The research
Ontario redesigned its electricity market on 1 May 2025. GridCast uses statistical testing and machine-learning models to study how prices, forecasting methods and battery strategies cope with the new market. Every question, data split and success criterion was registered before any model saw test data.
Pre-registration
Registered –, before any model touched data. Later changes are dated amendments inside the document.
Commit – · SHA-256 –
Research questions
- RQ1Did the redesign change how prices behave?
- RQ2Do zero-shot AI forecasting models match tuned specialists?
- RQ3After a redesign, which approach recovers fastest?
- RQ4Does better uncertainty make a battery more money?
Model comparison (validation, July–October 2025)
Five models in a rolling-origin backtest: gradient-boosted quantile regression (LightGBM) with conformal calibration, SARIMAX (statsmodels), LASSO-regularised autoregression (LEAR, scikit-learn) and two naive baselines. CRPS scores the whole forecast distribution; lower is better. These validation results were used to tune, choose and freeze each model. Test-year results are released after the data gate, so nothing could be tuned to them.
All metrics
How unusual is a "break"? A placebo check
Year-to-year price swings alone make standard break tests fire. To calibrate RQ1, the same four statistics (Chow F-tests on price level and volatility, a Kolmogorov–Smirnov distance and a Hill tail-index ratio) were computed at 48 dates in the old market when nothing changed. The redesign counts as a break only if it beats about 95% of these placebo dates on both level and volatility. Its own values appear here once the data gate is passed.