Check our record

We forecast one week ahead. The underlying data is the U.S. Energy Information Administration's weekly California regular retail price series, published each Monday, so the next print is the finest thing a forecast can honestly be graded against. We do not claim a specific day or a two hour window.

Every score is forward only. Each prediction was made using nothing but the weeks published before its target week, so no result on this page benefits from hindsight.

The length of the trend window was chosen by testing every value against the full record, so a directional accuracy scored on that same record would be flattered by the choice. The directional figure above is scored on the second half of the data instead. Being straight about the limit: that second half was still part of the record the window was chosen against, so treat it as a stability check across eras rather than a clean out of sample test. A proper holdout, choosing the window on the first half alone and scoring on the second, is the next thing on this list. Average error and the share of weeks within ten cents are scored on those same weeks, so every figure in the comparison describes one consistent sample.

Version 1 of the model, a twelve week trend, was retired on 27 July 2026 after the backtest showed it losing to the persistence baseline. It is left in the table above rather than deleted.

Our current model sees exactly one thing: how the price moved last week. It cannot use anything else, however predictive, which is the honest limit on how good it can get.

The clearest next input is Los Angeles CARBOB wholesale spot pricing. Wholesale moves reach the pump with a lag of roughly one to two weeks, so it leads the number we are forecasting. After that: refinery outages, and scheduled fuel tax and Low Carbon Fuel Standard changes, which are known in advance and currently show up as forecast error.

None of these are used yet. They are named here so this page describes what we actually do today, not what we intend to do.