Is a day-ahead electricity price forecast better than a naive one?

So far yes: over 27 scored days the model’s mean error is 28.3% below the naive forecast that is better on average, and it was the forecast with less error on 70.4% of those days.

The record is 33 days short of the 60 days it needs before it counts as evidence.

For people who work in the sectorIn plain languageFor machines (JSON)

The record, day by day

How big the differences are

The model made less error on 19 of the 27 scored days and more on 8.

The days the model made more error

Every day the model made more error than the naive forecast that made less error that day, ordered from the largest excess down.

The same question on the other three

When the mean and the median are far apart, a few unusual days are carrying the mean.

Month by month

September 2026

20 days were scored this month. The model made less error on 14 of them and more on 6.

What this record does and does not show

• An experimental evaluation is under way: 27 days so far, each recorded before its outcome was known, then scored. The day-count condition is not met yet — 33 days short. The numbers below are informational only — not yet enough for a reliable conclusion.
• Coverage — expected days: 43; forecast issued: 27; scored: 27. Days with no forecast issued: 16.
• One metric only: mean absolute error, over the same days and the same hours for both forecasts. A day whose 24 hours are incomplete is not scored at all.
• It says which forecast made less error. It does not say why the price moved, and it is not a forecast of how the coming days will go.
• Experimental and deliberately open — until all four conditions are met this record is not evidence.

The four publishing conditions

• Minimum days scored: measured 27, threshold 60. not met
• Minimum MAE improvement: measured 28.3%, threshold 10.0%. met
• Minimum per-day win rate: measured 70.4%, threshold 60.0%. met
• No losing month: measured 0, threshold 0. met

Method

Every day a forecast is produced for tomorrow’s 24 hours and written down before the outcome is known. When the realised price is published, that day is scored — the model against two naive forecasts, over the same hours. The forecast written first is never revised.

The difference in the answer above is measured against the single naive forecast that is better on average over the whole period, while the day table compares the model with whichever naive was better that day — the second is the harder test, so the difference you can work out from the table is the smaller one.

Tomorrow's forecast uses only what would be known before market gate closure.

What is compared

• Model (Gradient Boosting)
• Naive: Last Known Day
• Naive: Same Hour Last Week

Where this comes from

Source

EPİAŞ Şeffaflık Platformu (Piyasa Takas Fiyatı). The forecasts and their scores are recorded by Rasat.

Period

09.08.2026 to 20.09.2026

These figures were read at

21.09.2026 02:09 TRT

Open this record in Rasat

Related questions

• How has the day-ahead price moved over the last 30 days?
• How much water is in the reservoirs?
• All reports