Building an S&P 500 Stock Return Direction Prediction System with PyTorch: A Comparative Study of Traditional Machine Learning and Recurrent Neural Networks

This article introduces an S&P500 stock 5-day future return direction prediction system built using PyTorch, comparing the performance of traditional machine learning models (logistic regression, random forest) and recurrent neural networks (RNN, LSTM, GRU). The core of the project is to transform return prediction into a three-classification task (decline/flat/rise), focusing on solving data leakage issues, constructing technical features, and optimizing model performance using class-weighted loss. The following floors will detail the background, methods, results, and findings.

In the field of financial quantitative investment, accurate price prediction is difficult and has limited effectiveness, while predicting return direction (rise/decline/flat) is more practically valuable. This project (sp500-return-direction-rnn) transforms the problem into a three-classification task, aiming to predict the return direction over the next 5 trading days to provide decision-making references for investors.

Data processing aspects: 1. Obtain OHLCV data of S&P500 constituent stocks and construct 13 technical features (return rate, moving average ratio, volatility, momentum indicators, volume changes, etc.); 2. Use strict time-series splitting to avoid data leakage: training set (2010-2018), validation set (2019-2021), test set (2022-2026); 3. Define the target variable as the classification of 5-day future return (0=decline, 1=flat, 2=rise).

Model comparison and training: 1. Baseline models: majority class classifier, logistic regression, random forest; 2. Neural networks: RNN, LSTM, GRU, with the final choice being a single-layer GRU (128 hidden layers + 0.3 dropout); 3. Training parameters: batch size 1024, learning rate 0.001, training for 8 epochs; 4. Use class-weighted cross-entropy loss to solve the class imbalance problem.

Evaluation and key findings: 1. The main metric is macro-average F1 (balancing performance across all classes); 2. Weighted GRU achieves F1=0.39351 on the validation set and F1=0.368519 on the test set; 3. Ablation experiment: Adding SPY market features did not improve performance; 4. Key findings: Class weighting effectively improves F1, accuracy is misleading, the flat class is the hardest to predict, and simple market features are not helpful.

Project value and future directions: 1. Clear code structure (configs/notebooks/src/outputs) with strong reproducibility; 2. Demonstrates a real time-series ML workflow, emphasizing data leakage prevention and reasonable evaluation; 3. Limitations: For educational research only, does not constitute investment advice; 4. Future directions: Attention mechanisms, multi-task learning, ensemble learning, and actual trading backtesting.