Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 1, 2026
Key Takeaways for Energy Trading Teams
- Multi-model ensemble forecasts combine outputs from several independent atmospheric models, reduce bias, and quantify uncertainty. Lower uncertainty directly cuts imbalance costs in power markets.
- EPT-2e outperforms the 50-member ECMWF ENS on both RMSE and CRPS across almost the entire 0–240 hour range while using only 30 members.
- Jua’s rapid-refresh models update up to 24 times per day at a tiny share of traditional NWP cost (~0.25 kWh vs. 8,400 kWh), which enables higher-frequency ensemble updates for traders.
- Improved forecast accuracy on hub-height wind and surface solar radiation can save €1.5 M–€3 M per year for a 1 GW renewables portfolio under standard hedging structures.
- Run EPT-2e head-to-head against your current ensemble provider on your own region and variables.
Single-Model vs. Multi-Model vs. EPT-2e Performance
The latest technical report, arXiv:2507.09703, documents EPT-2e, the ensemble variant of Jua’s Earth Physics Transformer. The report shows a consistent performance edge over ECMWF ENS on RMSE (root mean square error, the average deviation from observed values) and CRPS (continuous ranked probability score, a measure of probabilistic sharpness and calibration) across the 0–240 hour forecast range. A companion report, arXiv:2410.15076, established that EPT-1.5 outperformed GraphCast, FuXi, Pangu-Weather, and ECMWF HRES on European wind and temperature. EPT-2 extends those gains across every lead time and on every variable that drives an energy P&L.
Energy teams choosing an ensemble provider decide between the long-standing NWP incumbent, newer AI peers that lack production-ready ensembles, and EPT-2e, which delivers probabilistic skill with lower operational cost. The comparison below quantifies that choice across the dimensions that matter for trading P&L, from accuracy and horizon to cost and workflow integration. All numeric claims are anchored to arXiv:2507.09703 and the Jua platform specification.
| Capability | EPT-2e (Jua for Energy) | ECMWF ENS (NWP incumbent) | Aurora / GraphCast (AI peers) |
|---|---|---|---|
| Ensemble members | 30 members | 50 members | No productised ensemble |
| RMSE vs. ECMWF ENS mean (0–240 h) | Beats ENS mean at virtually every lead time | Benchmark reference | Not applicable, no ensemble output |
| CRPS vs. ECMWF ENS mean (0–240 h) | Beats ENS mean at virtually every lead time | Benchmark reference | Not applicable, no ensemble output |
| Update frequency | 4 runs per day (EPT-2e), up to 24 runs per day via EPT-2 RR | 2–4 runs per day | Typically 4 runs per day, no productised operational schedule |
| Forecast horizon | Ensemble out to 60 days | 15 days (ENS) | Typically 10 days, research mode |
| Spatial resolution (native forecast) | Up to 5 km (EPT-2 HRRR, Europe) | 9 km (HRES), ENS coarser | ~25 km at published resolution |
| 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 |
| Productised agent layer | Athena: briefings, benchmarks, backtests, widgets (~90 s) | None | None |
Schedule a benchmark session to run EPT-2e against your current ensemble on your own region and variables.
Forecasting Hub-Height Wind and Solar for Power Markets
Hub-height wind speed, the wind velocity at turbine rotor height between roughly 80 and 150 m, drives wind-generation forecasts more than any other variable. Surface solar radiation (SSRD) plays the same role for solar-generation forecasts. Single-model deterministic forecasts struggle with both variables at the 12–72 hour lead times that matter most for day-ahead and intraday power trading.
EPT-2 outperforms ECMWF HRES on 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation across the full 0–240 hour range, as documented in arXiv:2507.09703. Microsoft Aurora exposes no SSRD output, so EPT-2 remains the only AI model with full coverage of the four variables that drive a renewables P&L. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves about €1.5 M per year under typical hedging and penalty structures. A 1 GW solar portfolio that gains four percentage points of forecast accuracy saves approximately €3 M per year under similar conditions.
How Multi-Model Ensembles Reduce Uncertainty in Trading
RMSE measures the average magnitude of forecast error against observed values, so lower scores indicate better point forecasts. Energy traders, however, need more than a single best-guess value and must understand the full range of plausible outcomes. CRPS addresses that need by measuring the full probabilistic skill of an ensemble forecast and rewarding sharp, well-calibrated distributions while penalizing diffuse or biased ones.
An ensemble creates that distribution by running multiple simulations with perturbed initial conditions or different physics, and the spread across those runs quantifies forecast uncertainty. The value of any ensemble also depends on timing, which includes lead time, the interval between forecast initialization and the valid time, and dissemination, the delay between computation finishing and the forecast reaching the user. For backtesting trading strategies against historical conditions, teams rely on hindcasts, which are forecasts run over past periods using the same configuration as the operational system.
Traditional numerical weather prediction (NWP) decomposes the atmosphere into grid cells and solves differential equations inside each cell, which locks performance to supercomputer economics. A single NWP simulation consumes approximately 8,400 kWh and costs €1,000–€20,000. The European supercomputer runs its full algorithm twice a day, and with supplementary runs the energy industry receives roughly four global forecasts per 24 hours. Between runs, traders work with stale numbers.
EPT-2 RR, Jua’s rapid-refresh model, updates up to 24 times per day. EPT-2e runs 4 times per day on the ensemble configuration. A single EPT-2 inference runs on a single GPU in minutes at approximately 0.25 kWh and $0.20–$15, which is roughly four orders of magnitude cheaper than an equivalent NWP run. This cost asymmetry makes high-frequency ensemble updates economically viable for the first time.
EPT-2e Ensemble Performance Benchmarks
EPT-2e is evaluated against more than 10,000 real ground stations using Jua’s open-source StationBench methodology, with no post-processing or station fine-tuning. This standard matches the evaluation used for ECMWF HRES, the leading NWP model for forty years. The results in arXiv:2507.09703 show EPT-2e delivering a consistent probabilistic edge over the 50-member ECMWF ENS mean on both RMSE and CRPS across the tested lead times.
A typical Jua run completes about 2.5 hours ahead of competing operational runs at the same cycle, which gives traders earlier access to the revised forecast signal. EPT-2 was trained on 8 × H100 GPUs over 10 days, while Microsoft Aurora required 32 × A100 GPUs over 18 days. Inference runs about 25 percent faster than Aurora.
This performance advantage stems from EPT’s architecture as a general physics foundation model that learns the governing dynamics of continuous, conservation-law-constrained physical systems. The atmosphere is the first system EPT has been fine-tuned for, and energy trading is the first market where these capabilities have been productised through Athena and Jua for Energy.
Request a live benchmark to see EPT-2e results on your highest-stakes region and variable in under five minutes.
Athena-Driven Natural-Language Workflows
Athena is Jua’s AI agent, currently instrumented with the Jua for Energy tool surface. A trader types a natural-language objective such as “show the 100 m wind forecast spread across models for northern Germany tonight” or “backtest a wind-ramp strategy on EPT-2e over the last two winters.” Athena then plans, calls tools, evaluates intermediate outputs, and returns a briefing, a benchmark, a backtest, or a custom widget. Athena turns raw physics predictions from EPT-2 into actionable trading context by reading market conditions and modeling the forecast landscape. Typical queries resolve in about 90 seconds, and backtests in about 5 minutes.
Athena keeps ECMWF ENS in the workspace alongside EPT-2e when that comparison adds value. The Jua platform hosts 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, Microsoft Aurora, and GFS GraphCast, all under a unified schema. A single Athena query can compare EPT-2e probabilistic spread against ECMWF ENS for any region, variable, and time window, and then return the result as a persistent workspace widget.
When to Retain ECMWF ENS Alongside EPT-2e
Jua for Energy complements rather than replaces ECMWF. Regulated utilities, physical trading houses, and quant funds typically keep their ECMWF subscription and run Jua for Energy alongside it. ECMWF AIFS, ECMWF’s own AI model, also 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 harness, morning-briefing analyst work, and spreadsheet stitching that fills the 7–9 a.m. window.
The practical guidance is straightforward. Use EPT-2e as the primary ensemble for RMSE- and CRPS-optimized probabilistic forecasts, especially for hub-height wind and solar radiation at lead times of 12–240 hours. That approach does not require abandoning ECMWF ENS, which remains useful as a reference signal and for regulatory or internal-risk frameworks that still mandate the incumbent benchmark. The real edge comes from running both, because the Jua platform’s divergence alerts fire the moment EPT-2e and ECMWF ENS disagree on a key variable, and that disagreement becomes a tradeable signal that flags heightened forecast uncertainty.
Frequently Asked Questions
What are multi-model ensemble forecasts, and how do consensus methods work?
Multi-model ensemble forecasts combine outputs from several independent atmospheric models, each built on different physics approximations, initialization schemes, or parameterizations, into a single probabilistic prediction. Consensus methods include the simple ensemble mean, which averages all member outputs, weighted superensembles, which assign weights based on each model’s historical skill, and rank-histogram calibration, which adjusts ensemble spread to match observed climatological variability. Each method trades computational cost against probabilistic sharpness.
The ensemble mean reduces random errors across members but can smooth out extreme events. Weighted methods preserve more of the signal from higher-skill members. Calibration methods correct systematic biases in ensemble spread. EPT-2e applies these principles within a foundation-model architecture constrained by conservation laws such as mass, momentum, and energy, which produces a probabilistic output that remains physically consistent by construction rather than through post-processing.
How often does EPT-2e update, and how does that compare to traditional NWP ensembles?
EPT-2e runs 4 times per day on the ensemble configuration, and EPT-2 RR, Jua’s rapid-refresh deterministic model, updates up to 24 times per day. Traditional NWP ensembles, including ECMWF ENS, update 2–4 times per day and remain constrained by the compute economics of running 50 members on HPC infrastructure. As described earlier, those simulations sit around four orders of magnitude more expensive per run than EPT-2. A typical Jua run also completes about 2.5 hours ahead of competing operational runs at the same cycle, which gives Jua for Energy customers earlier access to the revised forecast signal before the market re-prices.
How do I integrate multi-model ensemble outputs into my existing trading pipeline?
Jua for Energy exposes all models, including EPT-2e and third-party NWP and AI models, through a REST API and a Python SDK installable via pip install jua. The API uses Apache Arrow for large-payload forecast queries and applies a unified schema across all models, so switching or comparing models does not require re-engineering pipelines. Hindcast data is available across multiple Jua and third-party models for backtesting systematic strategies, and ENTSO-E grid data integrates directly for European power-market context. Quant developers and engineering teams typically complete the integration in days rather than the quarter it often takes to build equivalent infrastructure from raw research-model outputs. Athena can also run backtests programmatically in about 5 minutes or generate custom workspace widgets on a natural-language request in about 90 seconds.
Why does EPT-2e outperform ECMWF ENS with fewer ensemble members?
ECMWF ENS uses 50 members initialized with singular-vector perturbations to sample forecast uncertainty. EPT-2e achieves superior RMSE and CRPS at virtually every lead time with 30 members, as documented in arXiv:2507.09703. The efficiency gain comes from the underlying architecture, because EPT is a spatiotemporal transformer foundation model trained on more than 5 petabytes of observational data from over 120 sources, including more than 10,000 proprietary ground stations.
The model learns the governing physics, including conservation of mass, momentum, and energy, directly from observational data in a latent representation instead of discretizing differential equations over a grid. This approach yields an ensemble spread that matches real atmospheric uncertainty more closely and therefore requires fewer members to reach equivalent or superior probabilistic skill. Evaluation uses Jua’s open-source StationBench methodology against more than 10,000 real ground stations, with no post-processing or station fine-tuning.
Conclusion: Trading Impact of EPT-2e Ensembles
Multi-model ensemble forecasts define the production standard for probabilistic energy trading, and the benchmark has shifted. The benchmark results detailed earlier in this article show EPT-2e’s consistent edge over the incumbent ensemble, delivered with 30 members at inference costs around four orders of magnitude below traditional NWP and with a dissemination advantage of about 2.5 hours over competing operational runs at the same cycle. The financial impact detailed earlier, roughly €1.5–€3 M per year for a 1 GW renewables portfolio, scales linearly across the multi-GW portfolios operated by customers including Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec.
Jua is a foundation model and agent company, and Jua for Energy is the first applied product. The architecture learns physics, and the domain becomes a variable. The trading results follow from that design.
Schedule a demo to run live EPT-2e benchmarks against your current ensemble provider on your own region, variable, and time window, with results in under five minutes.