An independent power trading firm in central London had 12 traders active across EPEX UK day-ahead and GB intraday markets. It had used one weather provider for four years when the contract came up for renewal in autumn 2024. The head of quantitative analysis asked whether alternatives should be assessed. After a half-hour discussion, nobody could answer from actual data whether the provider's temperature and wind speed forecasts performed well for its position-sizing models. There was a renewal invoice, but no skill score.
This is more common than trading desks like to admit. Weather data is bought as a utility, then seldom reviewed. Forecast quality, which affects every weather-linked trading model, is therefore governed by institutional habit rather than evidence.
How Weather Shapes Power Positions
In European day-ahead power markets, prices form where forecast demand meets forecast generation. Temperature dominates demand: during cold periods, UK residential and commercial electricity demand varies by roughly 1-1.5 GW per degree Celsius nationally. Wind speed and solar irradiance shape supply, and UK wind and solar capacity is now large enough for errors in either forecast to move day-ahead clearing prices materially.
On a cold day, a 2-degree temperature error on the steep part of the sensitivity curve can create a 2-3 GW national demand forecast miss. A desk exposed to that signal and using the wrong temperature forecast sizes positions against a demand model that differs from the settled market. P&L depends on model coupling and cash-out price deviation, but the link from forecast error to position error is direct.
UK offshore wind capacity regularly covers 20-35% of national demand on high-wind days. At that scale, a 10% error in total wind generation represents a 2-4 GW generation miss. Day-ahead traders using wind forecasts for supply estimates face the same position-error chain as those using temperature for demand.
Procurement Reality
Procurement or IT usually negotiates weather contracts, not the desk or quant team using the data. A commercial API subscription costs a few thousand to a few tens of thousands of pounds annually, small beside trading desk operating costs. It consequently receives little proportional scrutiny.
Preferred-vendor lists and multi-year contracts retain providers through renewal cycles. Switching brings integration costs: a new API format, pipeline changes and model change management. Those costs are real. Yet scrutiny is asymmetric: leaving a weak provider is examined closely, while the cost of staying accumulates unseen.
"No one has complained about the forecasts" is dangerous. It does not mean the forecasts perform well. Results blamed on market conditions may partly reflect systematic forecast error. Without validation against actuals, an organisation cannot distinguish the two internally.
What a Forecast Score Measures
Climingo presents provider accuracy as a skill score normalised against a climatological baseline. For temperature at a given location and horizon, it shows how much avoidable error versus naive climatology the provider removes. A skill score of 0.7 means 70% of that avoidable error is eliminated. At 0.4, the provider beats random performance, but substantial avoidable error remains.
Scores are reported by variable, including temperature, wind speed, solar irradiance and precipitation, by forecast horizon, 6h, 12h, 24h and 48h, and by region. For GB intraday exposure, the relevant measure may be 6h wind skill in the offshore zones involved. For a day-ahead desk, 24h temperature skill in zones feeding its demand model is the key metric.
National or Europe-wide provider accuracy is not the right procurement measure. If exposure is concentrated in northern Scotland while marketed figures use a UK average, the headline number and operational reality may differ sharply.
The Regional Accuracy Gap
A provider may score well across the UK yet perform consistently worse in the trading zone containing your exposure. Our benchmarking data shows this frequently enough to matter.
Northern Scotland is structurally harder to forecast than the English Midlands. Complex terrain, active Atlantic systems and strong mesoscale variation mean UK-wide metrics can understate performance in high-wind exposure zones. A desk in the Scottish interconnector corridor using a provider's UK-wide accuracy claim for position sizing may therefore overestimate the quality it receives.
Central European markets present the same issue. Bavaria and northern Germany have distinct weather regimes. A provider calibrated for Western European temperature may show systematic cold bias in continental winter conditions that its headline European figure hides.
Building a Scoring Framework
Teams building an internal validation pipeline need three components: provider forecast files archived at original issuance, not hindcasts; verified actuals for the relevant variable and location from a reliable metered or station source; and standard calculations covering at least MAE and RMSE per horizon.
Archiving at issuance is the hardest requirement. Most commercial weather APIs make it difficult to access the forecast file received 18 months ago. Some retain historical archives, many do not. This makes retrospective comparison harder than it sounds and makes continuous independent archiving valuable for procurement.
Partial validation is still better than none. 6 months of operational forecasts archived in your systems, where API responses are logged, can produce indicative skill scores and expose systematic bias. It cannot show seasonal variation, but can identify a provider performing substantially below alternatives in your market.
What Scores Do Not Show
Climingo's benchmark has a defined boundary that matters in procurement discussions. Higher forecast skill does not automatically produce better trading P&L. A forecast feeds a position-sizing model, whose logic determines how much signal becomes position size. A desk using forecasts as context will respond differently to improved skill than one whose positions closely follow forecast output.
Benchmark scores can show that systematic bias creates a structural drag on dependent positions. If wind generation is consistently underestimated on high-wind days, positions sized from that forecast are misaligned with market outcomes on those days. Measuring and removing the bias is the starting point. Any P&L improvement depends on your trading logic, not the benchmarking method.