Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 9, 2026
Key Takeaways for Energy and Trading Teams
- EPT-2 beats ECMWF HRES on every lead time and energy-critical variable, setting a new global benchmark for forecast accuracy.
- Jua for Energy delivers up to 24 forecasts per day at a fraction of traditional NWP compute cost, removing the stale-run constraint that has limited traders for decades.
- The Athena AI agent turns raw forecasts into briefings, backtests, and dashboards in under two minutes, replacing manual analyst workflows for utilities and trading desks.
- Customers including Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec already rely on Jua for daily trading decisions across four continents.
- Run a live benchmark against your current provider in under five minutes and see the accuracy gains firsthand.
Latest Accuracy Benchmarks vs ECMWF and Aurora
The benchmarks below compare EPT-2 against ECMWF HRES and Microsoft Aurora using RMSE (Root Mean Squared Error), a measure of average forecast error where lower values indicate better performance. The evaluation uses Jua's open-source StationBench methodology against more than 10,000 real ground stations, with no post-processing or station fine-tuning applied.
| Variable | EPT-2 vs ECMWF HRES (0–240 h) | EPT-2 vs Microsoft Aurora (0–240 h) | EPT-2e vs ECMWF ENS mean (RMSE & CRPS) |
|---|---|---|---|
| 10 m wind speed | EPT-2 wins every lead time | EPT-2 wins full 0–240 h range | EPT-2e wins virtually every lead time |
| 100 m wind speed | EPT-2 wins every lead time | EPT-2 wins full 0–240 h range | EPT-2e wins virtually every lead time |
| 2 m temperature | EPT-2 wins every lead time | EPT-2 wins up to ~130 h and beyond | EPT-2e wins virtually every lead time |
| Surface solar radiation (SSRD) | EPT-2 wins every lead time | EPT-2 wins by default, Aurora has no SSRD output | EPT-2e wins virtually every lead time |
EPT-2 produces forecasts at native any-Δt, which means it predicts at arbitrary time steps instead of rolling forward in fixed 6-hour increments. Aurora and most AI peers roll forward in 6-hour steps, which compounds error across the forecast horizon. EPT-2 does not roll. EPT-2 delivers hourly global weather updates and outperforms leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE.
Test EPT-2 on your own region and variable against 25+ models in under five minutes.
High-Frequency Refresh Cadence Replaces Stale NWP Runs
Traditional NWP runs on a hard compute ceiling. A single NWP simulation consumes approximately 8,400 kWh and costs €1,000–€20,000 on high-performance computing infrastructure. These economics cap update frequency at two to four global runs per day, a constraint the energy industry has operated under for forty years. In Europe’s weather-driven energy markets, traders are turning to AI and machine-learning tools because stale NWP runs leave them reacting to weather only after it has already shown up in the price.
A single EPT-2 inference runs on a single GPU in minutes at approximately 0.25 kWh and $0.20–$15 per simulation. This profile is roughly four orders of magnitude cheaper than an equivalent NWP run and makes high-cadence refresh operationally viable.
The Jua for Energy refresh architecture operates at three cadences:
- EPT-2 RR (rapid refresh): up to 24 runs per day, delivering a new deterministic global forecast every hour.
- EPT-2e (ensemble): 4 runs per day, with 10 ensemble members beating the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time.
- Actual Generation power forecasts: refresh every 15 minutes with a 48-hour horizon across Germany, Great Britain, France, the Netherlands, and Belgium.
Intraday trading benefits directly from this cadence. Renewable energy systems require quarter-hourly or even minute-by-minute forecast updates because hourly forecasts are often insufficient for intermittent wind and solar generation. EPT-2 RR’s 24×/day cadence narrows the gap between the physics and the trade window. Power forecasts on the Jua platform cover solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load across all five covered countries, with a Fundamental Model running out to 20 days and natively forecasting at up to 5 km resolution over Europe.
A typical Jua run also completes approximately 2.5 hours ahead of competing operational runs at the same cycle. Customers who run Jua for Energy alongside existing NWP subscriptions receive the next forecast before the next traditional run lands. That speed advantage creates a new operational challenge, because trading teams must extract actionable intelligence from 24 forecast updates per day without drowning in data.
Athena Natural-Language Workflow for Quant and Trading Teams
Athena is Jua's AI agent built to solve that workflow problem. It is instrumented with the Jua for Energy tool surface and accepts a natural-language objective, plans, calls tools, evaluates intermediate outputs, and resolves to a deliverable. Typical queries resolve in approximately 90 seconds, and backtests complete in approximately 5 minutes.
Representative query examples from the Jua for Energy tool surface include:
- “What is the 100 m wind forecast spread across models for northern Germany tonight?” Athena returns a model-comparison widget with ensemble spread and divergence flags in about 90 seconds.
- “Backtest a wind-ramp strategy on EPT-2e over the last two winters.” Athena returns a full backtest report with RMSE and CRPS breakdowns by lead time in about 5 minutes.
- “Build me a workspace showing German solar generation overlaid with the model delta on surface solar radiation across EPT-2e and ECMWF ENS.” Athena assembles and persists the dashboard, with no analyst or BI team required.
Customer teams use Athena in different ways based on their structure and workflow maturity. Regulated utilities such as Axpo, EnBW, and EDF replace the manual 7–9 a.m. morning briefing routine. Day-Ahead and Intraday briefings auto-refresh on every new model run and cover model consensus across 25+ models, model delta since the previous run, convergence tracking, market spread, and price implications.
Physical trading houses such as TotalEnergies and Statkraft focus on speed around divergence and correction events. Athena surfaces divergence alerts the moment two models disagree and correction alerts the moment a model revises its own output, so trade windows open with a notification instead of a missed move.
Quant funds treat Athena as an analyst layer on top of programmatic access. Athena turns raw physics predictions from EPT-2 into actionable analysis by reading market context and modelling participant behaviour. Quant developers pipe forecasts directly into systematic models via pip install jua, with the REST API exposing 25+ models through a single schema and Apache Arrow support for large payloads. Hindcast data is available across multiple Jua and third-party models for backtesting, and documentation lives at docs.jua.ai.
Trading houses and quant desks describe Athena as “another headcount, for free.” Internal meteorologists shift from manual briefing production to deeper forecast research. Watch Athena resolve a query on your region in 90 seconds.
Decision Table: How Different Customers Use Jua for Energy
Jua for Energy runs alongside ECMWF rather than replacing it. ECMWF AIFS, ECMWF's own AI model, runs natively on the Jua platform. Jua for Energy instead displaces the plumbing around the incumbent feed, including the in-house grib pipeline, manual benchmarking, the morning-briefing analyst, and dashboard stitching.
The product surface maps to each customer archetype based on operational constraints and workflow maturity. Regulated utilities face the broadest stakeholder base, because meteorology, dispatch, and trading teams all need simultaneous access. They typically adopt the full workspace, including maps, briefings, power forecasts, and divergence and correction alerts.
Physical trading houses operate with smaller, more specialised teams, which pushes them toward API-first access to raw EPT-2 and EPT-2e outputs, ensemble data, and the benchmarking surface. Athena then acts as an analyst layer for embedded meteorologists and junior traders rather than as a replacement for manual workflows.
Quant funds have already automated their pipelines and therefore bypass the UI. They operate through the Python SDK and REST API, run hindcast backtests programmatically, and pipe Jua forecasts into their own systematic models. For this archetype, the benchmarking surface and Athena backtest capability serve as proof points and deal triggers rather than daily tools.
Across all three archetypes, the live benchmarking surface, with 25+ models on any region and variable and results in seconds, is the proof-of-value moment. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 million per year. A 1 GW solar portfolio at the same accuracy gain saves approximately €3 million per year.
Frequently Asked Questions
What defines the most accurate forecast system in 2026?
Accuracy in 2026 is measured against ground-truth observations, not against other models, using standardised metrics including RMSE and CRPS. These metrics are evaluated across multiple variables and lead times without post-processing or station fine-tuning. EPT-2 is benchmarked against more than 10,000 real ground stations via Jua's open-source StationBench methodology. A forecast system that cannot publish its evaluation methodology and data sources transparently cannot be considered best-in-class. Physics-constrained architectures that respect conservation laws such as mass, momentum, and energy produce outputs that are physically consistent by construction, which supports trading-grade reliability.
How does physics-constrained forecasting compare with traditional NWP on wind and solar variables?
Traditional NWP decomposes the atmosphere into three-dimensional grid cells and solves differential equations inside each one. It remains accurate and has been the industry standard for forty years, but its compute cost caps update frequency at two to four runs per day. EPT-2 learns the governing physics of the atmosphere directly from observational data in a latent representation that is integrated forward in time faster than the physics itself unfolds. The result is a model that outperforms ECMWF HRES on 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation at every lead time from 0 to 240 hours, while running on a single GPU at a fraction of traditional NWP cost. Physics-aware architectures also substantially reduce violations of radiative constraints compared with purely data-driven approaches, which is critical for solar irradiance forecasting.
Which energy forecasting tools support 24×/day updates for renewables?
EPT-2 RR on the Jua platform maintains this hourly update cadence in production. Traditional NWP incumbents remain structurally capped at two to four runs per day by HPC compute economics. AI weather peers including Microsoft Aurora and Google DeepMind GraphCast are typically updated four times per day in research mode without a productised operational refresh schedule. Actual-generation power forecasts on the Jua platform refresh every 15 minutes across Germany, Great Britain, France, the Netherlands, and Belgium, covering solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load.
How do agent-based workflows integrate with existing quant pipelines?
Athena integrates with existing quant pipelines through two surfaces. The natural-language layer accepts objectives in plain text and returns briefings, benchmarks, backtests, and custom widgets in approximately 90 seconds to 5 minutes, without requiring API calls or pipeline changes. The programmatic layer exposes 25+ models through a REST API with Apache Arrow support for large payloads. The Python SDK installs via pip install jua from PyPI and provides forecast access, hindcast and backtesting, and weather-parameter standardisation across all models under a single schema. ENTSO-E grid data integrates directly for European power-market data. Quant teams that previously spent a quarter building ingestion pipelines for raw AI-weather research outputs stand up the Jua integration in days.
Conclusion: Move Before the Next NWP Run Lands
These July 2026 benchmarks resolve the EPT-2 vs ECMWF HRES question on every energy-critical variable and every lead time. The update-frequency question is resolved by EPT-2 RR's 24×/day cadence. The workflow question is resolved by Athena's 90-second query latency and the Jua platform's 25-model benchmarking surface. Jua for Energy gives traders, meteorologists, and quant developers a foundation-model and agent platform that lets them act before the market does, not after the next NWP run lands.
See the benchmark results for yourself as EPT-2 runs against your current provider on your own region and variable.