Walk-forward backtest results with statistical significance analysis
| Period | Range | Months | Ann. Return | Ann. Vol | Sharpe | Sortino | Max DD | Significance |
|---|---|---|---|---|---|---|---|---|
TrainIN-SAMPLE | 1994–2015 | 264 | +4.77% | 4.20% | 1.14 | 1.62 | -9.48% | t=4.16 |
Validate | 2016–2017 | 24 | +4.30% | 2.72% | 1.58 | 5.57 | -1.25% | t=1.49 |
Test 1 | 2018–2019 | 24 | +2.28% | 3.03% | 0.75 | 0.92 | -2.49% | t=0.94 |
Test 2 | 2020–2021 | 24 | +8.65% | 6.54% | 1.32 | 2.05 | -6.37% | t=1.36 |
Test 3 | 2022–2025 | 48 | +2.52% | 5.80% | 0.43 | 0.69 | -7.64% | t=0.82 |
24-month out-of-sample windows have wide confidence intervals. A Sharpe of 0.75 over 24 months has SE ≈ 0.80 — the 95% CI includes zero.
Fama-French factors are long-short academic portfolios, not directly tradable strategies. Our stock selection uses these factors differently than the backtest.
Walk-forward validation prevents look-ahead bias but cannot fully eliminate survivorship bias in the underlying factor return data.
The aggressive profile shows Sharpe of -0.78 in bear markets. This is a known structural weakness — growth and momentum factors reverse in drawdowns.
Aggressive profile bootstrap CIs are extremely wide — growth weight CI [3%, 60%] indicates the optimal weight is not well-determined.
Past performance does not guarantee future results. This validation is a statistical exercise, not a promise of returns.
The conservative profile shows positive Sharpe ratios across all 4 out-of-sample periods, with the weakest being 0.43 over the most recent 48 months. While no individual OOS period reaches statistical significance (the windows are too short), the consistent positive direction has only a 6.25% probability of occurring by chance.
The aggressive profile is higher-return but structurally fragile — it collapses in bear markets and its bootstrap CIs are extremely wide. Use it for growth-seeking portfolios where you accept the drawdown risk.
This is a statistical exercise, not financial advice. The backtest uses academic factor returns as proxies, not our exact stock selection methodology. Real-world performance will differ.