BIST stock selection using combined LightGBM and ElasticNet predictions. 29 features, walk-forward training, monthly rebalance.
| # | Hisse | Skor | Ağırlık | Ay Getirisi |
|---|---|---|---|---|
| 1 | THYAO | 100 | 5% | — |
| 2 | DOCO | 95 | 5% | — |
| 3 | CLEBI | 90 | 5% | — |
| 4 | TABGD | 85 | 5% | — |
| 5 | TCELL | 80 | 5% | — |
| 6 | BIZIM | 75 | 5% | — |
| 7 | PGSUS | 70 | 5% | — |
| 8 | ARZUM | 65 | 5% | — |
| 9 | ISYAT | 60 | 5% | — |
| 10 | METEN | 55 | 5% | — |
| 11 | MOPAS | 50 | 5% | — |
| 12 | MASFN | 45 | 5% | — |
| 13 | GLCVY | 40 | 5% | — |
| 14 | AGESA | 35 | 5% | — |
| 15 | RODRG | 30 | 5% | — |
| 16 | GARAN | 25 | 5% | — |
| 17 | ORZAX | 20 | 5% | — |
| 18 | AKGRT | 15 | 5% | — |
| 19 | GLBMD | 10 | 5% | — |
| 20 | AKBNK | 5 | 5% | — |
The ML Ensemble portfolio uses the simple average of rank-normalized predictions from LightGBM and ElasticNet models. Each model is trained on 29 features (momentum, volatility, valuation, quality, sector, market regime). Walk-forward training ensures each month uses only data available up to that date — never looking ahead.
Model-level analysis showed that among 6 models (XGBoost, LightGBM, RF, Ridge, ElasticNet, CatBoost), LightGBM and ElasticNet individually delivered the highest cumulative returns. The stacking meta-model was removed because it overfit (66% worse than simple average). Simple averaging produces more reliable results than complex stacking.
Top features: sector average volatility, 21-day momentum, sector average E/P, market breadth, and 63-day momentum. Sector-level features ranked higher than individual stock features — suggesting sector rotation is a stronger signal than individual stock selection on BIST.
Related articles: ML Stock Selection, Factor Investing, 10-Year BIST Backtest