Daily guide · RO · 2026-08-18
The number your Romania price model has to beat this week
All numbers computed from the week's published auction results at publication and frozen — this article is an honest snapshot, not a page that rewrites itself. Reproduce via the API.
The benchmark to beat this week is a persistence mean absolute error of 27,35 EUR/MWh, the naive "tomorrow equals today" error averaged across the six days. That is the number your Romania model has to clear to justify itself. For context, our own 30-day track record sits at an MAE of 24,08 EUR/MWh against a baseline of 29,59, measured over 244 days, so the target is well within reach but not automatic.
What stands out is how uneven persistence is across the week. It ranges from 21,21 EUR/MWh on 16 August to 37,84 on 17 August, the hardest day, which also carries the highest day mean at 173,11 EUR/MWh. The 15 August case is the other trap: persistence error of 32,23 on a comparatively low day mean of 115,24, a sign of sharp day-to-day swings rather than a steady level.
The concrete takeaway: do not chase a single weekly average. Persistence is already competitive on 14 and 16 August, so focus your edge on 15 and 17 August, where naive carry-over breaks down and beating 27,35 depends on getting those two days right.
Before you train anything on RO prices, you need the number your model has to beat. In day-ahead forecasting that number is naive persistence: predict that tomorrow's curve equals today's. It costs nothing, needs no features, and it is embarrassingly hard to beat — because power prices are dominated by daily and weekly rhythms that persistence gets for free. We computed it on this exact week:
27.35 €/MWh
persistence MAE on this week's RO auctions — the bar. A model that can't get under this adds negative value.
24.08 vs 29.59
our live production model vs its persistence baseline, trailing 244 scored days — the same scoring, in public, on /accuracy
Why the bar moves: this week, day by day
| Target day | Day mean €/MWh | Persistence MAE |
|---|---|---|
| 2026-08-12 | 138.27 | 24.39 |
| 2026-08-13 | 163.18 | 26.22 |
| 2026-08-14 | 143.05 | 22.19 |
| 2026-08-15 | 115.24 | 32.23 |
| 2026-08-16 | 136.12 | 21.21 |
| 2026-08-17 | 173.11 | 37.84 |
Calm days make persistence look unbeatable; transition days (weather swings, weekend boundaries) are where a real model earns its keep. Averages hide this — always look at the daily distribution.
The part everyone gets wrong: the features
The model is the easy half. The silent killer is lookahead: training on weather reanalysis instead of the forecast that existed at issue time, on revised load data, on same-day flow averages that hadn't finished flowing at the auction. Backtests built that way look brilliant and die in production. Our Quant feature matrices serve one row per (issue day, 15-minute slot) where every column is documented with when it became knowable — the same rows our own models train on:
curl -H "Authorization: Bearer YOUR_KEY" \
"https://voltcast.com/api/v1/features/RO?issue_from=2026-07-01&horizons=1"
# columns include: price lags · previous-run weather (as forecast AT issue)
# · grid drivers (lagged to last COMPLETE day) · day-ahead load/RES forecasts
# · ensemble spread · target_price label (null until matured, never imputed)
Start with Home: the persistence bar above needs nothing but GET /v1/prices/RO
(Home covers one zone of your choice). When your walk-forward backtest beats it honestly, prove it in public on the
Forecast Bench — forward-blind, open scoring code, and our
own model is auto-entered under identical rules. If you'd rather buy the plumbing than build it,
contact us about ML-grade data access.
Method & citation. Prices are hourly means of published day-ahead auction results (native 15-minute periods averaged; ENTSO-E/SMARD, attributed). Wholesale-price component only — grid fees and taxes come on top and vary by supplier. Cite as "Voltcast Research, voltcast.com/guides/quant-features-ro-2026-08-18".
The monthly European power roundup
Negative-price records, the biggest spreads, which zones were hardest to forecast — every number computed from our production data, on the 2nd of each month. No filler, unsubscribe anytime.