Source: experiments.md — Experiment Roadmap¶

What it is¶

A five-item bullet list of modelling experiments that have not yet been tried. It is strictly a personal/team ideation note, not a specification or design document.

Section-by-section summary¶

The file contains five open experiment ideas:

First-differencing — try predicting the delta (change from last year) instead of the absolute value.
ARIMA/SARIMAX — try classical time-series regression approaches.
Nixtla — try the Nixtla forecasting framework.
PHENOWEIGHTING — try phenology-weighted features.
Irrigation feature — add an irrigation feature to account for bimodality in Texas.

Notable claims (the load-bearing ones)¶

The file is intentionally minimal. None of the five ideas is a current sprint item.
The irrigation feature for Texas suggests a known data quality issue: bimodal yield distribution in Texas counties is not currently captured by the model.
The phenoweighting idea hints at potential value from weighting weather/climo features by crop development stage rather than treating all growing-season days equally.

What this document is NOT¶

This is not a backlog of structural or code issues (that is TODO.md). It is not a design specification (that is DESIGN.md). It does not describe completed experiments.

Cross-references¶

TODO.md — structural backlog
DESIGN.md — pipeline contract including the feature-decoupling clause that allows changing regression_params.feature_cols without re-running features
features_builders_README.md — builder protocol that any new feature type must satisfy