Mean-variance λ=1 con shrinkage Ledoit-Wolf · backtest walk-forward vs stake plano
| métrica | Kelly (Σ⁻¹μ) | Flat |
|---|---|---|
| Bankroll final | 4.3235x | 4.4783x |
| Retorno total | 332.3% | 347.8% |
| Max drawdown | 57.1% | 57.1% |
| Volatilidad | 0.1866 | 0.1895 |
| Sharpe-like | 0.3093 | 0.3137 |
| Hit rate | 0.54 | 0.51 |
| Pasos | 37 | 37 |
| Δ bankroll | -3.5% vs flat | |
Nada cruza el umbral → Kelly = flat. Demuestra que la maquinaria es correcta, no que haya edge.
| métrica | Kelly (Σ⁻¹μ) | Flat |
|---|---|---|
| Bankroll final | 6.1293x | 4.4783x |
| Retorno total | 512.9% | 347.8% |
| Max drawdown | 37.4% | 57.1% |
| Volatilidad | 0.1574 | 0.1895 |
| Sharpe-like | 0.3946 | 0.3137 |
| Hit rate | 0.59 | 0.51 |
| Pasos | 37 | 37 |
| Δ bankroll | +36.9% vs flat | |
Solo demostrativo: bajar el umbral con n=16 no valida nada.
| estrategia | n | μ ret | hit |
|---|---|---|---|
| model-corners | 37 | +0.35 | 0.68 |
| model-goals | 37 | +0.19 | 0.59 |
| corners-mismatch | 23 | +0.04 | 0.52 |
| anti-narrativa | 3 | -0.33 | 0.33 |
| goals-under-suppressed | 2 | +1.00 | 1.00 |
Retornos even-money (±1u). Sin odds reales del book → estimación de edge limitada.
Método: pesos vía np.linalg.solve(Σ, μ) (sin inversa explícita), no-shorting,
cap por mercado 0.10, fracción Kelly 0.25, normaliza solo si el total > 1. El backtest re-estima
μ/Σ en cada paso usando solo partidos previos (out-of-sample; respeta la regla anti in-sample del charter).
Re-correr al acumular liquidaciones (meta n≥30 por estrategia).