# InelidaMarketScan — Scientific Analysis of Financial Markets with Machine Learning

> **Quantitative Finance Research: ICT, Ichimoku, XGBoost & Stochastic Calculus**
> By **Didier Vally / Reuniware Systems**
> Version 1.0.6 — 2026-07-19
> 📦 PyPI: [https://pypi.org/project/inelida-marketscan](https://pypi.org/project/inelida-marketscan) | GitHub: [https://github.com/reuniware/InelidaMarketScanner](https://github.com/reuniware/InelidaMarketScanner)

---

## Abstract

This document presents the results of 6 months of quantitative finance research applied to algorithmic trading. The InelidaMarketScan framework combines ICT (Inner Circle Trader) technical analysis and multi-timeframe Ichimoku with an XGBoost Machine Learning pipeline (123 features) and stochastic calculus functions (Ornstein-Uhlenbeck, Hurst, GARCH, Monte Carlo, Kelly). Model v18 achieves 64.0% cross-validation accuracy on 1,210 historical trades. Analysis of 27 successive models reveals that Ichimoku features (Kijun/Tenkan distances, T/K crosses) are more predictive than ICT patterns (FVG, Order Blocks). Capital simulation with ML% ≥ 50% filtering shows a profit factor of 2.27 on in-sample data, while out-of-sample validation identifies a ceiling at 1.173. This document serves as a comprehensive scientific reference on the intersection of ICT, Ichimoku, and Machine Learning.

---

## 1. Introduction

### 1.1 Context and Problem Statement

Algorithmic trading relies on identifying recurring patterns in financial time series. ICT (Inner Circle Trader) concepts developed by Michael Huddleston describe institutional market behavior: liquidity sweeps, Fair Value Gaps, Order Blocks, Market Structure Shifts. In parallel, the Ichimoku Kinko Hyo indicator provides a multi-timeframe reading of market equilibrium. This project explores the fusion of these two paradigms through an automated scanner coupled with an XGBoost model.

### 1.2 Research Objectives

1. Develop an automated scanner combining ICT and Ichimoku across 8 timeframes (M5 to MN). 2. Build an ML pipeline with 123 named features extracted from the scanner. 3. Validate predictive power via historical backtest (2025-2026) and out-of-sample testing. 4. Integrate quantitative finance functions (OU, Hurst, GARCH, Monte Carlo, Kelly). 5. Achieve a profit factor > 1.5 under real conditions.

## 2. ICT Theoretical Foundations

ICT concepts model the behavior of financial institutions (banks, hedge funds) that manipulate price levels to accumulate positions. The following 8 concepts are automatically detected by the scanner:

| ICT Concept | Description and Detection |
|:---|:---|
| **Liquidity Sweeps** | Brief break of a key level (Asian High/Low, London High/Low) to trigger stops before reversal. Detected on H1 with reversal confirmation. |
| **Fair Value Gaps (FVG)** | Gap between 3 consecutive candles where price did not trade. Detected on 6 TFs (M5 to D1) with mitigation percentage. |
| **Order Blocks (OB)** | Last candle before a directional impulse — zone where institutions placed orders. Detected via ATR(14) on 6 TFs. |
| **Breaker Blocks (BRK)** | Broken Order Block that reverses into opposite support/resistance. Derived from existing OBs, zero additional MT5 calls. |
| **Market Structure Shift (MSS/BOS)** | Change in market structure: Break of Structure and Change of Character on 6 TFs. |
| **Volume HVN/LVN** | High Volume Node and Low Volume Node based on MT5 tick volume. Detected on 6 TFs. |
| **Reversal Pipeline** | Complete sweep → reversal → confirmation chain with post-sweep bar counter. |
| **Fibonacci Extensions** | +1.618 to +4.0 extensions based on Asian range for TP targets. |

## 3. Multi-TF Ichimoku Architecture

The Diamond scanner analyzes 8 timeframes simultaneously. For each TF, 5 Ichimoku components are computed:

| Timeframe | Tenkan | Kijun | Kumo | T/K Cross | Flat History |
|:---|:---:|:---:|:---:|:---:|:---:|
| **M5** | X | X | - | X | - |
| **M15** | X | X | - | X | - |
| **M30** | X | X | - | X | - |
| **H1** | X | X | X | X | X |
| **H4** | X | X | X | X | X |
| **D1** | X | X | X | X | X |
| **W1** | X | X | X | X | X |
| **MN** | X | X | X | X | X |

## 4. Diamond Scanner Architecture

The Diamond Scanner aggregates 109 scoring criteria across 10 categories:

| Category | Criteria | Examples |
|:---|---:|:---|
| Ichimoku H1-D1 | 24 | Kijun/Tenkan distances, T/K crosses, Kumo positions, Flat bars |
| Ichimoku W1-MN | 8 | Kijun W1/MN, Kumo W1/MN, Cloud color, Double memory |
| FVG (6 TFs) | 12 | FVG bear/bull M5 to D1, mitigation %, time priority |
| Order Blocks (6 TFs) | 12 | OB bear/bull M5 to D1, position ABOVE/BELOW/AT |
| MSS/BOS (6 TFs) | 18 | Regime, BOS, CHoCH on M5 to D1 |
| Volume HVN/LVN (6 TFs) | 12 | Volume spike, HVN, LVN on M5 to D1 |
| Breaker Blocks (6 TFs) | 12 | BRK bear/bull M5 to D1 |
| Asian/London Sweeps | 4 | AH/AL sweep, London H/L sweep |
| Reversal Pipeline | 5 | Sweep→reversal→confirmation, post-sweep FVG, retest |
| M30/M15/M5 Kijun | 2 | Kijun M30/M15/M5 directional cross |

**Maximum score:** 109 criteria (107 without MN)

## 5. Machine Learning Pipeline

The ML pipeline transforms scanner results into win probability through 3 stages:

| Stage | Tool | Details |
|:---:|:---|:---|
| 1 | `ml_labeler.py` | Feature extraction from Diamond scan (DB / manual CSV / backtest). Generates labeled CSV with outcome column (1=won, 0=lost). |
| 2 | `ml_trainer.py` | XGBoost training with 5-fold cross-validation, scale_pos_weight for imbalanced classes, hyperparameter grid search. |
| 3 | `ml_predictor.py` | Real-time inference: extract_features() → predict_proba() → ML% displayed in live scan. Simplified MLPredictor: 1 joblib.load instead of 5. |

## 6. XGBoost Model v18 — Configuration

Model v18 is the active production model. Trained on 1,210 historical trades with 123 named features (including GARCH and stochastic).

| Hyperparameter | Value |
|:---|:---|
| `max_depth` | 6 |
| `learning_rate` | 0.03 |
| `n_estimators` | 200 |
| `subsample` | 0.8 |
| `min_child_weight` | 3 |
| `gamma` | 0.1 |
| `reg_alpha` | 0.5 (L1) |
| `reg_lambda` | 1.0 (L2) |
| `scale_pos_weight` | auto (based on win/loss ratio) |

**123 named features** include: scores/bias (6), Kijun 6 TFs (18), Tenkan 6 TFs (18), T/K crosses (6), Kumo (6), M5/M15/M30 crosses (12), Flat bars (2), Active FVGs (1), Order Blocks (6), Sweeps (5), Volume (6), MSS/BOS (6), Breaker Blocks (6), Compression (1), Stochastic (8: OU, Hurst, Monte Carlo, GARCH), Temporal (3), Asset type (3).

## 7. Model History

27 XGBoost models have been trained and evaluated since the project's inception. Here is the history of major versions:

| Model | Data | Features | Trades | CV | OOS PF | Status |
|:---|:---|---:|---:|:---:|:---:|
| v1 | 6 majors | 36 | 852 | 61.0% | — | Archived |
| v4 | 13 symbols | 111 | 1,210 | 62.9% | 1.26* | In-sample (leakage) |
| v7 | 13 symbols | 25 best | 2,923 | 60.1% | 1.173 | 🥇 Best unified OOS |
| v13 | 4 models/asset | 25 | 2,923 | 53-64% | 4.000 | 🏆 Per-asset (15 trades) |
| v15 | 2025-2026 | 25 + OU/Hurst | 5,009 | 59.8% | — | Stochastic integrated |
| v18 | 13 symbols | 123 (GARCH) | 1,210 | 64.0% | 0.503 F1 | [X] Production active |

## 8. Feature Importance (v4)

Feature importance analysis reveals which criteria carry the most weight in model decisions:

| Rank | Feature | Gain | Insight |
|:---:|:---|:---:|:---|
| 1 | `is_dxy` | 14.2 | The dollar follows a radically different logic from forex pairs |
| 2 | `rr` | 10.0 | Theoretical risk/reward ratio predicts actual outcome |
| 3 | `tkx_h1` | 9.2 | H1 Tenkan/Kijun cross is the single best predictor |
| 4 | `kj_h1` | 6.9 | Distance to H1 Kijun captures short-term equilibrium |
| 5 | `sl` | 6.4 | Stop-loss level is as important as take-profit |
| 6 | `day_wednesday` | 6.4 | Wednesdays (FOMC) show statistically distinct behavior |
| 7 | `d_kj_h4` | 6.3 | Distance to H4 Kijun indicates underlying trend |
| 8 | `d_kj_d1` | 6.0 | Distance to D1 Kijun anchors macro context |
| 9 | `tkx_h4` | 6.0 | H4 T/K cross confirms session trend |
| 10 | `is_forex` | 5.9 | FX pairs have distinct behavior vs indices/metals |

## 9. Out-of-Sample Validation

OOS validation (training January-June 2026, testing July 2026) is the decisive test. Model v4 shows an in-sample PF of 2.27 at ML% ≥ 50% threshold, but this includes data leakage (the model saw 80% of test data during training). The true OOS PF of v7 is 1.173 — a statistically significant but modest edge. Model v13 (per-asset) achieves PF=4.000 on 15 trades, but the sample is too small to be reliable. Conclusion: ML provides a real edge (PF > 1.0), but the current ceiling is ~1.17 with a unified architecture. Improvement will come from more data and potentially asset-specialized models.

## 10. Capital Simulation — ML% Filtering

Simulation on 1,210 historical trades with initial capital of €10,000 and 1% risk per trade:

| ML% Threshold | Trades | Win Rate | PF (RR) | P&L (€10k) |
|:---|---:|---:|:---:|---:|
| No filter | 1,210 | 37.6% | 1.02 | +€1,442 |
| ML% ≥ 45% | 661 | 54.3% | 1.91 🔥 | +€27,542 |
| ML% ≥ 50% | 504 | 59.7% | 2.27 🔥🔥 | +€25,842 |
| ML% ≥ 55% | 327 | 67.3% | 2.80 | +€19,242 |
| ML% ≥ 60% | 170 | 78.8% | 3.59 | +€9,340 |
| ML% ≥ 70% | 71 | 88.7% | 1.64 | +€514 |

## 11. Quantitative Finance & Stochastic Calculus

6 quantitative finance functions have been integrated into the scanner and ML pipeline. These functions compute risk, volatility, and market regime metrics:

| Function | Purpose | ML Feature | v18 Rank |
|:---|:---|:---|:---:|
| **Ornstein-Uhlenbeck** | Detects price overextension from long-term mean. Score 0-100%: >80% = likely reversal. | `ou_score` | Not tested |
| **Hurst Exponent** | Classifies market regime: H < 0.45 = RANGE (mean-reversion), H > 0.55 = TREND (momentum). | `hurst_h` | Not tested |
| **GARCH(1,1)** | Models conditional volatility. 4 features: persistence, long_run_vol, half_life, forecast_vol_1. | `garch_persistence` | #14 |
| **Monte Carlo** | Simulates 10,000 price trajectories. Computes P(SL hit) and P(TP hit). | `mc_p_sl, mc_p_tp` | Not tested |
| **Kelly Criterion** | Optimal position sizing by asset type. Gain=0.00 (redundant with already-optimized SL/TP). | `kelly_full` | #124 |
| **FTMO Ruin** | Ruin probability for a prop firm account with 10% drawdown over 100 simulated trades. | `ftmo_ruin_prob` | Not tested |

## 12. Kelly Criterion by Asset Type

The Kelly Criterion uses real win rates by asset type (from backtest) rather than a uniform win_rate=0.40. Results:

| Asset Type | Win Rate | Kelly Full | Kelly Half |
|:---|---:|:---:|:---:|
| **DXY** | 73% | 46.0% | 23.0% |
| **XAUUSD** | 42% | 13.0% | 6.5% |
| **Forex** | 31% | 3.5% | 1.8% |
| **Indices** | 38% | 7.0% | 3.5% |
| **Crypto** | 35% | 2.5% | 1.3% |

## 13. Simplified MLPredictor (2026-07-19)

The MLPredictor previously loaded the same model 5 times (1× per prediction method). The simplification of 2026-07-19 reduced this to a single `joblib.load()`. Result: 30 lines of code removed, loading time divided by 4, and the redundant `_ASSET_MODEL_MAP` eliminated. Model v18 is now the single unified model for all asset types.

## 14. Bug Fixes — 2026-07-19

A thorough code review on 2026-07-19 identified and fixed 4 bugs:

| # | Severity | File | Description |
|:---:|:---|:---|:---|
| 1 | CRITICAL | `diamond_scanner.py` | Dead code after `return` in `ml_available` — symbols were never activated. `initialize()` did not set `_initialized = True`. |
| 2 | HIGH | `stochastic_analytics.py` | GARCH: if no valid (α,β) pair found, silent fallback with invented parameters. Added `best_ll <= -1e100` check. |
| 3 | MEDIUM | `stochastic_analytics.py` | Monte Carlo: log-return volatility fallback was computed but never used when OU sigma was zero. |
| 4 | LOW | `stochastic_analytics.py` | FTMO: `pnl_avg`/`roi_avg` outside try block — risk of NameError if code added between computation and return. |

## 15. Final Decision Matrix

The decision matrix combines Diamond score, ML%, safety guards, and macro context:

| Condition | Action | Rationale |
|:---|:---:|:---|
| ML% ≥ 60% + TKx H1 aligned + no HIGH NEWS | 🟢 GO | Triple confirmation: ML, Ichimoku, macro |
| ML% ≥ 60% + TKx H1 conflict | 🟡 WAIT | ML is contradicted by T/K H1 cross |
| ML% 50-60% + TKx H1 aligned | 🟡 CAUTION | Moderate ML edge, needs additional confirmation |
| ML% < 50% | 🔴 NO GO | Model has no statistical edge |
| HIGH NEWS DAY + BULL | 🔴 WAIT | Macro days (≥2 major events) break correlations |
| Wednesday (FOMC day) + Forex | 🟡 High Risk | Wednesdays show statistically distinct behavior |

## 16. Conclusion and Perspectives

The InelidaMarketScan project demonstrates that an ML pipeline trained on Ichimoku features can generate a statistically significant edge (OOS PF = 1.173). The 4 critical bugs fixed on 2026-07-19 strengthen live scanner reliability. The 27 successive models revealed key insights: Ichimoku features (Kijun/Tenkan distances) are more predictive than ICT patterns, asset type (is_dxy, is_forex) is a major predictor, and stochastic calculus functions (GARCH, OU, Hurst) provide complementary volatility signal. The next performance leap will require a larger dataset (2024-2026, >5,000 trades/asset) and potentially deep learning architectures (LSTM, Transformer) to capture temporal dependencies that XGBoost does not model.

## References

- Huddleston, M. — Inner Circle Trader (ICT) Concepts, 2022-2024
- Chen, T. & Guestrin, C. — XGBoost: A Scalable Tree Boosting System, KDD 2016
- Bollerslev, T. — Generalized Autoregressive Conditional Heteroskedasticity, Journal of Econometrics, 1986
- Hurst, H.E. — Long-term Storage Capacity of Reservoirs, Trans. American Society of Civil Engineers, 1951
- Uhlenbeck, G.E. & Ornstein, L.S. — On the Theory of Brownian Motion, Physical Review, 1930
- Kelly, J.L. — A New Interpretation of Information Rate, Bell System Technical Journal, 1956
- Hosoda, G. — Ichimoku Kinko Hyo, 1968

---

> **Quantitative finance research work. Reproduction permitted with citation.**

## 🌐 Community

Join the **Trading Pro** Discord server to connect with the community: https://discord.gg/u9JApXqhXy

---

> **Didier Vally / Reuniware Systems** — InelidaMarketScan v1.0.6
> 2026-07-19