Risk & Volatility Forecasting, Published in the Open
Almost no crypto data vendor publishes real, measured accuracy for its risk & volatility forecasts. We do. Here is exactly what each model answers, how accurate it is per asset, and what it does not do.
What the Four Horsemen Answer
Four independent classifiers, each built on a distinct math principle. Plain-language questions, not jargon.
Tail Risk
Will the next bar move further than the recent 90th-percentile move?
Power Law Distributions
Regime State (High Vol)
Is the next bar in a high-volatility regime vs. the 50-bar median?
Volatility Regime
Regime Change
Is the next bar the first high-vol bar after a calm stretch?
Topological Persistence
Regime Emergence
Is short-horizon (10-bar) volatility spiking vs. the 30-bar median?
Complex Systems Emergence
None of the four call price direction. They flag risk and volatility regime - not up or down.
Validated Across 30+ Assets, Daily & Hourly
748 model/asset/timeframe combinations. Ranges shown, not a single cherry-picked number.
| Signal | Model types | Avg accuracy | Range |
|---|---|---|---|
| Tail Risk | XGBoost / LightGBM / CatBoost | 75.5% | 70.1–87.4% |
| Regime Emergence | LogisticRegression / XGBoost / etc. | 84.8% | 77.8–88.5% |
| Regime State | LogisticRegression / XGBoost | 92.5% | 85.8–95.6% |
| Regime Change | LightGBM / CatBoost / XGBoost | 93.3% | 71.1–96.6% |
Source: high_accuracy_models.csv, machineLearning-models repo. n = 748 across 30 assets, daily + hourly.
Core Alpha: Volatility Models
A different model family - regression and classification on volatility magnitude, not regime.
Future Volatility Regressor
XGBRegressor. Forecasts how much the next bar is likely to move - not which way price goes.
Avg RMSE ≈ 0.0088 across clean assets.
Volatility Expansion Classifier
XGBClassifier. Probability that volatility expands beyond its 20-day baseline.
Accuracy avg 75.4% - this is our strongest, most audit-ready claim today.
Check Our Work
Every number on this page is pulled directly from our public methodology repository - no cherry-picking, no marketing rounding beyond one decimal.
Read this before you trade on any of it
- None of the Four Horsemen or Core Alpha models predict price direction. They forecast risk and volatility regime - pair them with your own directional thesis or entry signal.
- Tail Risk's positive class is roughly the top 10% of bars by definition, so raw accuracy alone understates how it compares to a naive baseline - weigh it alongside the published confidence calibration, not accuracy in isolation.
- Two low-history assets (CC-USD, NIGHT-USD) showed degenerate metrics and are excluded from every figure on this page until they have enough backtest history.
- All figures are measured on historical data. Markets change; methodology is under continuous improvement. Not investment advice.