# Cross-Sectional Momentum Crypto Strategy

Mid-frequency cross-sectional momentum strategy backtested across 7 cryptocurrencies on 1H OHLCV data. Built from scratch — data pipeline, feature engineering, signal generation, backtest engine, and performance reporting.

---

## Strategy Overview

Every hour, all 7 coins are ranked by relative momentum (each coin's return vs the cross-sectional average). The strategy goes **long the top 2** ranked coins and **short the bottom 2** — but only when three independent confirmation filters simultaneously agree. Position sizing adapts dynamically to a daily market regime score that detects whether the broader crypto market is in a bull or bear phase.

The core insight: crypto prices are driven by collective sentiment and capital momentum rather than fundamentals. This creates persistent relative strength effects across coins that can be systematically captured.

---

## Universe

| Coin | Exchange |
|---|---|
| AVAXUSDT | Binance |
| BNBUSDT | Binance |
| BTCUSDT | Binance |
| DOGEUSDT | Binance |
| ETHUSDT | Binance |
| LINKUSDT | Binance |
| SOLUSDT | Binance |

**Common start date:** 2020-09-22 (determined by AVAXUSDT listing date)  
**Timeframes used:** 1d, 1h, 15m, 5m

---

## Performance Results

| Period | Dates | Return | Final ($10k) | Sharpe | Max DD | Win Rate | Trades |
|---|---|---|---|---|---|---|---|
| Full Dataset | Sep'20 – Dec'24 | +3,869% | $396,851 | 1.55 | -62.4% | 49.8% | 8,905 |
| Train | Sep'20 – Sep'23 | +726% | $82,553 | 1.29 | -62.4% | 49.7% | 6,153 |
| Validation | Sep'23 – May'24 | +98% | $19,768 | 2.10 | -20.4% | 49.6% | 1,383 |
| **Test** | **May'24 – Dec'24** | **+143%** | **$24,318** | **2.81** | **-34.2%** | **50.6%** | **1,369** |

> Sharpe ratio improves from Train (1.29) → Validation (2.10) → Test (2.81). Strategy is profitable in all three periods including the completely unseen test set.

> All returns are gross of transaction costs.

---

## Signal Architecture

### Three-Layer Entry Filter
All three layers must agree simultaneously — no signal fires if any layer fails.

| Layer | Condition | Purpose |
|---|---|---|
| L1 — Momentum Rank | Long: rank >= 6 of 7 / Short: rank <= 2 of 7 | Only top/bottom coins by relative momentum qualify |
| L2 — ATR Expansion | ATR(14) > ATR baseline(20) | Ensures market has energy to sustain directional moves. Avoids ranging markets |
| L3 — Volume Confirmation | volume_ratio > 1.0 | Confirms genuine capital participation behind the move |

### Regime Score
Daily score (0 to 1) = proportion of 7 coins above their 20-day moving average.
- **Above 0.5** → bull regime → long positions scaled up, short positions scaled down
- **Below 0.5** → bear regime → short positions scaled up, long positions scaled down
- **Distance from 0.5** → total capital deployed (near 0.5 = uncertainty = reduce exposure)

### Exit Framework

| Exit | Status | Logic |
|---|---|---|
| Signal Expiry | ✅ Implemented | Position closes when 3-layer filter no longer fires |
| Momentum Deterioration | ✅ Implemented | Exit when both return_6 and return_12 simultaneously turn negative |
| ATR Trailing Stop | 🔲 Designed | Exit if price moves against position by X × ATR |
| Time Cap | 🔲 Designed | Force exit after 12 candles maximum |

---

## Feature Engineering

| Feature | Timeframe | Description |
|---|---|---|
| upper_wick_ratio | 1h, 15m, 5m | Selling pressure within candle |
| lower_wick_ratio | 1h, 15m, 5m | Buying pressure within candle |
| body_ratio | 1h, 15m, 5m | Conviction of the candle move |
| close_location_value | 1h, 15m, 5m | Where price closed within candle range |
| ATR | 1h | 14-period average true range |
| ATR_baseline | 1h | 20-period rolling mean of ATR |
| volume_ratio | 1h | Current volume vs 20-period rolling mean |
| return_6, return_12, return_24 | 1h | Rolling returns for momentum and exit signals |
| relative_momentum | 1h | Each coin's return minus cross-sectional mean |
| cross_rank | 1h | Ordinal rank of coin by relative momentum (1–7) |
| dispersion | 1h | Standard deviation of returns across all 7 coins |
| regime_score | 1d | Proportion of coins above 20-day MA |

> All rolling features use `shift(1)` before calculation to prevent lookahead bias.

---

## Data Format

Parquet files with MultiIndex — level 0 is `symbol`, level 1 is `open_time`.

```
Columns: open, high, low, close, volume
Files:   1d.parquet, 1h.parquet, 15m.parquet, 5m.parquet
```

Expected directory structure:
```
project/
├── step_1.py
├── 1d.parquet
├── 1h.parquet
├── 15m.parquet
└── 5m.parquet
```

---

## How to Run

**1. Install dependencies**
```bash
pip install pandas numpy matplotlib pyarrow reportlab
```

**2. Place parquet files in the same directory as the script**

**3. Run the strategy**
```bash
python step_1.py
```

**4. Output**
- Prints performance metrics for Train, Validation, Test, and Full Dataset
- Displays equity curves and drawdown charts for all periods

---

## Tech Stack

| Library | Usage |
|---|---|
| pandas | Data loading, feature engineering, signal generation |
| numpy | Numerical calculations, weight construction |
| matplotlib | Equity curves and drawdown charts |
| pyarrow | Parquet file reading |
| reportlab | PDF report generation |

---

## Key Design Decisions

**Fixed universe** — all 7 coins from common start date. Cross-sectional rankings are inconsistent when universe size changes. Fixed universe is the professional standard.

**Absolute sum weight normalization** — prevents long and short weights from cancelling each other during normalization. Ensures total gross exposure never exceeds 1.0.

**Returns winsorized at ±10%** — AVAXUSDT produced a 474% single-candle return at listing due to near-zero liquidity. This is a data artifact not a tradeable return and is capped accordingly.

**Regime neutral fill** — NaN regime values during warmup filled with 0.5 (neutral). Ensures no position is taken with undefined regime rather than corrupting the equity curve with NaN propagation.

---

## Limitations

- Transaction costs not modelled — gross Sharpe reported above
- Execution assumes fills at open price — real slippage would reduce returns
- ATR trailing stop and time cap exits are designed but not yet implemented
- 15m and 5m features computed but not yet integrated into execution timing

---

## License

MIT