A 6-model ensemble — LinReg, LogReg, GBT, LightGBM, LSTM, and Regime Classifier — tested on 68 S&P 500 stocks across 16 quarterly windows from 2022 to 2025. Zero lookahead bias. Full transparency.
Traditional backtests often cheat — they train on future data. Alpha Signal uses walk-forward validation: models see only what a real investor would have known at each point in time.
No single model has an edge in all market regimes. The ensemble combines classical statistics, gradient boosting, and deep learning — each votes, confidence-weighted.
These results come from a walk-forward backtest — 16 quarterly windows where models were trained strictly on past data before each test period. This is how practitioners in quantitative finance validate strategies; traditional backtests that train on the full dataset are not credible.
The ensemble outperformed SPY across a period that included the 2022 bear market, the 2023 tech recovery, and 2024–25 AI-driven growth. It performed best as drawdown protection — reducing losses in falling markets while capturing upside in rising ones.
BUY signal accuracy sits at —%, with an average return of —% per signal. Not every call is right — but the edge compounds over thousands of decisions.
Stock-by-stock drill-down. Walk-forward windows. Live signal track record. Paper portfolio builder. All of it, right here.
Can a 6-model ML ensemble, tested with zero lookahead bias, beat the market?
Walk-forward backtest on S&P 500 stocks — built entirely with Claude Code.
| Date | Signal | Price | Buy Prob | Day Return | Correct |
|---|
| Stock | Signals | BUY | SELL/SHORT | BUY Acc | SELL Acc | BUY Avg Ret | Trades | $ Traded | P&L |
|---|
| Window | Train End | Test Start | Test End | Signals | BUY Acc | SELL Acc |
|---|
Pick S&P 500 stocks, set your capital, and track how Buy & Hold compares to following Alpha Signal's recommendations — starting today.