Skip to content
EPF Docs

Intraday Forecastability — Tested and Rejected

Date: September 1, 2026 | Status: Investigated — rejected on evidence

Why this version exists

Day-ahead prices are fixed once the gate closes, so the only place a recurrent forecast — one that keeps updating as new information arrives through the day — could add value is intraday. We backfilled a full year of OMIE’s public intraday files (IDA1, IDA2, IDA3, and the Continuous Intraday Market) for the Spanish zone and tested, with real backtested evidence, whether the gap between the day-ahead price and each intraday market’s cleared price can be predicted.

What we tested

Three checks, in increasing order of how hard they are to argue away:

  1. Does the spread even matter? Yes — it’s 8.7% to 21.2% of the average price depending on the market, so this isn’t a “nothing happens” result.
  2. Can anything beat guessing “no change”? We compared the naive baseline (assume the intraday price equals the day-ahead price) against an hour-of-day average, yesterday’s spread, and a 7-day rolling average.
  3. Does hindsight help? We correlated the spread against the realised renewable generation forecast error — using the actual error rather than a forecast of it, which is the most favorable case any real predictive model could hope to approach.

What we found

Nothing beat naive. Not persistence, not a rolling average, and — the sharpest version of this test — not even an hour-of-day average fitted and scored on the same data, which is the most optimistic result any model could produce and it still lost. The renewable-error correlation explained about 1% of the variance (R² ≈ 1%), despite the underlying forecast errors themselves being large. Full numbers, method, and the exact backtest window are in the linked analysis below.

Decision

No intraday price model will be built. This reads as an efficiently-priced market: the day-ahead price already reflects the public information available at gate close, and the residual is driven by information that only arrives afterward — which is, by construction, not available to any forecast made ahead of that gate. The negative result is published rather than shelved, because ruling this out with evidence is itself useful — it closes a line of investigation the project would otherwise have kept revisiting.

Key files