2026-11-14" ], "seeds": [ 0 ], "sharpe": -0.059767, "sharpe_range": null, "steps": 44, "total_return": 0.02437, "total_return_range": null, "turnover": 2.409155, "usage": { "calls": 44.0, "input_tokens": 694182.0, "list_price_usd": 0.0754, "output_tokens": 12036.0 } }, "cutoff": "2026-06", "memory": null, "model": "claude-haiku-5.5", "open_weights": false, "provider": "Anthropic", "slug": "claude-haiku-5.5", "status": "partly" } ], "may_remember": [], "title": "2026 to date: a rising market with a spring sell-off", "window": { "end": "2026-11-11", "start": "2026-01-02" } } }, "seeds": [ 0, 1, 2 ], "settings": { "cash": 100000.0, "commission_per_share": 0.0, "costs": "realistic", "decide_every": 5, "execution": "plain", "max_news": 3, "max_positions": 10, "mode": "long_only", "news": true, "position_size": null, "slippage_bps": 5.0, "universe": "megacap" } }