Methodology
What this site claims — and what those claims rest on
The front page shows numbers without error bars, because that is not where they belong. This page has what they were measured against: which baseline, how large a sample, and what is deliberately not claimed.
Every figure is fixed with the date it was measured rather than fetched live — a methodology page that silently changes is not a methodology page. — 3 September 2026.
Price forecast
Testing timing, not price levels
The model does not predict what electricity will cost. It predicts which part of a day is cheaper — and that is measured through the decision: how much someone loses by scheduling a two-hour load on the model, against someone who knew the prices in advance.
Point-in-time audited: the ordering model uses only lagged A44 prices (the latest visible delivery day is target D−2) and calendar features. A44 is publish-once, so a later pull returns the issued value. A69 and A65 are not inputs to this model, so the revision trap does not apply to it.
The same model's level error is 44–51 €/MWh against a mean near 82 €/MWh. The same model's level error is 55–62% of the mean. That is why no forecast euro figure appears anywhere on the site: the ordering is usable, the level is not. These are held-out backtest figures, not a live record — no public live record exists yet, and none is presented as if it did.
Letting weather touch the intraday shape made the resulting decision 27–40% worse; confined to the daily level it gave an 11–16% improvement. The model therefore uses weather for the level only, never the shape.
Solar forecast
Here the number may be published
The same weather input forecasts solar generation well. Ground truth is ENTSO-E A75 measured Estonian solar output (July 2026); the predictor is Open-Meteo tilted irradiance from the run issued N days earlier. Fitted on the first half of the month, scored on the second. Sample: 744 hours, peak 871 MW.
| Lead | MAE (daylight) | % of peak | Persistence baseline | r |
|---|---|---|---|---|
| D+1 | 77.5 MW | 8.9% | 91.2 MW | 0.908 |
| D+2 | 78.4 MW | 9.0% | 91.2 MW | 0.904 |
| D+3 | 73.4 MW | 8.5% | 91.2 MW | 0.906 |
The error is small against capacity, beats the persistence baseline, and does not grow from D+1 to D+3. That is the opposite of price, where skill is also flat across horizons — but because there is nothing there to decay.
This experiment evaluates a historical national-generation model. The homepage’s rooftop estimate uses Puhub irradiance at a different input and spatial scale. Its accuracy has not yet been validated against real roofs; the national experiment’s error percentage does not transfer to it.
Validation
Against instruments, not against another model
Hourly radiation observations from the Estonian Environment Agency are archived immutably through a public interface (Puhub). Eight stations measure global horizontal irradiance with pyranometers; Tartu-Tõravere belongs to the international BSRN network. Forecast and observation are compared on the same plane — horizontal against horizontal, not against a tilted surface, or the comparison would measure geometry rather than skill. The archive begins 3 September 2026.
Publishing gate: at least 30 days and 1,200 daylight point-hours, and any published figure must name the window it was measured over. Estonian sun elevation changes so much across the year that a September figure does not describe December.
Market observation
The hour is too coarse a unit
The European day-ahead market settles in quarter-hours, and Estonian prices genuinely move within the hour: measured on the date below, all four quarters differed from each other in all 24 hours of both published days. The median within-hour spread was 15–23 €/MWh and the largest 137 €/MWh — at 20:00 the price moved from 21.1 to 158.2 €/MWh inside one hour, or 87% of that hour's peak.
That has a practical consequence: hourly scheduling leaves money on the table even when it picks the right hour. Both the price chart and the forecast strip on this site are therefore 15-minute, not hourly.
Data
Sources and their lag
| Source | Used for | Lag |
|---|---|---|
| ENTSO-E A75 | generation mix, carbon intensity | ≈ 2 h |
| ENTSO-E A11 | sea-cable direction and magnitude | 2–5 days |
| Nord Pool / Elering | exchange price, windows | published D−1 ~12:45 CET |
| Elering | Estonian consumption | ≈ 1–1.5 h |
| Open-Meteo | solar and wind forecast | model run, cached 30 min |
| Keskkonnaagentuur / Puhub | forecast validation | 1 h |
Distribution-tier (10–35 kV) flows are modelled, not measured — ENTSO-E publishes no per-line data for them, and the map says so itself. Transmission 330 kV flows are a model solution over measured inputs.
Forecast model verification record
The full research record includes published results, pending experiments and findings of no improvement. These experiments are not the homepage’s published market prices or evidence of household savings.
Checking the evidence record…
Three sites
What lives where
- puhub.ee ↗Weather and measurement — observations, forecasts, the radiation archive.
- wattography.com ↗European grid and flow analytics — 44 bidding zones, a public accuracy record.
- ElekterThis site: both combined, from an Estonian consumer's point of view.