Across tickers
The same feature pipeline and model registry applied to every trained ticker: 32,453 daily bars in total. Consistent results across stocks matter more than one lucky backtest.
Summary by ticker
Naive accuracy = always predicting the more common direction in the test period. The model for each ticker is picked by training-period CV ROC AUC, so its test accuracy and Sharpe (5 bp cost, long/short) are genuinely out-of-sample.
| Ticker | Test period | Test days | Naive accuracy | CV-selected model | Its test accuracy | Strategy Sharpe | Buy & hold Sharpe |
|---|---|---|---|---|---|---|---|
| S&P 500 ^GSPC | Jun 2021 – Sep 2026 | 1332 | 53.5% | Random Forest | 52.9% | 0.60 | 0.69 |
| Amazon AMZN | Jun 2021 – Sep 2026 | 1332 | 50.9% | Gradient Boosting | 50.6% | -0.18 | 0.24 |
| Microsoft MSFT | Jun 2021 – Sep 2026 | 1332 | 52.0% | Logistic Regression | 50.4% | 0.15 | 0.52 |
| Alphabet (Google) GOOGL | May 2022 – Sep 2026 | 1100 | 52.8% | Bagging (KNN) | 50.9% | 0.05 | 0.78 |
| Oracle ORCL | Jun 2021 – Sep 2026 | 1332 | 51.7% | Extra Trees | 49.6% | 0.19 | 0.25 |
Models across stocks
Does any model work consistently, or only by luck on one ticker?
The stocks themselves
Full-history behaviour of each stock and the index
Takeaway Pairwise correlations range from 0.35 to 0.70: big tech moves largely with the market, so these names diversify each other less than it might seem.