Regime ensemble

Forecaster id regime-ensemble · tycheon.routing.EnsembleForecaster

Combines several forecasters with the weights of the volatility regime the market is in now. The random walk is always a member, so the ensemble can never do much worse than "no skill".

Source

Implemented in Tycheon. A causal regime detector (VolatilityRegimeDetector: rolling volatility clustered by k-means or quantile cuts, re-fitted on a trailing window at each bar; or MarkovSwitchingRegimeDetector: filtered, not smoothed, probabilities from a statsmodels two-regime model) labels every past forecast origin. RegimeRouter scores each member on its past forecasts with a proper score (the quantile score on a level grid) and turns recent in-regime relative loss into weights, exp(-sensitivity * relative loss), shrunk toward equal weights. Regimes with too few origins use the overall weights. mode="pool" draws sample paths from members in proportion to their weights (joint paths kept intact); mode="quantile" averages quantile functions (Vincentisation) for members without paths.

License

Tycheon: Apache-2.0. statsmodels (BSD-3-Clause) is used by the Markov-switching detector.

Training data

None of its own. Weights are learned from the members' replayed forecasts at past origins of the series being forecast, each using only data known at that origin.

Intended use

Per-asset forecasting where no single model dominates. Its replayed score history is what the conformal calibrator is fitted on, so the published intervals describe the ensemble, not its members.

Limitations

• Weights are only as good as the score history: few origins mean noisy weights, and the router says so in its notes instead of pretending.
• Regimes are detected from volatility alone; a regime that does not show up in volatility is invisible to it.
• A member the router did not score is refused, not silently ignored.
• mode="quantile" has no joint paths, so VaR/ES and drawdown cannot be computed from it.

Calibration

Uncalibrated as built. Wrap it with ConformalCalibrator; the calibrated forecast carries the holdout coverage that justifies (or withdraws) the calibrated label.

Baseline comparison

The random walk is a member and its weight is reported per regime. When it carries most of the weight, that is the honest finding. Measured in the small benchmark (benchmarks/results/small/0.1.0.json: Tycheon 0.1.0, three synthetic series, horizon 5, 108 pooled walk-forward origins). Pooled MASE against the no-change forecast: 1.041. Achieved coverage of the nominal 90% interval: 83.3%. One-sided Diebold-Mariano p-values that it has lower loss than the random walk: 0.991 (squared error) and 0.343 (CRPS), so it is indistinguishable from the random walk. The same ensemble calibrated with adaptive conformal prediction (tycheon-calibrated) reached 86.1% coverage of the 90% interval, the closest to nominal in that run, with the widest interval; it was also indistinguishable from the random walk (p = 0.787 and 0.797). Synthetic data: this shows how the machinery behaves, not whether the model forecasts real markets.

For research and risk analytics. Not investment advice.