Daily guide · IT-SUD · 2026-09-02
The number your Italy South 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 number to beat this week is 21,77 EUR/MWh, the average error a naive persistence model carries across the five days in Italy South. That figure hides a wide spread. On 28 August persistence lands within 9,14 EUR/MWh, but by 31 August its error climbs to 39,14 EUR/MWh, more than four times as large. The quiet start-of-week days, with means of 204,01 and 206,08 EUR/MWh, are easy; the harder work sits at the back end.
For the reader, this means a flat weekly scorecard understates the real challenge. The day-mean path falls from 206,08 to 175,76 EUR/MWh by 30 August, then rebounds to 194,59 on 31 August. Persistence struggles precisely on those turning points, where its MAE jumps to 25,61 and then 39,14 EUR/MWh.
The concrete takeaway: any model claiming value has to earn it on 29 to 31 August, not on the calm opening. Our own 30-day track record sits at 21,61 EUR/MWh against a baseline of 23,70 across 361 days, so a small structural edge is realistic, but this week the margin lives in the volatile tail.
Before you train anything on IT-SUD 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:
21.77 €/MWh
persistence MAE on this week's IT-SUD auctions — the bar. A model that can't get under this adds negative value.
21.61 vs 23.7
our live production model vs its persistence baseline, trailing 361 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 |
|---|---|---|
| 2026-08-27 | 204.01 | 11.32 |
| 2026-08-28 | 206.08 | 9.14 |
| 2026-08-29 | 190.99 | 25.61 |
| 2026-08-30 | 175.76 | 23.62 |
| 2026-08-31 | 194.59 | 39.14 |
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/IT-SUD?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/IT-SUD
(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-it-sud-2026-09-02".
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.