Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: June 26, 2026
Key Takeaways
- Energy trading relies on eight core hourly forecast types: day-ahead price, intraday price, load, wind, solar, residual load, DART spread, and probabilistic ensembles. Each type maps to specific inputs, horizons, accuracy metrics, and settlement windows.
- Traditional NWP models update only 2–4 times per day on expensive supercomputers, which leaves traders exposed to stale data between runs and stuck in manual briefing routines during critical pre-market hours.
- AI foundation models like EPT-2 beat ECMWF HRES on every lead time for the four variables that drive P&L (10 m/100 m wind, 2 m temperature, surface solar radiation) while running at about 0.25 kWh and supporting up to 24 refreshes per day.
- High-frequency refresh combined with divergence and correction alerts turns model revisions into clear trading signals for intraday positioning and DART spread capture, which converts forecast accuracy gains into multi-million-euro annual savings for wind and solar portfolios.
- Book a personalized benchmark on the Jua platform to see EPT-2 against your current provider in under five minutes: schedule your demo here.
Why Weather Forecasts Drive Energy Trading
In Europe's weather-driven energy markets, traders now use AI and machine-learning tools to forecast the forecast. They focus on revisions in the reference models that reprice risk around heating demand, renewable output, and system tightness. Weather is the single largest driver of short-term electricity and gas prices. The gap between when a model updates and when a trader acts on it is where money is made or lost.
The existing forecasting stack does not close that gap. Numerical weather prediction (NWP) decomposes the atmosphere into three-dimensional grid cells and solves differential equations inside each one. These simulations run two to four times per day on supercomputers consuming roughly 8,400 kWh and €1,000–€20,000 per run. Between runs, traders work with stale numbers. The manual morning routine of downloading raw grib files, processing them through brittle in-house pipelines, and waiting for the meteorologist's briefing consumes the hours before the market opens.
Jua is a foundation model and agent company. Its EPT (Earth Physics Transformer) family is a general spatiotemporal transformer foundation model that learns the governing physics of complex systems directly from observational data. Athena is an AI agent that plans, reasons, and calls tools to turn natural-language objectives into deliverables. Jua for Energy is the first applied product built on both. It delivers atmospheric forecasts that outperform leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE. Forecasts refresh up to 24 times per day at up to 1 km resolution in the product surface.
How Energy Market Forecasts Work
An energy market forecast is a quantitative estimate of a future state variable such as price, load, generation, or system balance. It is produced at a defined horizon and resolution to support a specific trading or operational decision. The forecast type sets the inputs, the model class, the accuracy metric, and the settlement window.
The European Centre for Medium-Range Weather Forecasts' two-week outlook is the definitive reference point for traders repricing risk around heating demand, renewable output, and system tightness. Price forecasts depend on generation forecasts. Generation forecasts depend on weather forecasts. The dependency chain runs from atmospheric physics to settlement price, and an error at any layer propagates through the rest.
Modern energy market forecasting draws on five input categories. Meteorological variables include wind speed at hub height, surface solar radiation, temperature, and precipitation. Installed-capacity data covers turbine and panel fleets by zone. Grid topology and constraint data capture transmission limits and interconnector flows. Market microstructure signals include auction results, imbalance prices, and fuel spreads. Historical observational data supports model calibration and backtesting. The accuracy of the atmospheric layer, the first input, sets the ceiling on every downstream forecast.
Executive Summary and Evaluation Lens
Evaluating an hourly forecasting platform for energy trading rests on five criteria that you apply together. Model capability asks whether the deterministic flagship beats ECMWF HRES on RMSE across the 0–240 hour range and whether the ensemble beats the ECMWF ENS mean on CRPS. Operational usability asks whether the platform auto-generates briefings, alerts, and power forecasts without manual assembly. Reliability asks whether outputs are physically constrained, peer-reviewed, and benchmarked against ground-truth observations rather than reanalysis. Scalability asks whether the platform can refresh up to 24 times per day without HPC infrastructure. Integration fit asks whether a REST API with Apache Arrow support and a Python SDK provide programmatic access to hindcasts, ensembles, and 25+ models under a unified schema.
EPT-2, documented in arXiv:2507.09703, satisfies all five criteria. It outperforms ECMWF HRES on every lead time and on 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation across the full 0–240 hour range. EPT-2e, documented in arXiv:2410.15076, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. A single EPT-2 inference runs at approximately 0.25 kWh and $0.20–$15 on a single GPU, which is roughly four orders of magnitude cheaper than a traditional NWP simulation.
How Traders Turn Hourly Forecasts into P&L
Hourly forecasts map directly to the settlement windows that define a trader's P&L. In the day-ahead market, traders submit generation and consumption bids for each hour of the following day before the auction closes, typically at noon. The accuracy of the weather forecast at that moment determines whether the bid reflects actual generation capacity or leaves imbalance exposure on the table.
In the intraday continuous market, traders adjust positions as new information arrives. A wind ramp that the morning run misses becomes a trading opportunity for whoever sees the model revision first. Divergence between two models on a key variable, such as 100 m wind over a wind-rich zone or surface solar radiation over a high-penetration solar region, acts as a signal rather than noise. On the Jua platform, divergence alerts fire the moment two models disagree, and correction alerts fire the moment a model revises its own output. The trade window opens with a notification.
A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves about €1.5 M per year, which scales to hundreds of millions for large portfolios. A 1 GW solar portfolio at the same accuracy gain saves approximately €3 M per year.
Energy Forecasting Providers and Jua’s Role
The energy forecasting landscape divides into three categories. NWP incumbents such as ECMWF HRES, ECMWF ENS, NOAA GFS, and DWD ICON are the universal benchmarks. They run on supercomputers, update two to four times per day, and have defined the accuracy standard for forty years. AI weather peers such as Microsoft Aurora, Google DeepMind GraphCast, and ECMWF AIFS are research-grade models consumed as raw outputs without productised ensembles, operational refresh schedules, or analyst layers. Point-solution SaaS vendors and meteorology consultancies resell processed NWP outputs or produce analyst reports without owning an underlying model or running cross-vendor benchmarks.
Jua for Energy sits in a different category as a foundation model and agent company's first applied product. EPT and Athena are horizontal and domain-agnostic by architecture. The Jua platform exposes more than 25 models, including 10 proprietary AI models from the EPT family and 15 third-party NWP and AI models such as ECMWF HRES, ECMWF ENS, ECMWF AIFS, NOAA GFS, DWD ICON Global, DWD ICON-EU, Microsoft Aurora, and GFS GraphCast, through a single schema and a single API. Aurora and GraphCast run on the Jua platform as guests on the comparison surface. The benchmark is built into the product.
Compare Jua’s multi-model surface against your current stack in a live session. Book a demo.
Day-Ahead vs. Intraday Hourly Forecasting
Day-ahead and intraday forecasting differ in horizon, input freshness, and the decisions they support. Day-ahead forecasting targets the 12–36 hour window before physical delivery. The primary inputs are the overnight NWP runs, such as ECMWF's 00Z and 12Z cycles, combined with installed-capacity data and market microstructure signals. Accuracy at this horizon is dominated by the quality of the atmospheric model. A one-percentage-point improvement in wind forecast accuracy at 24 hours translates directly into reduced imbalance exposure at settlement.
Intraday forecasting targets the 15-minute to 6-hour window. At this horizon, the value of forecast refresh frequency dominates. EPT-2 RR updates up to 24 times per day, while traditional NWP updates two to four times. The trader who sees a wind ramp revision at hour 14 of a 24-run cycle acts before the trader who waits for the next 6-hourly NWP run. EPT-2 HRRR delivers the same high-cadence refresh at up to 5 km native resolution over Europe, which provides the spatial granularity required to resolve local orographic effects on wind generation and cloud-shadow effects on solar output.
The Jua platform's power forecast surface covers solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load across Germany, Great Britain, France, the Netherlands, and Belgium. The Actual Generation model refreshes every 15 minutes with a 48-hour horizon. The Fundamental model runs out to 20 days.
Key Physical Drivers and System Components
Four physical drivers govern hourly energy price formation. Wind speed, particularly at hub height between 80 m and 160 m for modern turbines, drives wind generation. Surface solar radiation determines photovoltaic output. Near-surface temperature drives heating and cooling demand. Precipitation affects hydro dispatch and gas demand. Each driver requires a specific atmospheric variable, spatial resolution, and forecast horizon to become actionable.
EPT-2 covers 25 variables, including wind at 11 height levels from 10 m to 200 m, which spans the full range of commercial turbine hub heights. It produces forecasts at native any-Δt, which means it is trained to predict at arbitrary time steps rather than rolling forward in fixed 6-hour increments. Aurora and most AI peers roll forward in 6-hour steps and compound error at each step. EPT-2 does not roll. The architecture learns physics, and the domain becomes a variable.
Load forecasting adds calendar effects such as hour-of-day, day-of-week, and public holidays, along with economic activity proxies and temperature-demand elasticity, to the atmospheric inputs. Residual load, defined as total load minus wind and solar generation, determines thermal dispatch and gas positioning. Errors in wind and solar forecasts propagate directly into residual load errors and from there into gas spread positioning.