nordic-price-forecast

Nordic + NL Electricity Price Forecast — 13 bidding zones

Day-ahead electricity price forecasts 2 to 9 days out, hourly, in EUR/MWh, for every bidding zone in Denmark, Norway, Sweden, Finland and the Netherlands:

DK1, DK2 · NO1, NO2, NO3, NO4, NO5 · SE1, SE2, SE3, SE4 · FI · NL

Retrained and re-uploaded daily, so the weights here are what production runs. A companion model specialised to Denmark alone lives at Elpriser/denmark-price-forecast.

Design: one pooled model, not thirteen

A single LightGBM quantile model covers all 13 zones, with zone as a categorical feature. The zones share nearly all their structure — the same calendar effects, the same weather physics, the same lag behaviour — so pooling lets a thin zone borrow strength from a busy one, and leaves one thing to maintain instead of thirteen.

Pooling only works because the model can identify a zone's price level before it starts explaining deviations. NO4 averages 22 EUR/MWh and NL 90; a pooled model on raw prices would otherwise spend its capacity rediscovering that. Three features carry it: the zone id, last_day_mean, and level_30d (the zone's own trailing 30-day average, measured to the last known day).

Read the metrics carefully — most of them lie across zones

This is the single most important thing to understand about this model.

Percentage MAE punishes cheap zones. NO4's mean absolute error is 14.1 EUR/MWh — lower than DK1's 25.6 — but it reads as 64% because NO4's mean price is only 22 EUR/MWh. Nothing is wrong with the forecast there; the denominator is small.

Min-hour hit-rate measures the daily price spread, not forecast skill. Hit-rate tracks how far apart a day's cheapest and dearest hours are:

When the day's hours differ by 12 EUR/MWh, ranking them is close to random — and getting it wrong costs almost nothing.

Use regret instead. Regret is what following the forecast's three cheapest hours costs versus perfect timing. It stays meaningful whatever the zone's price level or spread.

Performance

Rolling-origin backtest, monthly retrain, tested April 2025 → July 2026.

"Captured" is the share of the available timing saving the forecast captures versus running at a random hour.

Two things to take from this. The model beats a seasonal baseline in all 13 zones on regret. And absolute regret is strikingly uniform — 6 to 8 EUR/MWh almost everywhere, FI's 10.7 the only outlier — so in terms of what a user actually loses by following it, the model performs comparably across a market where mean prices differ fourfold. The lower "captured" figures in NO3, NO4, SE1 and SE2 reflect there being less to capture, not worse forecasts.

Pooled MAE 22.2 EUR/MWh, against 25.7 for the seasonal baseline and 28.2 for persistence. MAE by horizon rises gently from 20.1 (h=2) to 23.6 (h=9).

Features

30 per row, at horizon h (2-9):

• Zone & calendar — zone (categorical), h, hour, weekday, month, is_weekend, doy_sin, doy_cos
• Price history — lag1/2/3/7/14/21/28, filled only when known at issue time (k >= h-1), seasonal4w, last_day_mean, level_30d
• Own-zone weather at the right lead — wind, rad, temp, precip, precip_24h
• Regional state — reg_wind, reg_temp, hydro_precip, hydro_precip_7d
• Hydro reservoirs — reservoir_pct, reservoir_anom, nordic_reservoir_anom

Weather comes from Open-Meteo's previous-runs archive: the forecast as issued N days before delivery, so training rows carry the same information quality the live model gets. Precipitation is included because NO and SE prices are set by reservoir hydrology, where rainfall is a genuine driver in a way it is not in Denmark.

Why reservoirs are in

The first backtest had the model losing to a plain seasonal average in NO1 (+1.3%) and NO2 (+8.1%) — the signature of a missing level driver. Nordic prices follow water value: producers hold water back when reservoirs are low and release when full. Adding weekly reservoir levels (ENTSO-E A72, lagged 14 days, expressed as share-of-maximum and as deviation from the zone's median for that week of year) turned NO1 to −1.7% and NO2 to −0.3%. The Nordic aggregate goes to every zone including DK and NL, because the market is coupled and a dry Norwegian year lifts prices well beyond Norway.

Post-processing

Shape blend. The model's daily level is kept, its intraday shape blended 50/50 with a 4-week seasonal profile. Carried over from the Danish model, where that was measured to beat both parents on both accuracy and hour-ranking.

Bands are additive empirical residual quantiles, not scaled model quantiles. Scaling was tried first, as the Danish model does it, and broke down: DK, NL and SE3/SE4 needed a sensible ~1.8-2.0x, but NO4 needed 8.0x and still missed 80% coverage, NO5 up to 7.5x, SE1/SE2 up to ~5x. A band inflated eightfold is not measuring uncertainty, it is being stretched to pass a check. Taking the 10th and 90th percentiles of actual residuals per zone and horizon gives correct coverage by construction.

Limitations

• The bands are only informative in the structured zones. At 80% coverage the band width is 70-96% of the mean price in DK, NL, NO1/NO2/NO5 — wide but usable. In SE1 and SE2 it is 243-253%, and in FI and NO4 around 200%. Those zones sit near a stable level and then spike hard, so an honest 80% interval is nearly uninformative about level. The point forecast and the hour-ranking still work there; the interval does not.
• One weather point per zone. Deliberate, to keep 13 zones maintainable, but coarser than the 14-point setup the Denmark-only model uses. Large zones like NO4 and SE1 are poorly represented by a single coordinate.
• No interconnector or congestion modelling. The Denmark-only model includes month-ahead transfer capacity; this one does not.
• Reservoir data is weekly and lagged. Deliberately shifted 14 days to stay clear of publication timing, so fast hydrological swings arrive late.
• The autumn fall-back hour is averaged away. Prices are keyed on a naive local timestamp, which cannot hold the two 02:00s of the autumn clock change, so those days carry 24 hours instead of 25. Two hours per zone per year. Removing it means re-keying the pipeline on UTC.
• Not financial or trading advice. These are wholesale day-ahead prices, not retail bills, which add grid tariffs, taxes and VAT.

Data correction, August 2026

Every price series was refetched after a parser bug was found. ENTSO-E publishes day-ahead prices as curveType A03 — variable-sized blocks, where a point holds until the next published position — and the parser read only the explicit positions. Each hour whose price repeated the hour before it was therefore dropped: 2,569 hours, and not at random, because the loss scaled with how flat a zone is (2.8% of NO4 against 0.34% of DK1). Expanding the blocks also corrected the hourly mean under the 15-minute market design introduced in October 2025, where the old code weighted each published block equally regardless of its length; that moved 4,656 further hours by 0.27 EUR/MWh on average.

The correction was verified against a source outside this code path: Denmark's correlation with Energinet's independent EDS feed rose from 0.99999 to 1.00000, so the restored hours match an external record rather than being interpolated.

It did not make the model more accurate. Retrained and backtested on the corrected data, MAE rose 1.6% while regret — the cost of following the forecast's three cheapest hours — improved from 5.28 to 5.26 EUR/MWh and the hit-rate fell from 51.7% to 51.3%. The training is deterministic, so these are real differences rather than run-to-run noise; they are simply small. The value of the fix is that the data and the published prices are correct, not that the forecast improved.

Intended use

Consumer and household decision support — when in the coming week to run flexible load — and research. Not validated for trading or hedging.

Sources & license

CC BY 4.0. Prices and reservoir levels: ENTSO-E Transparency Platform. Weather: Open-Meteo. Code: elpriser.org on GitHub, scripts/forecast_model/nordic/.