Written by: Olivier Lam, Physical AI Team, Jua.ai AG
Key Takeaways for Quant Energy Teams
- European atmospheric data covering wind speed, temperature, solar irradiance, precipitation, and pressure drives signals for TTF gas, EUA carbon, spark spreads, and intraday power markets.
- Forecast accuracy relies on RMSE for deterministic skill and CRPS for probabilistic skill, both critical for day-ahead positioning and tail-risk strategies.
- Legacy NWP models like ECMWF deliver high accuracy at high compute cost and limited update frequency, while many AI models cut cost yet often lack production-grade hindcasts and APIs.
- Jua’s EPT-2 and EPT-2e models outperform ECMWF HRES and the 50-member ENS on RMSE and CRPS across key variables and lead times, with updates up to 24 times per day and native hindcast archives for backtesting.
- Run your first head-to-head comparison against your current provider and complete your first backtest in under five minutes.
How the European Atmospheric Data Landscape Breaks Down
The European atmospheric data landscape splits into two generations. Legacy numerical weather prediction (NWP) includes ECMWF HRES, ECMWF ENS, NOAA GFS, and DWD ICON, which solve differential equations across three-dimensional grid cells. A single NWP simulation consumes approximately 8,400 kWh and costs €1,000–€20,000 on HPC infrastructure, which constrains the industry to two to four global forecast updates per day. ECMWF Open Data retains only a rolling archive of the most recent forecast runs, so multi-year hindcast backtesting is unavailable on the free tier.
The second generation is AI-based and includes Microsoft Aurora, Google DeepMind GraphCast, and ECMWF’s AIFS. These models reduce inference cost by several orders of magnitude but usually arrive as research outputs, meaning raw model files without ensembles, documented hindcasts, or productised refresh schedules. Quant teams that subscribe to these outputs must build ingestion pipelines, ensemble logic, and benchmarking harnesses themselves.
Jua operates as 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, such as mass, momentum, and energy conservation, directly from observational data. Athena is Jua’s AI agent that plans, calls tools, and turns natural-language objectives into concrete deliverables. Jua for Energy is the first applied product built on EPT and Athena, and it is the surface through which quant funds access European atmospheric data today. The relationship mirrors Anthropic and Claude Code, a horizontal AI platform with a flagship vertical product.
Core Concepts for Trading With Atmospheric Data
Two accuracy metrics govern atmospheric data evaluation for trading. RMSE (root mean square error) measures deterministic forecast skill, so lower RMSE against ground-truth observations means a more accurate point forecast. CRPS (Continuous Ranked Probability Score) measures probabilistic skill, or how well the full forecast distribution matches the observed outcome. A model can have low RMSE but poor CRPS if its ensemble spread is miscalibrated. Both metrics matter: RMSE drives day-ahead positioning, while CRPS drives options and spread strategies where the tails of the distribution carry the P&L.
An ensemble is a set of perturbed model runs that samples forecast uncertainty. The ECMWF ENS runs 50 members, and EPT-2e, Jua’s ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time. A hindcast is an archived set of historical forecast vintages, meaning forecasts issued at a specific time in the past for verification against what actually happened. Hindcasts form the raw material of backtesting, because without them a quant team cannot evaluate whether a signal would have been tradeable at the time of issuance.
EPT-2 introduces native any-Δt forecasting. The model is trained to predict at arbitrary time steps rather than rolling forward in fixed 6-hour increments. Aurora and most AI peers use a fixed 6-hour roll-forward, which compounds error at each step. EPT-2 does not roll and avoids that compounding. Athena, instrumented with the Jua for Energy tool surface, turns a natural-language query into a benchmark, backtest, or custom widget in approximately 90 seconds, which removes the pipeline-assembly step that usually consumes quant engineering capacity.
Strategic Trade-offs When Selecting Atmospheric Data Sources
With these technical capabilities established, the next decision concerns which data source to select. Three trade-offs define the source-selection decision for quant teams building European atmospheric signals.
Accuracy versus latency. The ECMWF two-week outlook is the definitive reference for repricing risk in European energy markets, yet a typical Jua run completes approximately 2.5 hours ahead of competing operational runs at the same cycle. For intraday strategies, EPT-2 RR updates up to 24 times per day compared with the industry standard of two to four. The team that sees the next forecast revision first has the trade window, while the team that sees it second reacts to a price that has already moved.
Ensemble depth versus compute cost. Deeper ensembles improve CRPS and tail-risk estimation but increase data volume and pipeline complexity. EPT-2e provides 30-member probabilistic output updated 4 times per day and beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time, at a fraction of the infrastructure cost. A single EPT-2 inference runs at approximately 0.25 kWh and $0.20–$15 on a single GPU, compared with approximately 8,400 kWh and €1,000–€20,000 for a traditional NWP run.
Raw research outputs versus productised APIs. ERA5 reanalysis provides consistent hourly data from 1990 onward at 0.25° resolution and suits climatology and regime normalisation, but it cannot emulate tradable intraday forecast timing because reanalysis is not a forecast vintage. AI research subscriptions such as Aurora and GraphCast deliver raw outputs without hindcast archives or schema stability guarantees. The engineering cost of building a production pipeline on top of raw research outputs is measured in quarters, not days.
Implementation and Operational Best Practices for Quant Teams
A production-grade European atmospheric data stack for quant trading has four components. First, a live forecast source with documented accuracy provides the real-time signal. Second, a hindcast archive enables walk-forward backtesting against past market conditions. Third, a benchmarking harness maintains ongoing model surveillance so teams can detect accuracy drift before it hits P&L. Fourth, a Python SDK with schema stability and large-payload support ties these components together into a single integration point.
Jua for Energy addresses all four through one integration surface. The Python SDK installs via pip install jua and exposes more than 25 models, including 10 proprietary AI models from the EPT family plus 15 third-party NWP and AI models such as ECMWF HRES, ENS, AIFS, NOAA GFS, DWD ICON, Aurora, and GraphCast, through a unified schema. The REST API (POST /v1/forecast/data and related endpoints, documented at query.jua.ai/docs) supports Apache Arrow payloads for large continental, multi-variable, multi-model queries. Hindcast data is available across multiple Jua and third-party models for backtesting. ECMWF’s own infrastructure adheres to FAIR principles, meaning Findable, Accessible, Interoperable, and Reusable, and Jua’s unified schema applies the same principle across all models on the platform.
To illustrate how these four components work together in practice, consider a standard backtesting workflow on Jua for Energy. The team installs the SDK, queries hindcast data for the target variable and region, runs a walk-forward evaluation against ERA5 or station observations, and compares RMSE and CRPS across EPT-2, EPT-2e, ECMWF HRES, and any third-party model in the same call. Athena executes the same workflow from a natural-language prompt in approximately five minutes. EPT-2 and EPT-2e benchmarks are validated against more than 10,000 real ground stations on open-source StationBench, with no post-processing or station fine-tuning, and documented in peer-reviewed technical reports on arXiv:2507.09703 and arXiv:2410.15076.
Test the integration yourself and pipe Jua forecasts into your own models to run your first backtest in under five minutes.
Readiness and Opportunity Assessment for Funds
The checklist below identifies when a quant fund is ready to integrate European atmospheric data as a systematic signal.
Technical fit. The team has a Python-based data pipeline and can consume REST or SDK outputs. Apache Arrow or other columnar payload formats are supported downstream. A walk-forward backtesting framework exists or can be built quickly. The team can evaluate RMSE and CRPS against a reference dataset independently.
Operational fit. At least one energy-sensitive contract, such as TTF, EUA, spark spread, or intraday power, sits in the book. The strategy horizon maps to a forecast lead time where atmospheric skill is demonstrable, such as 0–240 hours for EPT-2 or ensemble horizons out to 60 days for EPT-2e. The team can act on intraday signal updates rather than only daily closes.
Strategic fit. The compliance carbon market is projected to reach approximately $1.21 trillion in 2026, with Europe as the largest regional market accounting for roughly one-third of the global share. CBAM certificates were priced at €75.36 per ton of CO2 in Q1 2026, which makes temperature-driven demand signals directly material to EUA positioning. A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves approximately €1.5 million per year, and a 1 GW solar portfolio at the same accuracy gain saves approximately €3 million per year. Funds operating multi-GW exposures scale these economics linearly.
Common Pitfalls When Operationalising Atmospheric Data
Relying on vendor graphics instead of self-run benchmarks. Accuracy claims from atmospheric data vendors are only as credible as the methodology behind them. EPT-2 benchmark numbers are reproducible on open-source StationBench against more than 10,000 real ground stations. Any vendor that cannot provide the same reproducibility should be evaluated with caution.
Building brittle GRIB pipelines. ECMWF’s earthkit-data library converts GRIB files into NumPy arrays, pandas DataFrames, or xarray datasets, yet the pipeline still requires format-specific handling, version management, and schema translation across models. A unified API with Apache Arrow support removes this layer entirely. Integration that takes a quarter to build on raw GRIB outputs stands up in days on the Jua for Energy SDK.
Lacking documented hindcasts. As noted earlier, ECMWF’s free tier lacks historical archives, and full access requires separate service agreements. ERA5 reanalysis provides data from 1990 onward but cannot substitute for time-stamped forecast vintages when backtesting intraday strategies. A strategy backtested on reanalysis alone will overestimate signal quality because reanalysis does not replicate the information set available at trade time.
Treating AI research outputs as production data. Aurora, GraphCast, and ECMWF AIFS are research-grade outputs without productised refresh schedules, ensemble variants, or hindcast archives. Subscribing to them as raw feeds and building the production layer internally is a significant engineering commitment with no accuracy guarantee. EPT-2 outperforms Aurora on 10 m wind, 100 m wind, and 2 m temperature across the full 0–240 hour range, and Aurora has no surface solar radiation output.
Frequently Asked Questions
How does EPT-2 differ from ECMWF HRES for energy trading purposes?
ECMWF HRES is the 40-year deterministic benchmark for global NWP, running at 9 km resolution with two to four updates per day. EPT-2 outperforms HRES across the full 0–240 hour range on the four variables most material to energy P&L: 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation. EPT-2 also delivers forecasts earlier, reflecting the timing advantage noted in the trade-offs section, and EPT-2 RR updates up to 24 times per day. Jua for Energy does not replace ECMWF, because serious customers keep their ECMWF subscription and run Jua for Energy alongside it. The product instead displaces the plumbing around the ECMWF feed, including the GRIB pipeline, manual benchmarking, and morning-briefing assembly, with both models available on the same platform under a unified schema.
What hindcast depth is available for backtesting TTF gas and spark-spread strategies?
Hindcast data is available across multiple Jua and third-party models on the Jua for Energy platform. ERA5 reanalysis, available from 1990 onward at 0.25° resolution, provides the long-horizon climatological baseline. For walk-forward backtests that require time-stamped forecast vintages, which is the correct methodology for intraday and day-ahead strategies, Jua’s hindcast archive covers the variables and lead times relevant to TTF, EUA, and spark-spread positioning. Athena executes a full backtest from a natural-language prompt in approximately five minutes, or the same workflow runs programmatically through the Python SDK.
How does EPT-2e compare to ECMWF ENS for probabilistic energy strategies?
EPT-2e is Jua’s ensemble variant, updated 4 times per day with a 60-day horizon. As established earlier, EPT-2e outperforms the ENS mean on both metrics across the forecast horizon. For strategies that depend on the tails of the forecast distribution, such as options on spark spreads, EUA volatility positioning, and wind-ramp derivatives, CRPS skill at extended lead times is the relevant metric. EPT-2e delivers that skill with a 30-member ensemble at inference costs several orders of magnitude below traditional NWP ensemble runs.
What is the integration path for a quant developer with an existing Python pipeline?
The integration starts with pip install jua. The REST API exposes more than 25 models through a single schema with Apache Arrow support for large continental, multi-variable queries. Hindcast and backtesting access is available through the same SDK. The developer dashboard sits at developer.jua.ai and full API documentation at docs.jua.ai. Integration that takes a quarter to build on raw GRIB or research-model outputs stands up in days. ENTSO-E grid data is available natively for European power-market cross-referencing.
How are EPT-2 accuracy claims validated independently?
EPT-2 benchmarks are evaluated against more than 10,000 real ground stations on open-source StationBench, with no post-processing or station fine-tuning applied. Results are published in peer-reviewed technical reports on arXiv, documented in the reports cited earlier. The live benchmarking surface on the Jua platform allows any evaluator to run a head-to-head comparison on their own region and variable against ECMWF HRES, ENS, Aurora, GraphCast, or any of the models on the platform, with results in under 30 seconds. The numbers are reproducible by the customer, not just by the vendor.
Conclusion: Turning Atmospheric Skill Into Trading Edge
European atmospheric data acts as a primary driver of TTF gas demand, EUA carbon pricing, spark-spread dynamics, and intraday power balancing rather than a commodity input. The quality of that data, measured by RMSE and CRPS against real observations, the depth of hindcast archives available for backtesting, the frequency of model updates, and the engineering cost of integration, determines whether a systematic strategy delivers edge or noise.
Legacy NWP delivers accuracy at the cost of update frequency and pipeline complexity. AI research outputs reduce inference cost but shift the engineering burden onto the quant team. Jua for Energy delivers hourly global forecasts at up to 5 km resolution natively, outperforming leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE, through a documented Python SDK and Athena agent that turns backtests and benchmarks into five-minute operations rather than quarter-long engineering projects.
The performance advantages outlined in the key takeaways, including EPT-2’s superiority over HRES and EPT-2e’s edge over the ENS mean, are documented in the peer-reviewed reports referenced above and are reproducible on open-source StationBench. Customers including Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec execute daily trading decisions on the platform across four continents.
See the benchmark on your data and run EPT-2 head-to-head against your current European atmospheric data provider on your own region and variables.