MechAI← All projects

Open source · PyPI

Dependence-Forecastability

A deterministic pre-modeling layer that asks whether a time series contains exploitable structure — before you spend weeks on model search.

Most forecasting projects start too late: data → model search → tuning → more features → more compute → still poor results. Dependence-Forecastability inserts the question that should come first:

Does this series contain exploitable structure, and what should a model be allowed to use?

It is not another forecasting library. It's a deterministic pre-modeling layer that inspects readiness, informative horizons, target lags, seasonality structure, covariate usefulness, and leakage risk — before you reach for Darts, MLForecast, StatsForecast, Nixtla, Prophet, or a custom model.

What it does

  • Horizon-wise forecastability using interchangeable dependence scorers (AMI / pAMI and KSG-II estimators).
  • True Schreiber transfer entropy for covariate screening — which external signals actually carry predictive information, and at which lag.
  • Statistical rigor by default — Romano–Wolf FWER correction across lags so you don't fool yourself with multiple-comparison noise.
  • Rolling-origin benchmarking and reproducible reporting, with a typed agent boundary and a physical domain/ layer.

Why it matters

It turns the most expensive guesswork in a forecasting engagement — is this even learnable? — into a fast, defensible, reproducible diagnostic. The companion examples repo walks through it on real datasets.