Refit the model yourself

The backtests use SARIMAX models fitted offline in Python. The lab uses their lightweight cousin, an ARX model fitted by least squares, which is fast enough to refit a few hundred times every time you move a slider: next month’s change is predicted from the last p changes plus one climate reading k months back, re-estimated on a rolling window, and compared with the same model without the climate term.

The TypeScript code is a port of the Python reference and is tested against it number for number (see the method notes). Try making the result look good, then ask whether you would have picked those settings in advance.