Method and limitations

Everything on the live page is built from feeds that are free and need no account: Elexon Insights (settled System Price and net imbalance volume, market index price, accepted balancing actions, day-ahead indicated margin, system frequency, generation outturn by fuel, interconnector flows) and NESO wind and demand forecasts. No scraped or licensed market data is used.

Elexon publishes the settled price for a half-hour roughly fifteen to twenty minutes after that half-hour ends, so the two most recent periods are always unknown — those are exactly the ones being predicted.

The forecast is deliberately transparent rather than a black box. It has two parts.

Anchor. The traded exchange price for the period sets the level. The within-day market index (EPEX SPOT UK and Nord Pool, volume weighted, published by Elexon) normally prints for the period in flight within minutes, so the current-period forecast rides a genuine traded price; when only part of the usual volume has traded so far, that price is blended toward the estimate below in proportion to the volume behind it. Periods that have not traded yet take the freshest traded index and step it along the day-ahead auction shape — last traded price plus the auction’s price difference between that period and this one — fading toward the plain auction price the further ahead we look. Where no auction price exists either, the day-ahead price plus today’s basis is used, and failing that we carry the last traded price forward and widen the range to say so. The band tightens when the anchor is a well-traded index and widens when it is carried.

Gate closure. Trading for a Settlement Period stops fifteen minutes before it begins, so the price printed inside a period cannot be traded on for that period. It still matters for the forecast — it tells us what the market paid while the period ran — but any position must be struck earlier. The backtester enforces this by default: no plan is allowed to use a period’s own print, only prices struck before its gate closed. The in-flight forecast is therefore about your imbalance exposure and the periods you can still trade, not about trading the period you are standing in.

Spread. The System Price differs from that anchor mainly because the system is short or long. A piecewise linear model fitted on the last thirty days maps net imbalance volume onto that spread, separately for short and long conditions, because the two sides behave very differently. The model is fitted twice — once against the within-day anchor and once against the day-ahead anchor — so a forecast always carries the error band that belongs to the price it was built on. Further-ahead periods also have their imbalance estimate pulled back toward balanced, since imbalance mean-reverts.

For the in-flight period the imbalance volume itself is estimated from the last three settled periods, the ten-minute average system frequency, and the day-ahead indicated imbalance. Further out, frequency stops being informative and the forecast leans on the indicated position, which is why the range widens with each step.

The range shown is an 80% interval derived from the fitted model's own residual spread, scaled by horizon — not a guess.

Half-hourly imbalance prices are heavy-tailed. Extreme spikes come from a handful of very expensive offer acceptances late in a period, and no model anchored on average conditions catches all of them. Treat the upper bound of the range as the meaningful spike signal, not the central number.

Clock-change days have 46 or 50 settlement periods; those two days a year are handled approximately.

The gap the free feeds leave is live wholesale trading data — EPEX day-ahead and intraday prices, volumes and order book. The accuracy page compares the model against a baseline that simply uses the market index price. If the model's error stays close to that baseline, better trading data is the missing ingredient and a Montel/EnAppSys trial is worth running. If the error is well below it, the balancing-side signals are doing the work and a subscription buys little.