Forecasters often know an event is imminent but lack information about its shape, magnitude, or timing. We introduce Event Signature Transfer (EST), a training‑free, model‑agnostic operator that converts a completed historical event into an explicit forecast scenario. EST first removes the source event’s trend and seasonality, then scales and retimes the remaining signature onto a target forecast, preserving the forecast’s linked structure and collapsing back to the original forecast when the transfer strength is zero. Because it reads only output quantiles, EST works with any quantile forecaster without accessing model internals or requiring additional training. We evaluated EST on Chronos‑2, TimesFM‑2.5 and Toto‑2.0 across twelve real episodes and ten synthetic scenarios; a manually tuned EST reduced in‑sample weighted quantile loss (WQL) by 21.7%–90%. On Chronos‑2, EST outperformed covariate conditioning, activation editing and raw replay in eleven of twelve matched comparisons. Note that the operator builds a scenario but does not estimate its likelihood.
Review