Historical performance analysis of our trading strategies. All results are based on out-of-sample backtesting with no lookahead bias.
Machine learning models predicting the probability of significant drawdowns across multiple time horizons and severity levels.
Percentile rank of current crash risk relative to all historical readings
Composite score of 17.7 — higher than 91% of all historical readings
S&P 500 at 6,581 — -5.7% from ATH of 6,979 (Jan 27, 2026)
Based on calibrated drawdown probabilities across 3M/6M/12M horizons and 5%/15%/30% severity thresholds. Updated daily.
Historical crash risk percentile rank alongside actual S&P 500 drawdown from all-time high. Background bands indicate risk zones.
Percentile values are smoothed using a 10-day Exponential Moving Average (EMA) to reduce daily noise while remaining responsive to regime changes. The EMA weights recent observations more heavily (α = 0.182).
Probability of S&P 500 experiencing a maximum drawdown exceeding each threshold within each horizon. Calibrated via isotonic regression; monotonicity enforced across severity levels.
| Drawdown | 3-Month | 6-Month | 12-Month |
|---|---|---|---|
| ≥5% | 19.9% | 26.6% | 47.2% |
| ≥15% | 0.7% | 0.9% | 41.2% |
| ≥30% | 0.7% | 0.9% | 16.0% |
Model-predicted probability of drawdown ≥5% over time. Grey area shows actual S&P 500 drawdown from all-time high. Shaded regions mark crisis periods.
When the model predicted >30% crash probability, how often did a drawdown of that severity actually occur? Higher hit rates indicate more trustworthy signals.
| Drawdown | 3-Month | 6-Month | 12-Month |
|---|---|---|---|
| ≥5% | 47% 1574/3366 alerts | No alerts | 65% 2900/4487 alerts |
| ≥15% | No alerts | No alerts | 64% 911/1432 alerts |
| ≥30% | No alerts | No alerts | 81% 223/277 alerts |
An "alert" is any date where the model predicted >30% probability for that cell. A "hit" means the actual S&P 500 maximum drawdown within the horizon exceeded the threshold.
Average permutation importance across 9 crash models. Features appearing in more models are more universally predictive of drawdown risk.
Methodology
These probabilities are based on maximum drawdown within each horizon window, not point-to-point returns. This means a 6-month/15% probability reflects the chance that the S&P 500 will drop 15% or more at any point within the next 6 months, even if it subsequently recovers.
Each of the 9 models is a CatBoost classifier trained on ~50 years of market data with purged time-series cross-validation. Raw probabilities are calibrated via isotonic regression and monotonicity-enforced so that P(≥30%) ≤ P(≥15%) ≤ P(≥5%) for each horizon. The composite risk score is a weighted average across all 9 cells, tilted toward near-term moderate crashes (most actionable for portfolio hedging).
Important: With only ~5-6 major crashes since 1970, models for extreme scenarios (-30%) have limited training data. Interpret extreme-severity probabilities with appropriate caution.