Walk-forward and baselines
Three naive baselines
Section titled “Three naive baselines”A naive baseline is a simple rule that needs no model at all. EPF compares every forecast with three of them:
| Baseline | Rule |
|---|---|
| Same as yesterday | Each quarter-hour tomorrow costs what it cost today (technical name: persistence, T-1). |
| Same day last week | Each quarter-hour costs what it cost on the same day a week earlier (seasonal, T-7). |
| Average of the last 4 same weekdays | Each quarter-hour costs its average on the same weekday over the last four weeks (similar-day, mean of 4). |
They are produced every day for every market, at the same resolution and horizons as the real forecast, and scored the same way. “Best naive” in our tables is whichever of the three did best over the period.
Why the naive is the bar
Section titled “Why the naive is the bar”Electricity prices repeat strongly from day to day and week to week, so simple rules are surprisingly hard to beat. A forecast with an impressive-sounding error figure may still be worse than copying yesterday. Reporting MAE without a baseline hides that. On every page where EPF shows its own accuracy, it shows the naive alongside it, including where EPF loses.
Walk-forward testing
Section titled “Walk-forward testing”A walk-forward rebuild forecasts the past again, the way production would have lived it.
- Pick a month.
- Train a model only on data from before that month.
- Forecast every day of that month, using only inputs that would have been published at each forecast time.
- Move forward one month and repeat.
Over twelve monthly folds this gives a year of out-of-sample forecasts, each made by a model that never saw the period it is scored on. It covers different seasons and price regimes, which a single test window cannot.
A candidate model is promoted only if it beats the current production model and the naive baselines in a walk-forward test, wins on most individual days rather than on a few large ones, and does not trade away the economic measures. A human approves every promotion.
Live versus walk-forward
Section titled “Live versus walk-forward”EPF reports two records, and they answer different questions.
- Walk-forward asks: how good is the current method, measured fairly over a year?
- Live asks: what did production actually deliver on each day?
The live record includes every model that was in production at the time, and any problem in how forecasts were served. When the two disagree, the gap is itself a finding. In 2026 that gap led to audits which found that some inputs available in testing were missing at forecast time in production. Those issues were fixed in September and October 2026; the live record still includes the months before the fixes and will take time to reflect them.