📈 LSTM-Based Stock Forecasting & Backtesting Pipeline

A modular research framework for building, fine-tuning, and rigorously backtesting deep learning models for financial time series forecasting.

---

---

🚀 Overview

This repository demonstrates a complete, production-grade machine learning pipeline for daily stock price prediction, alpha signal generation, and systematic trading backtesting.

Combining:

• ✅ Modern LSTM and Attention-LSTM architectures
• ✅ Rich technical indicator engineering
• ✅ Strong baseline benchmarking (trees, regressors)
• ✅ Transparent backtesting with real trade simulation
• ✅ Full Weights & Biases (W&B) experiment tracking

---

🎓 Academic Motivation

Designed to meet postgraduate research standards in quantitative finance and applied machine learning:

• Trains sequence models on engineered features that capture price momentum, trend, volatility, and cyclical structure.
• Supports stock-specific fine-tuning and multi-asset pretraining.
• Evaluates prediction quality with both statistical loss (RMSE) and risk-focused metrics (directional accuracy, Sharpe ratio, drawdown).
• Provides fully auditable trade signals, charts, and logs.

---

💼 Industry Relevance

Built to illustrate practical steps for real-world systematic trading signal development:

• Converts raw forecasts into actionable Long/Short trading signals via robust threshold sweeps.
• Simulates realistic equity curves, entry/exit logic, and PnL paths.
• Outputs easy-to-interpret trade logs (CSV) for third-party review.
• Provides clear benchmark baselines for measuring ML edge.

---

⚙️ Pipeline Modules

Module	Purpose
pretrain_general.py	Pretrain a general-purpose LSTM on multiple tickers to learn shared dynamics.
finetune_single.py	Fine-tune pretrained weights on a specific stock over any custom date window.
feature_engineering.py	Adds robust technical signals: SMA, EMA, RSI, MACD, Bollinger Bands, OBV, and more.
train.py	Core training loop with scaling, batching, loss tracking, and W&B logging.
baseline_models.py	Runs classical regressors (Linear, Ridge, Lasso, SVR, Trees) for benchmark comparison.
evaluate.py	Computes test RMSE, directional accuracy, and mean predicted signal return.
backtest_strategy.py	Turns predictions into realistic trade positions, simulates PnL, exports trade logs.
threshold_sweep.py	Sweeps thresholds to find robust Long/Short signal breakpoints.

---

📊 Key Metrics & Plots

• RMSE – prediction error for price forecasts.
• Directional Accuracy – % of up/down calls correct.
• Win Rate – % of trades that finish profitably (Long/Short).
• Sharpe Ratio – return per unit risk.
• Max Drawdown – worst peak-to-trough loss.
• Equity Curve – how the strategy capital grows over time.
• Price Chart – marks every entry/exit directly on the asset price.

All results log to Weights & Biases for reproducibility and tracking.

---

📈 Example: Realistic Backtest

A typical run:

1. Train on any stock for any date range.
2. Predict next-day price, convert to implied return.
3. Apply asymmetric thresholds:
   • Long: if predicted return > +X%.
   • Short: if predicted return < –X%.
4. Simulate trade PnL, visualize real entry/exit points on the price chart.
5. Export a trade log CSV for transparency.

---

✅ Quick Start

# Fine-tune on AAPL
python finetune_single.py

# Backtest with realistic trade simulation
python backtest_strategy.py

# Sweep thresholds to find robust signal breakpoints WIP
python threshold_sweep.py