News
DIANA: The Full Trading Cycle — From Atmosphere to Portfolio
TradeWpower AS presents
DIANA
Digital Intelligent Artificial Numerical Advisor
The Full Trading Cycle — From Atmosphere to Portfolio
Weather → Physical Prices → Financial Risks → Portfolio Strategy → Repeat
|
v5.7 Engine |
94 Research Sections |
130+ Climate Variables |
|
35 Years Calibration |
282 Automated Tests |
23 Regime Rules |
The Problem Is Clear
ECMWF is excellent at Day 6 — 81% directional accuracy on wind. By Day 14, it drops to 60%. By Week 3–4, it is barely above a coin flip at 55%.
Weeks 2–4 are where positions are built, hedges are placed, and risk is managed. The market is uncertain. ECMWF skill collapses. That is exactly where DIANA operates — holding 67–68% accuracy, where the best NWP model in the world drops to 55%.
That 12–15-percentage-point gap — sustained over thousands of weeks — is the edge. In trading, a consistent edge compounds.
The best part: DIANA is strongest when it matters most — the extreme weeks above and below normal that drive 50–80% of annual P&L. Where ECMWF drops to coin-flip territory, DIANA delivers 87–97% accuracy on the very weeks that make or break portfolios.
The Core Edge: DIANA vs ECMWF
Verified against real EC forecasts — 312 weeks, 2020–2026. 50% = coin flip.
| EC Day 6 | EC Day 14 | EC W3–4 | DIANA | vs D14 | vs W3–4 | |
| Wind (avg) | 81.4% | 60.4% | 54.7% | 67.1% | +6.7pp | +12.3pp |
| Wind (winter) | — | 61.2% | 52.7% | 68.0% | +6.8pp | +15.3pp |
| Precip (avg) | 81.7% | 64.5% | 57.4% | 65.8% | +1.3pp | +8.4pp |
Week 1
Market already knows. Priced in. Limited edge.
Week 2–4 — THE CORE EDGE
Market uncertain. EC skill drops sharply. DIANA holds 67%+. Positions built, hedges placed, risk managed.
Beyond Week 4
Monthly/quarterly regime assessment via cross-season bridge signals. Far superior to EC seasonal forecasts (SEAS5).
Extreme Week Accuracy — Where the Money Is
Extreme weeks (~15% of all weeks) drive 50–80% of annual P&L. A single extreme wind week in Germany swings portfolio value by millions. Backtested 1991–2025.
| Market | Ext. Low | Below | Normal | Above | Ext. High |
| NP Wind | 87.2% | 76.0% | 60.2% | 73.0% | 79.4% |
| DE Wind | 86.8% | 82.7% | 57.6% | 63.3% | 82.4% |
| FR Wind | 96.7% | 84.2% | 60.9% | 55.8% | 74.8% |
| ES Wind | 82.4% | 76.9% | 57.8% | 49.1% | 66.9% |
96.7% — FR Wind extreme calm. Dunkelflaute detection. 29 out of 30 extreme calm weeks caught across 35 years.
87–97% — Accuracy range across all four markets at the extremes — the 15% of weeks that drive 50–80% of annual P&L.
Out-of-Sample Extreme Accuracy (walk-forward, never trained on test data):
| Market | Ext. Low | Below | Normal | Above | Ext. High |
| NP Wind | 77.8% | 75.5% | 61.6% | 74.0% | 80.8% |
| DE Wind | 89.2% | 80.3% | 55.4% | 66.2% | 87.0% |
| FR Wind | 94.1% | 80.9% | 61.8% | 55.3% | 76.1% |
| ES Wind | 88.1% | 79.1% | 57.9% | 45.3% | 63.9% |
DE Wind extreme below improves OOS: 89.2% vs 86.8% in-sample. The signal is structural.
Market-by-Market Accuracy
All-season and winter — 1991–2025 (2,392 weeks). 50% = no skill.
| Variable | All-Season | Winter | Plain Terms |
| NP Wind | 70.5% | 74.9% | 7 out of 10 correct |
| DE Wind | 70.0% | 75.0% | 7 out of 10 correct |
| FR Wind | 69.7% | 74.1% | 3 of 4 in winter |
| ES Wind | 62.8% | 66.3% | Hardest market |
| NP Hydro | 72.9% | — | Best signal |
| ES Hydro | 65.6% | — | Rain/drought |
| Solar (DE/FR/NP/ES) | 60–65% | — | Cloud proxy |
| Wind Average | 68.3% | 72.6% | 4-market avg |
Walk-Forward Validated — Not Overfit
4-fold cross-validation (2006–2025). Each fold was trained on 15 years, tested on 5 years, never seen during training.
| Metric | In-Sample | Out-of-Sample | Drop |
| Wind avg | 68.9% | 67.9% | −1.0pp |
| Winter Wind | 72.3% | 71.4% | −1.0pp |
| NP Hydro | 72.0% | 71.3% | −0.7pp |
| OOS/IS Retention | — | 98.5% | Minimal |
Per-Fold Stability (wind avg, all markets):
| 2006–10 | 2011–15 | 2016–20 | 2021–25 |
| 68.5% | 67.5% | 68.7% | 66.2% |
Stable across all folds. The 2021–25 fold (most recent climate) still delivers 66.2%.
What DIANA Is
DIANA closes the full trading cycle — from atmospheric regime signals through energy weather impact (GWh) to price response and position signal. Physics-first, not a black box. Not a neural network. Not a chatbot. Every parameter traces to a physical mechanism.
The Trading Cycle: Weather → Physical Prices → Financial Risks → Portfolio Strategy → Repeat
1. Weather: Is a regime shift coming? Probability-based forecasts vs physical weather risks. Market consensus vs DIANA analysis. WRM and polar vortex/jet stream framework.
2. Physical Prices: What is already priced in? Stack model. Weekly to quarterly. Track spot vs forward curve disconnect.
3. Financial Risks: Where are the flows? Track large fund positions, rolling behaviour, financial position risks. Optimal strategy based on risk-reward.
4. Portfolio Strategy: Scale or exit. Optimal entrance volume and ramp-up. Early stop losses if analysis fails.
Independent risk assessment at each stage. Stop loss or ramp-up can be triggered at any of Steps 1, 2, or 3 — weather fails → stop; physical prices don’t follow → stop; financial positioning turns against → stop. Confirmation at each step → ramp up. Sometimes Step 3 runs prices, not Step 1 — fund positioning and short-covering can move markets independently of weather. The cycle repeats on confirmation.
WEATHER LAYER — THE PROVEN ENGINE (130+ variables, 2,392 weeks, ERA5 1991–2026):
Stratosphere — Early Warning
Polar cap height profiles, polar vortex position/shape/trajectory (displaced, stretched, split), SSW descent signatures. Signals arrive 2–6 weeks before surface weather.
Troposphere — Engine Room
North Atlantic and European pressure patterns. Jet stream position and strength. Blocking detection and classification. Mid-atmosphere circulation anomalies.
Surface — Ground Truth
Regional temperature anomalies across European energy zones. SST anomalies in key ocean regions (seasonal lead signals). Energy generation anomalies — wind, solar, hydro, temperature.
Teleconnections — Remote Controls
AO/NAO and related indices. MJO phase and amplitude. Pacific and tropical pathways. ENSO state and seasonal modulation.
Verified Signal Hierarchy:
Stratospheric signals give 2–6 weeks lead. Polar vortex configuration wins for directional prediction (2.6× vs surface alone). When stratosphere and surface agree: +27pp accuracy boost. When they disagree: surface wins this week, stratosphere wins next week.
Additionally: 23 regime-conditional rules · 47 unified regime atlas sections · 282 automated tests passing
High-Value Signals DIANA Detects
Polar Vortex Disruptions
SSW events are detected before reaching the surface. 2–4 weeks lead over surface-only analysis. Nordic wind swings 2,400+ GWh/week. German wind swings 6,400+ GWh/week.
Tropical Wave Teleconnections
MJO phase transitions trigger European blocking/zonal flow 2–4 weeks later. Single-phase transition swings German wind by 6,700+ GWh. Most trading desks do not track this systematically.
Trapped Low / Conflicting Signals
Stratosphere says “warm and calm,” but surface stays bearish — trapped configuration. 66 documented cases with cross-market cascade effects. 51–57pp accuracy spread between correctly and incorrectly classified.
Cross-Season Bridges (3–6 Month Lead)
Summer SST → autumn temperatures (4-month lead, r>0.59). Autumn Arctic blocking → winter stratospheric state (3-month, r>0.60). Winter polar vortex weakness → spring Nordic wind. Far superior to EC SEAS5.
Boundary Layer Analysis NEW v5.7
Detects lower atmosphere disagreeing with surface — cold advection under ridges, warm advection above cold surfaces, and inversions. Temperature spreads 0.58°C, hydro spreads 325 GWh between opposing configurations.
Cross-Market Contrarian Setups
NP and ES are structurally opposite — the same teleconnection regime drives strong Nordic winds while suppressing Iberian winds. Blocking benefits ES, punishes NP/DE/FR. DE and FR are co-dependent. High conviction on one market automatically informs all others.
Why DIANA Is Uncopyable — The Moat
1. Proprietary Research Outside AI Training Data
94 original research sections from ERA5 analysis. None in the training data of ChatGPT, Claude, Gemini. Cannot be reproduced by prompting. Cannot be bought. Trade secrets.
2. Physics-First, Not Black-Box
Strat-trop coupling, Rossby wave propagation, polar vortex dynamics, boundary layer structure, SST leads. Interpretable, debuggable, trustworthy. A black-box ML model cannot explain WHY. DIANA can.
3. 94 Documented Research Sections
Teleconnection-energy correlations, stratospheric propagation, polar vortex trajectories, blocking stability, cross-market cascades, trapped low dynamics, SST leads, cross-Atlantic coupling, boundary layer diagnostics, US weather risk, LNG modules. Every claim tagged: VERIFIED, PARTIALLY VERIFIED, HYPOTHESIS, or PROCESS RULE.
4. Continuous Learning and Adaptation
17 monotonic climate trend corrections. Decadal era-correction. Daily loop: ingest → classify → compare → learn → adapt. Engine adapts to the 2020s atmosphere, not the historical average.
5. Cross-Market Intelligence
NP↔ES structurally opposite. Blocking benefits ES, punishes NP/DE/FR. DE↔FR co-dependent. A high-conviction NP call automatically informs all four markets simultaneously.
6. Proven Out-of-Sample
98.5% OOS/IS retention. DE extreme below improves OOS (89.2% vs 86.8%). Stable across all 4 folds. The signal is structural, not overfit.
DIANA Learns and Improves Over Time
Not a static model. Not a locked product. New research goes from idea → tested against 35 years → verified → wired into engine → live in production. This cycle happens in hours, not months. From v2.4 to v5.7: 19 major versions, 94 research sections, 23 regime rules — built in months, not years.
Research Growth: v3.0: 42 → v5.0: 63 → v5.3: 74 → v5.5: 91 → v5.7: 94 (growing)
Rule Growth: v5.0: 8 → v5.3: 14 → v5.4: 18 → v5.6: 22 → v5.7: 23 (each verified)
Targetable High-Value Setups:
DIANA can be applied to specific atmospheric configurations: Dunkelflaute events, sudden wind ramps, cold spells during low-wind periods, hydro drought onsets, and NP-vs-ES spread trades. When a trading desk says, “We need better accuracy on THIS event,” DIANA can research, test, and deploy a targeted rule. Rapidly.
The Business Case
VS HIRING A SENIOR ENERGY METEOROLOGIST:
| Met Hire | DIANA | |
| Deploy | 6–12 mo hire + 6–12 mo onboard | Weeks |
| Cost/year | ~2.5M NOK all-in (one person) | Fraction of a single hire |
| Key-person risk | HIGH | ZERO |
| Coverage | 1 person, sick days, holidays | 130+ vars, 4 markets, 24/7 |
| Output | Qualitative, human variance | Quantified GWh + probability |
VS STANDARD WEATHER SERVICE:
Standard services give raw forecasts + narrative — same data as competitors. DIANA gives regime detection + GWh quantification + AI vs EC comparison. Proprietary intelligence, not commodity data. Week 3–4: EC ~55%, DIANA 67%+.
VS BUILDING IN-HOUSE:
2–5 years research from zero vs immediate deployment. Requires PhD met + quant + developer. May not work (expensive failure risk). DIANA: 94 sections done, proven 67.9% OOS, 98.5% retention. Continuously improved.
OPERATIONAL RISK REDUCTION:
Zero handoff degradation (met → analyst → trader chain eliminated). Zero key-person dependency. Zero subjectivity variance. Quantified confidence on every signal. Complete audit trail. 24/7/365 operation.
The Green Shift — Structural Tailwind
Every GW of new renewable capacity makes weather regimes matter more to price formation. European capacity grows 5–10% per year. By 2030, wind and solar dominate marginal price setting.
| Market | Wind 2030 | Solar 2030 |
| Germany | 100 GW | 215 GW |
| Spain | 62 GW | 77 GW |
| France | 45 GW | 54 GW |
| Nordic | 40+ GW | — |
More capacity = more GWh variance per regime shift = more P&L impact = more value from DIANA. The edge does not erode — it grows.
DIANA Price Stack — Closing the Trading Cycle
In Active Development
The Weather layer is proven and operational. The Price Stack completes the Trading Cycle — connecting atmospheric regime signals directly to forward curves and portfolio positioning.
Weather: What will the atmosphere do to energy generation?
Physical Prices: What is already priced in — and where is the market wrong?
Financial Risks: Which positions carry weather-driven risk the market has not accounted for?
Portfolio Strategy: Where should conviction be deployed?
Live interaction with forward power and gas prices. Decomposes fuel costs, renewable expectations, demand, interconnector flows, plant availability, and merit order. Automatically detects mispriced weeks and months by comparing DIANA’s weather confidence against what the forward curve implies.
THE FULL LOOP:
1. DIANA Weather detects regime state and forecasts energy weather per market
2. Price Stack reads the forward curve and decomposes what is priced in
3. Where DIANA’s confidence disagrees with market pricing — flag it
4. Trader reviews, applies market judgement, and executes
DIANA does not trade. Traders trade. DIANA adds the quantified confidence layer no human can maintain consistently. Together, the combination is more powerful than either alone.
Hydro Reservoir Intelligence — Beyond Normal-Period Assumptions
The industry standard for reservoir filling simulations, inflow forecasts, and water value calculations is to run scenarios based on a fixed normal period — typically 30 years of historical weather applied uniformly. This treats every future week as equally likely to resemble any historical year. It is wrong.
DIANA changes this fundamentally. Because DIANA knows the current atmospheric regime state — polar vortex configuration, teleconnection phase, blocking patterns, SST anomalies — it can select the subset of historical years that are genuinely representative of the current outlook. Instead of simulating reservoir filling with 30 equally-weighted years, DIANA constructs regime-conditional weather ensembles: validated weather means drawn from years that actually resemble the atmospheric setup ahead.
Reservoir Filling Simulations
Regime-weighted year selection produces filling trajectories that reflect the actual atmospheric outlook — not a climatological average. When DIANA detects a blocking-dominated regime, it selects dry historical analogs. When it detects a zonal flow setup, it selects the wet ones. The result: tighter, more realistic filling scenarios.
Inflow Forecasting
Inflow projections based on regime-conditional precipitation and temperature means — not flat climatology. Structurally better than normal-period assumptions because the selected weather scenarios reflect the actual atmospheric regime ahead.
Water Value Calculations
Water values depend on expected future inflow and price. Both are weather-driven. When the inflow scenarios used in the water value model are regime-conditioned rather than normal-period-based, the resulting water values better reflect the actual risk landscape — leading to better dispatch decisions and more accurate forward pricing.
The principle is simple: the atmosphere is not random. The current regime state constrains what comes next. DIANA uses that constraint to produce weather scenarios that are validated, regime-consistent, and structurally superior to using a normal period.
Who Benefits from DIANA
Hydro Producers / Utilities
Regime-conditioned reservoir filling simulations, inflow forecasts, and water value calculations — structurally better than normal-period assumptions. DIANA selects historically representative years for the current atmospheric outlook, producing tighter and more realistic scenarios for dispatch, hedging, and forward pricing. Critical for seasonal production planning and storage optimisation.
Fund Managers / Institutional Investors
Lower VaR through extreme-week detection (87–97%). Systematic edge: +12–15pp over EC at W3–4. One system covers NP, DE, FR, ES (+ TTF, HH adaptable). Defensible moat — 94 proprietary sections, 35yr calibration. Fraction of a met team cost. Roadmap: Price Stack connects weather intelligence directly to forward curves.
Energy Trading Companies
Replace 2–3 person met/analyst chain. Data-to-signal in minutes, not hours/days. No quality variance from who is on shift. Cross-market intelligence. Interactive — traders can ask DIANA to investigate specific setups. Targetable: point DIANA at highest-risk positions. Full audit trail.
Risk Managers
Quantified confidence on every call — probability, not narrative. Extreme-week early warning for VaR-blowing events. Cross-season risk signals with 3–6 month lead. Walk-forward validated. Climate-adapted: works in the 2020s atmosphere.
TradeWpower AS
Energy Weather and Trading Intelligence · Lillehammer, Norway
The full trading cycle — from atmospheric regime signals through energy weather impact to price response and position signal.
Physics-based. 35 years of calibration. 94 research sections. Zero key-person dependency. 24/7.
The system adapts. The edge compounds. That is the moat.
www.tradewpower.no · post@tradewpower.no · +47 928 46 276
