JM.Back to work ↗
01 / Forecasting · energy markets · WebMCP

Can you check a forecast without rewriting its history?

I built a dashboard that records forecasts before the outcome is known, keeps missed days visible, and lets an AI assistant explain the same evidence you can inspect.

6page-native WebMCP tools
25 daysscored in the submitted snapshot
8 dayswithout a forecast in that snapshot
31 Octformal evaluation gate
Decision

Use the workspace to inspect the record. Whether the forecasting model improves on the baseline remains a separate, unfinished evaluation.

Forecast Ledger artwork over wind turbines and electricity transmission lines
Model Context ProtocolOpenAI WebMCP Challenge
Hackathon entry
Forecast Ledger · submitted 3 September 2026 DE-LU day-ahead electricity
Why it mattered

A missed forecast is part of the record

A historical result can look better if missing forecasts are quietly filled in later. This record saves forecasts before the auction closes, scores them after settlement and leaves failed runs visible.

What I built

One workspace for the analyst and the agent.

I connected six WebMCP tools to the dashboard’s existing filters and record views. When an assistant inspects a date or a missed run, the page shows that same evidence. The tools cannot rewrite forecasts.

AuditCheck the complete prospective record
CompareMeasure against two persistence baselines
InspectOpen misses, weak days and cutoff rules
Technical evidence

The submitted challenge snapshot contains 25 scored days and eight missed days. Every miss remains visible and no forecast was backfilled. Running price skill was +37.1% against the same-hour-last-week baseline, but that number is descriptive only: the preregistered assessment remains pending until 31 October 2026.

Open submitted workspace ↗Watch two-minute demo ↗Inspect challenge repository ↗View Devpost entry ↗
What the result means

Auditability is not forecast accuracy

The workspace makes the record easier to inspect. A positive running score does not settle whether the forecasting model reliably wins; the planned assessment remains pending.

What comes next

Finish the evaluation without changing the past

The next research milestone is the preregistered assessment after 31 October 2026. The challenge submission stays frozen for judging, with continuing research kept separate.

Boundary

This is an audit of forecasting research, not a trading system. It makes no trading-edge or P&L claim. The submitted application and challenge repository are frozen for judging; continuing research is kept separate.