From Index to Equity: Pre-Training Transformers for Stock Return Prediction

Battle Count: 329

Highly relevant to quantitative trading. The paper directly addresses stock return prediction for trading decisions (buy/sell/hold), implements a long-short portfolio strategy (top 5 / bottom 5 stocks), and develops a real-time prediction application. The transfer learning approach from index to individual stocks is a practical technique for quantitative strategies. However, the modest returns and disconnect between MSE and portfolio performance suggest practical trading viability needs further improvement.

Moderate complexity. Uses standard deep learning frameworks (likely TensorFlow/Keras given the architecture description) and scikit-learn/XGBoost. Feature engineering with 15 technical indicators over 10-day windows is straightforward. The multi-stage training pipeline (pre-train on index → fine-tune on stocks → adapt for regression) adds complexity. The Streamlit application for real-time predictions adds deployment complexity. No custom architectures or novel algorithms are introduced.

The paper describes methodology in detail including feature engineering, model architecture, hyperparameters (Table 1 for XGBoost), and data splits. Data is sourced from Yahoo Finance API via yfinance Python package (publicly accessible). However, no code repository is provided, specific random seeds are not mentioned, and the exact preprocessing pipeline details (e.g., class weighting strategy) are not fully specified. The Streamlit application is described but not shared.