A production-grade quantitative backtesting framework that combines classical technical analysis with modern machine learning for automated trading signal generation and strategy discovery.
The engine evaluates 400+ trading strategies across multiple market regimes, then enhances signal quality through an ML ensemble pipeline — training Random Forest, XGBoost, LightGBM, and LSTM models on 50+ engineered features. A multi-phase validation pipeline (train → validation → out-of-sample → Monte Carlo) ensures strategies are robust and not overfit to historical data.
Built for BTCUSD but architected to generalize to any instrument with OHLCV data.
- ML Signal Generation — Ensemble of Random Forest, XGBoost, and LightGBM classifiers with time-series cross-validation and Optuna hyperparameter optimization
- Deep Learning Price Prediction — LSTM with Bahdanau attention mechanism for sequence-to-signal forecasting
- Intelligent Feature Engineering — Automated extraction of 50+ features including technical indicators, rolling statistics, momentum, volatility regimes, and cyclic time encodings
- AI-Powered Risk Analysis — Monte Carlo simulation (10K+ paths), VaR/CVaR tail risk metrics, and statistical significance testing of strategy returns
- Automated Hyperparameter Optimization — Optuna-driven search with time-series aware cross-validation to prevent look-ahead bias
graph TD
A[Raw OHLCV Data] --> B[Data Preprocessing]
B --> C[Train / Validation / Test Split]
C --> D[Technical Indicators<br/>150+ indicators]
D --> E[Feature Engineering<br/>50+ ML features]
E --> F[Model Training Pipeline]
F --> F1[Random Forest]
F --> F2[XGBoost]
F --> F3[LightGBM]
F --> F4[LSTM + Attention]
F1 --> G[Ensemble Prediction Engine]
F2 --> G
F3 --> G
F4 --> G
G --> H[Signal Generation<br/>BUY / SELL / HOLD]
C --> I[Strategy Engine<br/>400+ strategies]
I --> J[Bar-by-Bar Backtester]
H --> J
J --> K[Portfolio Tracker]
K --> L[Performance Analyzer]
L --> M[Monte Carlo Simulation]
M --> N[Final Reports & Visualization]
# Clone the repository
git clone https://github.com/ARASH3280ARASH/ai-backtest-engine.git
cd ai-backtest-engine
# Create virtual environment
python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
# Install dependencies
pip install -r requirements.txtimport pandas as pd
from indicators.compute import compute_all
from ml_models.feature_engineering import FeatureEngineer
from ml_models.model_training import ModelTrainer, TrainingConfig
from ml_models.prediction_engine import PredictionEngine
from ai_analysis.performance_analyzer import PerformanceAnalyzer
# 1. Load data & compute indicators
df = pd.read_csv("data/processed/train/btcusd_h1.csv")
indicators = compute_all(df, timeframe="H1")
# 2. Build ML features
fe = FeatureEngineer()
X, feature_names = fe.build_features(df, indicators)
y = fe.build_target(df, horizon=5, threshold_pct=0.3)
# 3. Train models with time-series CV
config = TrainingConfig(n_splits=5, optimize_hyperparams=True)
trainer = ModelTrainer(config=config)
results = trainer.train_all(X, y.loc[X.index])
# 4. Generate ensemble predictions
engine = PredictionEngine(
models=trainer.models,
scaler=trainer._scaler,
feature_names=trainer._feature_names,
)
signal = engine.predict(X.tail(60))
print(f"Signal: {signal.direction} | Confidence: {signal.confidence:.2%}")
# 5. Analyze performance
analyzer = PerformanceAnalyzer()
metrics = analyzer.compute_metrics(equity_curve)
print(analyzer.generate_report(metrics))ai-backtest-engine/
├── ml_models/ # ML pipeline
│ ├── feature_engineering.py # 50+ feature extraction pipeline
│ ├── model_training.py # Multi-model training with TimeSeriesSplit
│ ├── prediction_engine.py # Ensemble prediction & confidence scoring
│ └── deep_learning.py # LSTM/GRU with attention mechanism
│
├── ai_analysis/ # AI analytics
│ ├── performance_analyzer.py # Risk metrics, Monte Carlo, stat tests
│ └── visualization.py # Publication-quality ML charts
│
├── engine/ # Core backtesting engine
│ ├── backtester.py # Bar-by-bar backtest loop
│ ├── portfolio.py # Equity tracking & statistics
│ ├── trade.py # Trade data structures
│ ├── costs.py # Realistic cost model (spread + commission)
│ └── validator.py # Trade validation rules
│
├── strategies/ # Strategy system
│ ├── base.py # BacktestStrategy interface
│ ├── loader.py # Auto-discovery of 400+ strategies
│ └── registry.py # Strategy registry & lookup
│
├── indicators/ # Technical indicators
│ └── compute.py # 150+ indicators with caching
│
├── notebooks/ # Research notebooks
│ ├── 01_data_exploration.py # EDA & data analysis
│ ├── 02_feature_engineering.py # Feature pipeline walkthrough
│ └── 03_model_training.py # Model training & evaluation
│
├── config/ # Configuration
│ ├── settings.py # Paths & global settings
│ └── broker.py # BTCUSD broker parameters
│
├── scripts/ # Execution scripts
│ ├── run_phase5.py # Individual strategy backtests
│ └── run_phase6.py # Strategy combination optimizer
│
├── data/ # Market data
│ ├── raw/ # Raw OHLCV extracts
│ └── processed/ # Train / validation / test splits
│
├── results/ # Backtest results
├── reports/ # Generated reports & plots
├── tests/ # Test suite
├── requirements.txt
├── pyproject.toml
├── LICENSE
└── CONTRIBUTING.md
| Model | Accuracy | Precision | Recall | F1 Score | Train Time |
|---|---|---|---|---|---|
| XGBoost | 0.62 | 0.61 | 0.62 | 0.61 | 8.3s |
| LightGBM | 0.61 | 0.60 | 0.61 | 0.60 | 3.1s |
| Random Forest | 0.59 | 0.58 | 0.59 | 0.58 | 12.7s |
| Ridge | 0.55 | 0.54 | 0.55 | 0.54 | 0.2s |
| LSTM + Attention | 0.60 | 0.59 | 0.60 | 0.59 | 45.2s |
Results on BTCUSD H1 validation set — 5-fold TimeSeriesSplit CV.
| Category | Stack |
|---|---|
| ML / AI |      |
| Data |     |
| Finance |   |
| Tools |    |
Contributions are welcome! Please see CONTRIBUTING.md for guidelines.
This project is licensed under the MIT License — see LICENSE for details.