Does the corrected FinBERT + XGBoost + CVaR strategy remain effective during both favorable economic conditions and recessions or market crises?
- Strategy: corrected walk-forward backtest produced on 2026-07-13.
- Evaluation window: 2020-01-03 through 2024-12-30.
- Economic labels: monthly NBER-based USREC indicator from FRED. A value of 1 denotes recession.
- Business-cycle interpretation follows the NBER Business Cycle Dating Committee.
- Market regimes use deterministic, reproducible thresholds:
- crisis: VIX >= 30 or SPY 20-day return <= -10%;
- stress: VIX >= 20 or SPY 20-day return <= -5%;
- calm: all remaining observations.
- Annualized return in the segmented tables is daily arithmetic mean multiplied by 252. Conditional compound return chains only the days belonging to that regime.
- Confidence intervals are deterministic 2,000-sample bootstrap intervals.
| Regime | Days | Annualized return | 95% CI | Volatility | Sharpe | Conditional drawdown | SPY-relative return |
|---|---|---|---|---|---|---|---|
| Expansion | 1,213 | 11.66% | −1.11% to 23.63% | 13.99% | 0.83 | −17.85% | −3.98 pp |
| Recession | 43 | 29.46% | −238.46% to 278.44% | 54.66% | 0.54 | −20.77% | +13.58 pp |
| Calm market | 648 | 30.47% | 19.96% to 40.79% | 8.56% | 3.56 | −3.69% | −14.86 pp |
| Stress market | 470 | 13.94% | −9.27% to 38.01% | 16.27% | 0.86 | −19.12% | +1.38 pp |
| Crisis market | 138 | −78.90% | −177.30% to 18.16% | 36.84% | −2.14 | −38.12% | +34.32 pp |
The recession-minus-expansion daily mean difference has a Welch-test p-value of 0.894. There is no statistically useful evidence that average performance differs between those economic labels. The recession estimate is especially weak because it contains only the March–April 2020 observations.
Calendar-year returns were positive in 2020, 2021, 2023, and 2024. The strategy lost 13.11% on an annualized-mean basis in 2022, while SPY lost 17.20%, indicating relative protection but not positive absolute-return stability during that bear-market year.
The model works well in calm markets and remains positive during ordinary stress, but it is not stable as an absolute-return strategy during crisis conditions. Its crisis loss was smaller than SPY's, yet a −78.90% annualized conditional mean and −38.12% conditional drawdown are too large to call the strategy recession- or crisis-proof.
The evidence supports adding an explicit crisis defense—such as volatility targeting, exposure scaling, or a cash/hedging overlay—and rerunning the same fixed regime analysis. More historical recession cycles are required before making a strong economic-cycle robustness claim.
python -m src.cli analyze-regime-stability \
--backtest-dir reports/pipeline-corrected-2026-07-13/backtest \
--output reports/regime-stability-2026-07-13Generated artifacts include daily labels, economic and market-regime metrics, yearly metrics, the recession-versus-expansion test, and a generated Markdown summary.