Forecast workbench · Snapshot 06 May 2026
UK Renewable Delivery Outlook
Explore 6,189 active projects, compare aggregate capacity outlooks, rank delivery evidence, map the pipeline and stress-test external conditions. The public two-year signal is a validated relative ranking—not a literal project probability.
Loading 6,189 forecast records…
Aggregate forecast
Pipeline Outlook
Filters update every metric and chart. Forecast capacity is aggregated across projects; exact project probabilities remain outside the public decision surface.
Expected Capacity by Technology
3-year aggregate outlook · highest-capacity technologies
Expected Capacity by Stage
3-year aggregate outlook · current public planning stage
Research queue
Project Delivery Rankings
The score compares projects with one another. A score of 80 means stronger two-year model evidence than roughly 80% of the current forecast universe.
Loading projects…
| Project | Technology | Region | Stage | Capacity | 2-year signal | Coverage | Evidence |
|---|
Page 1
External conditions
Delivery Scenario Lab
Apply bounded economic, grid and policy shocks to the filtered aggregate outlook. These sensitivities are intentionally separate from the trained project ranking.
Change from the reference financing environment
Real cost shock relative to the reference case
Portfolio-wide connection delay stress
Revenue-support shock relative to the reference case
Bounded qualitative support assumption
3-year scenario outlook
Reference capacity—
Stress-adjusted capacity—
—
100 conditions indexReference assumptions
Scenario coefficients are transparent, bounded sensitivities for comparison—not trained causal effects or investment-grade forecasts. They do not alter the validated project ranking.
Spatial intelligence
Forecast Project Map
Marker size represents capacity and marker tone represents the public two-year delivery-signal band. Filters remain active across the map.
Preparing located projects…
LowerTypicalStronger
Validated challenger
Forecast Model Evidence
The richer model is used only where it proved useful: resolving ties inside the stable empirical ordering.
Primary orderingEmpirical survival score
+AI tie-break evidenceStage age · milestones · capacity revisions · developer history · regional track record · CfD · portfolio pressure
→Public outputRelative percentile only
The standalone CatBoost challenger was rejected as a replacement because its overall ranking was weaker. Its within-bucket ordering improved both AUC and average precision, passed the pre-declared ranking gate and is the only AI component promoted to Forecast v2.2.
Product governance
Horizon Release Matrix
Each horizon is released independently. Ranking quality cannot justify publishing an uncalibrated percentage.
| Horizon | Test sample | Ranking AUC | Observed | Raw mean | Probability error | Public output |
|---|---|---|---|---|---|---|
| 2 years8 rolling cohorts | 4,01963 observed events | 0.791 | 1.6% | 8.9% | +33.9% worse | Ranking signal |
| 3 years5 rolling cohorts | 4,03577 observed events | 0.693 | 1.9% | 10.6% | +40.2% worse | Research only |
| 5 years2 rolling cohorts | 1,22289 observed events | 0.519 | 7.3% | 9.8% | +3.5% worse | Withheld |
The two-year score is exposed only as a percentile rank across the current forecast universe. Three-year values remain methodology research. Five-year values are removed from the public decision surface.
Held-out evidence
Ranking & Reliability Evidence
Higher ROC-AUC means better ordering. The rate comparison explains why the same scores are not released as probabilities.
Ranking Discrimination
ROC-AUC by horizon · dark marker at the 0.5 random baseline
Predicted and Observed Delivery Rates
Raw mean score versus realised outcome rate on held-out rows
2 years4,019 held-out rows
3 years4,035 held-out rows
5 years1,222 held-out rows
Decision workflow
Delivery-Signal Usage
The signal narrows the research queue; the underlying evidence remains the decision input.
Rank
Sort comparable projects by the two-year delivery signal. A score of 80 means stronger model evidence than roughly 80% of the current forecast universe.
Verify
Open the project record and inspect its planning stage, dated milestones, authority, CfD evidence and recorded constraints.
Stress
Use the Scenario Lab to test how rates, construction costs, grid delay and policy assumptions change portfolio delivery pressure.
Leakage controls
Temporal Validation Controls
Fully Observed Outcomes
An origin is evaluated only when the entire forecast horizon has elapsed. Unresolved recent projects are censored rather than labelled as failures.
Project-Purged Temporal Splits
Test projects are removed from survival-training rows. Features such as developer track record count only information dated before the forecast origin.
Calibration Governance
Log-odds, Platt and isotonic calibration are fitted on earlier out-of-time cohorts and tested on the latest complete cohort. They are not promoted using in-sample fit.
Release Abstention
The public product now abstains: a failed horizon is shown as research-only or withheld instead of publishing the least-bad candidate as a probability.
Feature roadmap
External Driver Register
Project evidence enters the trained challenger now. Market, political and news variables enter only after they have dated histories, source QA and a measurable out-of-time lift.
| Variable | Source | Current treatment | Status |
|---|---|---|---|
| Prior portfolio size and delivery track recordDeveloper | Derived from dated DESNZ REPD snapshots ↗ | Trained feature | Live in AI challenger |
| Connection agreement, queue position and target connection dateGrid | National Energy System Operator connections registers ↗ | Project-match feature after entity-resolution QA | |
| Planning stage, stage age and milestone historyProject delivery | DESNZ Renewable Energy Planning Database ↗ | Trained feature | Live in AI challenger |
| Capacity revisions, data completeness and observation countProject delivery | Derived from dated DESNZ REPD snapshots ↗ | Trained feature | Live in AI challenger |
| CfD award, strike price, delivery year and allocation roundSupport | DESNZ allocation results and Low Carbon Contracts Company register ↗ | Project-match feature and scenario input | Award flag live; contract detail next |
| Infrastructure construction pricesCosts | Office for National Statistics Construction Output Price Indices ↗ | Manual stress input | Live context |
| Copper, aluminium, iron ore and energy pricesCosts | World Bank Commodity Markets Pink Sheet ↗ | Technology-weighted scenario input | |
| Bank Rate, gilt yields and credit conditionsFinance | Bank of England statistical database ↗ | Live context and manual stress input | Bank Rate live; yield curve next |
| Dated policy events, planning reform and technology-specific supportPolicy and news | GOV.UK Content API and UK Parliament APIs ↗ | Source-linked event features; no free-form LLM probability adjustment | Research pipeline |
| Wholesale electricity price level and volatilityRevenue | Elexon Insights market-index price API ↗ | Manual stress input until dated history is archived | Scenario-ready |
| Accounts recency, insolvency flags and filing statusDeveloper finance | Companies House API ↗ | Entity-linked feature for UK companies | Research pipeline |
| Curtailment, negative-price hours and regional constraint exposureSystem value | Elexon Insights and NESO data ↗ | Regional scenario and portfolio-risk feature | Research pipeline |
| Developer share return, volatility and drawdownMarkets | Licensed market-data provider ↗ | Publicly listed developers only; never imputed to private firms | Research only |
Every external feature must be dated, source-linked and available at the forecast origin. Revisions and publication lags are retained so future information cannot leak into historical tests.
Politics, prices and external conditions
Macroeconomic & Policy Scenarios
Bank Rate, wholesale electricity prices, construction and material costs, grid delay, developer-financing stress, policy support and CfD assumptions can materially change delivery conditions. However, 15 independent REPD dates are not enough to estimate credible macroeconomic coefficients. The Scenario Lab therefore exposes them as bounded sensitivities, separate from the trained project ranking. This avoids pretending that thousands of projects sharing one date are thousands of independent observations of politics or inflation.