Residual-corrected forecaster

Forecaster id residual(<base id>) · tycheon.covariates.ResidualCorrectedForecaster

A base forecaster plus a model of its errors from as-of-stamped covariates (news sentiment, fundamentals, macro series). It sits behind a ResidualCorrector interface so a better fusion method can replace the first implementation without touching callers.

Source

Implemented in Tycheon. FeatureBuilder computes each feature row only from covariate versions published by that row's time. LightGBMResidualCorrector trains LightGBM (native API) to predict the base model's realised log-return residual, per horizon step, on a chronological train/validation/test split. The correction is accepted only if it beats a zero correction on the untouched test split by min_improvement; otherwise the base forecast passes through unchanged and the report says so.

License

Tycheon: Apache-2.0. LightGBM: MIT.

Training data

The caller's covariate store and bars. Covariate values are numeric only, stored bi-temporally (timestamp, available_at, optional key; revisions are later versions). News text is never read by this code: only numeric scores derived elsewhere are accepted, so instructions inside a headline cannot reach it.

Intended use

Testing whether covariates add information beyond price, with an honest answer when they do not.

Limitations

• Residual correction is the simplest fusion; it cannot represent interactions with the base model's internals.
• Small samples overfit: the acceptance rule guards this but does not remove it.
• It refuses an as_of earlier than its training data (that would be lookahead).
• Sentiment scores, fundamentals and macro series are only as point-in-time as the caller stamped them.

Calibration

Uncalibrated as built; wrap it with ConformalCalibrator. A correction shifts the centre of the distribution, and calibration then measures whether the intervals around it hold.

Baseline comparison

The zero-correction (base) forecast is the baseline in the acceptance test: the correction is only used if it beats no correction out of sample. It is not in the published benchmark run. Nothing on this card claims that the residual-corrected forecaster beats the random walk. Run it yourself with make benchmark-small after adding it to a config; every result is reported against the random walk with Diebold-Mariano tests. See benchmark methodology.

For research and risk analytics. Not investment advice.