Stock Jump Prediction and Strategy Backtest

This project builds and evaluates a machine learning model to predict post-earnings stock price jumps and implements a trading strategy based on the model's predictions. It outperforms traditional buy-and-hold approaches in backtests.

🚀 Project Goals

• Predict whether a stock will significantly jump or drop after earnings
• Backtest a threshold-based and confidence-weighted trading strategy
• Compare against a buy-and-hold benchmark

🔍 Problem Framing

The model is trained as a binary classifier to predict jump_label:

• 1 → stock makes a significant positive move (e.g., > +3%)
• 0 → stock does not make a significant move or drops

The output is a probability, which is translated into trading actions:

• Buy if probability > 0.63
• Sell (short) if probability < 0.37
• Hold cash otherwise

🧠 Features Used

Final model uses the following features:

• eps_surprise_pct: Earnings surprise percentage
• positive_count: Count of positive headlines
• negative_count: Count of negative headlines
• pos_total_ratio: Ratio of positive headlines
• sentiment_label_encoded: Encoded polarity label of sentiment (POS/NEG/NEUT)
• pos_neg_ratio: Ratio of positive to negative headlines

🧪 Model

• Model: LightGBM Classifier
• Optimized with Optuna for AUC
• Class-weighted, depth-limited (max_depth=4)
• Threshold Strategy: 0.63 / 0.37 split for confident trades

📈 Strategy Performance

Strategy	Cumulative Return
Threshold-Based Model	5.5x
Confidence-Weighted	2.3x
Buy & Hold Benchmark	3.5x

The model outperforms both passive investing and probabilistic scaling.

🧰 Tech Stack

• Python
• LightGBM
• Optuna
• Pandas, NumPy, Matplotlib
• FinBERT Sentiment Scoring

📂 Project Structure

├── subset_trn.py           # Full training and evaluation script
├── feature_subset_backtest_results.csv   # Dataset with features and labels
├── best_model_final_subset.pkl  # Trained LightGBM model
├── strategy_comparison_final_features.png  # Strategy performance plot
├── README.md                    # This file

🧠 Key Learnings

• Simpler feature sets often outperform over-engineered ones
• Calibration and threshold tuning matter more than raw AUC
• Feature selection has a large impact on trading performance

🔄 Next Steps

• Extend model to other sectors (e.g., defense, banks)
• Incorporate macro indicators (VIX, SP500 trend)
• Add Sharpe ratio, drawdown, and capital allocation simulation

📬 Author

Aniket Mandal — PhD in Theoretical Chemistry, specializing in ML + scientific modeling

---

For questions or collaboration inquiries, please contact or connect via LinkedIn.