Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 10, 2026
Key Takeaways for Energy Traders
- EPT-2 delivers the most accurate atmospheric forecasts in production, outperforming ECMWF HRES on every lead time from 0 to 240 hours across wind, temperature, and solar radiation variables.
- Jua for Energy provides up to 24 daily model updates and 15-minute power forecast refreshes, compared to the 2–4 daily runs typical of legacy NWP systems.
- Athena, Jua’s AI agent, converts natural-language queries into trading briefings and benchmarks in approximately 90 seconds, replacing manual grib processing and morning prep routines.
- Native power forecasts for solar, wind, and load across five European countries are available through the Jua platform, REST API, and Python SDK with direct ENTSO-E integration.
- Book a demo with Jua to see EPT-2 head-to-head against your current forecast provider and unlock measurable P&L gains in energy trading.
2026 Accuracy Benchmarks vs ECMWF HRES, Aurora, and GraphCast
The July 2026 EPT-2 technical report (arXiv:2507.09703) evaluates forecast skill against more than 10,000 real ground stations using StationBench, an open-source benchmarking methodology with no post-processing or station fine-tuning. RMSE (root mean square error, the average magnitude of forecast error, lower is better) and CRPS (continuous ranked probability score, a measure of probabilistic forecast sharpness and reliability, lower is better) are the evaluation metrics. Lead time refers to the number of hours ahead of the forecast initialization that a prediction covers.
The results across the four variables that drive energy P&L, 10 m wind speed, 100 m wind speed (hub height for most onshore turbines), 2 m temperature, and surface solar radiation downward (SSRD), are unambiguous across the full 0–240 hour range. The table below shows EPT-2’s performance against each competing model across all four variables, demonstrating consistent superiority.
| Variable | EPT-2 vs ECMWF HRES (0–240 h) | EPT-2 vs Microsoft Aurora (0–240 h) | EPT-2e vs ECMWF ENS mean |
|---|---|---|---|
| 10 m wind speed | EPT-2 beats HRES on every lead time | EPT-2 beats Aurora across full range | EPT-2e beats 50-member ENS mean on RMSE and CRPS at virtually every lead time |
| 100 m wind speed | EPT-2 beats HRES on every lead time | EPT-2 beats Aurora across full range | EPT-2e beats 50-member ENS mean on RMSE and CRPS at virtually every lead time |
| 2 m temperature | EPT-2 beats HRES on every lead time | EPT-2 beats Aurora up to ~130 h; Aurora has no SSRD output | EPT-2e beats 50-member ENS mean on RMSE and CRPS at virtually every lead time |
| SSRD (surface solar radiation) | EPT-2 beats HRES on every lead time | EPT-2 wins by default, Aurora publishes no SSRD output | EPT-2e beats 50-member ENS mean on RMSE and CRPS at virtually every lead time |
EPT-2e is the ensemble variant of EPT-2. An ensemble is a set of perturbed model runs used to quantify forecast uncertainty and produce probabilistic outputs. EPT-2e beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. The financial translation is direct: a 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 million per year in European energy markets, with solar portfolios at the same accuracy gain saving ~€3 million per year.
Update Frequency and Operational Refresh for Trading Desks
NWP (numerical weather prediction, the method of solving differential equations across a three-dimensional atmospheric grid) has operated on a hard compute ceiling for forty years. Government-backed large-scale NWP models generally refresh a few times per day, with most legacy systems running on 6-hour cycles. A single NWP simulation consumes ~8,400 kWh and costs €1,000–€20,000 on HPC infrastructure, which caps operational cadence at 2–4 runs per 24 hours.
EPT-2 inference runs on a single GPU in minutes at ~0.25 kWh and $0.20–$15 per simulation, which is roughly four orders of magnitude cheaper. That cost asymmetry makes 24× daily refresh operationally viable. The following table breaks down how these economics translate into operational advantages across update frequency, resolution, and cost.
| Capability | Jua for Energy (EPT family) | ECMWF HRES / ENS | Aurora / GraphCast (AI peers) |
|---|---|---|---|
| Deterministic update frequency | Up to 24×/day (EPT-2 RR); 4×/day (EPT-2 flagship) | 2–4×/day | Typically 4×/day research cadence, no productised operational schedule |
| Ensemble update frequency | 4×/day (EPT-2e) | 4×/day (ENS, 50 members) | No productised ensemble equivalent |
| Power forecast refresh | Every 15 minutes (Actual Generation); 20-day horizon (Fundamental Model) | Not a native product | Not a native product |
| Inference cost per simulation | ~$0.20–$15, ~0.25 kWh, single GPU | ~€1,000–€20,000, ~8,400 kWh, HPC cluster | Similar order of magnitude to Jua for inference |
| Native spatial resolution (models) | Up to ~5 km (EPT-2 HRRR, Europe) | 9 km (HRES) | ~25 km at published resolution |
Jua for Energy’s EPT-2 RR delivers up to 24 updates per day, and actual-generation power forecasts refresh every 15 minutes at up to 1 km resolution on the product surface. Between those updates, divergence alerts and correction alerts fire the moment any model in the 25+ model fleet revises its output, so traders are not waiting for the next scheduled run to learn that the market has moved.
Workflow Replacement: From Grib Files to Athena Briefings
Those frequent updates create value only when traders can act on them quickly, which is where Jua for Energy removes the workflow bottleneck. Grib (gridded binary) files are the raw output format of NWP models, binary archives of atmospheric variables on a three-dimensional grid that require specialist software and in-house pipelines to decode and process. The standard energy-trading morning routine begins at 6 a.m. with grib downloads, proceeds through brittle in-house processing scripts, and ends, if the meteorologist is available, with a briefing assembled from a dozen sources. By the time a coherent view of the day exists, the market has already moved.
Athena, Jua’s AI agent currently instrumented with the Jua for Energy tool surface, replaces that sequence. A trader types a natural-language objective, for example, “what is the 100 m wind forecast spread across models for northern Germany tonight?” Athena then plans, calls the relevant forecast and benchmarking tools, evaluates intermediate outputs, and returns a written briefing with the underlying widget. Athena converts EPT-2 physics predictions into trading decisions by reading market context and modeling likely participant behavior. Typical queries resolve in approximately 90 seconds. Backtests, hindcast runs against years of archived forecasts used to evaluate a strategy against past data, complete in approximately 5 minutes.
Day-Ahead and Intraday briefings auto-refresh on every new model run, covering model consensus across 25+ models, model delta since the previous run, convergence tracking as lead time shortens, and price implications. Because all of this information updates automatically in a single workspace, the 7–9 a.m. manual prep routine, previously spent aggregating data from multiple sources, compresses into simply opening that workspace before the market does. The time savings are significant enough that trading houses and quant desks describe Athena as “another headcount, for free.”
Power Forecasts and Energy-Specific Integrations
Jua for Energy ships native power forecasts, not weather variables translated post-hoc into generation estimates, but forecasts built from EPT-2 atmospheric outputs combined with installed-capacity data and grid topology. Coverage spans solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load across five European countries: Germany, Great Britain, France, the Netherlands, and Belgium. New country coverage is added on a weekly basis.
Two complementary models run on the same surface. The Fundamental Model combines the EPT-2 weather forecast with installed-capacity data and runs out to 20 days. The Actual Generation Model refreshes every 15 minutes with a 48-hour horizon and lower error in the near term. Both are available through the Jua platform workspace and programmatically via the REST API and Python SDK.
For quant developers and engineering teams, pip install jua installs the Python SDK from PyPI. The REST API exposes 25+ models, 10 proprietary AI from the EPT family plus 15 third-party NWP and AI models including ECMWF HRES, ECMWF ENS, ECMWF AIFS, NOAA GFS, Microsoft Aurora, and GFS GraphCast, through a single unified schema with Apache Arrow support for large payloads. ENTSO-E grid data integrates directly for European power-market data. The integration that takes a quant team a quarter to build elsewhere stands up in days. Documentation is at docs.jua.ai.
Jua serves major utilities across five continents, including some of Europe’s largest energy companies, as well as commodity traders and hedge funds, including Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec.
Evaluation Framework for Meteorologists and Quant Teams
The live benchmarking surface on the Jua platform puts 25+ models on a single screen. A meteorologist evaluating Jua for Energy during procurement selects their most stakes-relevant region and variable, typically a wind-rich region of the home market, and runs a head-to-head accuracy comparison against EPT-2 and any third-party model on the platform. Results return in seconds. The same surface is available post-procurement for ongoing model surveillance.
For quant teams, hindcast data is available across multiple Jua and third-party models. A backtest against years of historical forecasts runs in approximately 5 minutes via Athena, or directly through the SDK for teams that prefer programmatic access. The evaluation methodology, StationBench, against more than 10,000 real ground stations with no post-processing or station fine-tuning, is open-source and reproducible. EPT-2 results are published in the peer-reviewed technical report at arXiv:2507.09703.
Jua for Energy does not replace ECMWF. Most serious customers keep their ECMWF subscription and run Jua for Energy alongside it. ECMWF AIFS, ECMWF’s own AI model, runs on the Jua platform as a guest model, available in the same workspace as EPT-2. What Jua for Energy displaces is the plumbing around the incumbent feed, the in-house grib pipeline, the manual benchmarking, the morning-briefing analyst, and the dashboard stitching. NOAA’s WFIP-driven improvements in wind forecast accuracy at turbine heights saved utility companies more than $95 million per year, so the financial case for accuracy improvements at hub height is well-established. EPT-2’s margin over HRES on 100 m wind across the full 0–240 hour range represents a material and measurable edge on top of that baseline.
Frequently Asked Questions
How do 2026 EPT-2 results compare to ECMWF HRES on wind and solar variables?
EPT-2 outperforms ECMWF HRES on every lead time from 0 to 240 hours across all four energy-critical variables, 10 m wind speed, 100 m wind speed, 2 m temperature, and surface solar radiation downward (SSRD). The evaluation is conducted on StationBench, an open-source methodology benchmarked against more than 10,000 real ground stations with no post-processing or station fine-tuning. Results are published in the peer-reviewed technical report at arXiv:2507.09703. EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. ECMWF HRES remains the universal industry benchmark and Jua for Energy runs alongside it, so the comparison is now a comparison, not a hierarchy.
What update frequency do energy traders actually receive from legacy NWP?
Legacy NWP systems, including ECMWF HRES and NOAA GFS, run their full global algorithm twice per day, with smaller supplementary runs bringing the total to approximately 4 updates per 24 hours. The compute cost of a single NWP simulation makes higher cadence economically infeasible at scale, because each run requires HPC infrastructure that is orders of magnitude more expensive than GPU-based inference. Between runs, traders are working with stale numbers. EPT-2 RR updates up to 24 times per day, and actual-generation power forecasts on the Jua platform refresh every 15 minutes. A typical Jua run also completes approximately 2.5 hours ahead of competing operational runs at the same cycle, so customers see the next forecast before the market does.
Can Athena replace internal meteorology teams?
Athena is an AI agent instrumented with the Jua for Energy tool surface. It turns natural-language objectives into briefings, benchmarks, backtests, and custom widgets, and delivers most queries in under two minutes and complex backtests in about five. For trading houses and quant desks without dedicated meteorologists, Athena functions as an analyst that works continuously without manual briefing production. For organizations with internal meteorology teams, Athena handles the repetitive production work, daily briefings, model delta summaries, and convergence tracking, which frees meteorologists to focus on deeper forecast research and desk-specific downscaling. Athena does not replace the judgment of a trained atmospheric scientist; it removes the manual assembly work that consumes a disproportionate share of that scientist’s time.
Does Jua replace ECMWF or run alongside it?
Jua for Energy runs alongside ECMWF, not in place of it. Most serious customers maintain their ECMWF subscription and add Jua for Energy to the stack. ECMWF AIFS, ECMWF’s own AI forecasting system, runs as a guest model on the Jua platform, available in the same workspace as EPT-2 and EPT-2e. What Jua for Energy displaces is the infrastructure around the ECMWF feed, the in-house grib processing pipeline, the manual benchmarking harness, the morning-briefing production routine, and the fragmented dashboard stack assembled from a dozen vendor contracts. The result is a single workspace where ECMWF, EPT-2, Aurora, GraphCast, AIFS, and 20+ other models are on the same screen, under a unified schema, refreshed on the same cycle as the underlying physics.
Conclusion
The 2026 StationBench results documented in arXiv:2507.09703 establish EPT-2 as the global state of the art in atmospheric prediction. The benchmark results show EPT-2’s consistent superiority over ECMWF HRES across all lead times and energy-critical variables, and a clear advantage over Aurora on 10 m wind, 100 m wind, and 2 m temperature across the full range. The ensemble variant matches this performance advantage against the 50-member ECMWF ENS mean. EPT-2 RR updates up to 24 times per day against the 2–4 daily runs available from legacy NWP. The estimated P&L impact is €1.5 million per gigawatt annually in European energy markets, which scales linearly across multi-GW portfolios to hundreds of millions of dollars.
Jua is a foundation model and agent company. EPT is a general physics foundation model and Athena is an AI agent. Jua for Energy is the first applied product, in the same way Anthropic relates to Claude Code. The Jua platform pairs those accuracy gains with Athena’s natural-language analyst layer, 15-minute power forecast refresh, live cross-model benchmarking, and a Python SDK that stands up in days. The 7–9 a.m. manual prep routine becomes a single workspace open before the market does. You act before the market does.
See EPT-2 head-to-head against your current forecast provider. Book a demo.