Machine Learning Stock Selection Model

Problem Addressed

In A-share multi-factor stock selection, factor construction, model training, backtesting, and reporting are often scattered across separate scripts. ML-Alpha connects real akshare price and fundamental data, 20 factor features, rolling training, quintile backtesting, and HTML reporting into one reproducible workflow for studying 1/5/10/20-day forward return predictions, not direct trading signals.

How It Works and Limits

• Data layer: Fetches forward-adjusted daily prices, index constituents, and indicators such as PE, PB, and ROE; builds a date x stock x features panel; applies outlier clipping, cross-sectional standardization, and suspension imputation.
• Feature layer: Covers technical, volume-price, fundamental, momentum, and volatility factors, including moving-average ratios, RSI, volume ratio, turnover changes, PE percentile, 1/3/6-month momentum, and 20-day volatility.
• Model layer: Supports XGBoost, LightGBM, and Ridge; trains on the past N months and predicts the next period, using TimeSeriesSplit to reduce leakage.
• Backtesting layer: Sorts stocks by predicted score, long the top group, short the bottom group, rebalances monthly, and calculates portfolio return and turnover.
• Evaluation and reporting: Outputs IC, rolling IC, group monotonicity, Sharpe ratio, max drawdown, feature importance, and self-contained HTML with equity curves, heatmaps, and scatter plots.

Limits: A-share only, no HK or US stocks; backtests exclude commissions, slippage, and market impact; financial data may lag; models can fail in crises or policy shocks. It is intended for quantitative research and academic analysis, not investment advice.