JevGym
Evaluate Jev forecasts against timestamped historical prediction markets and resolved outcomes.
Added to Jevfast
How it uses Jev
JevGym renders each historical market snapshot as a timestamped state plus a typed Choice question, then sends it through a provider adapter and reads the returned YES/NO probability distribution. The default forecast asks which way the contract resolves. Its arena can ask a second Choice question about buying YES, buying NO, or waiting; application code turns the answer into a simulated trade and scores it against the recorded resolution.
What Jev decides
Forecast a historical FOMC market (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · q
YOUR APP
INSTRUCTION
Which outcome will this contract ultimately resolve to?
STATE
Repository-owned synthetic offline fixture: KXFED-26JAN-CUT, the January 2026 Fed decision, with YES defined as a cut in the federal funds target range. The rendered state names a historical checkpoint and its available evidence. Market-aware mode also reveals the market probability. This is an illustrative fixture, not a captured live forecast.
JEV · CHOICE
- YES
- NO
Choose an arena action (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · q
YOUR APP
INSTRUCTION
Choose whether to buy one YES contract, buy one NO contract, or wait for a later timestep.
STATE
The same fixture snapshot includes market state; the rendered action state adds the current YES and NO purchase prices. The question offers the repository's three fixed arena actions.
JEV · CHOICE
- bet_yes
- bet_no
- wait
App workflow
- Prepare historical checkpoints
Ingest and parse market data, build timestamped snapshots, and validate temporal leakage and dataset splits. A checked-in synthetic fixture supports a reproducible offline pipeline.
- Render a Jev question
Render each eligible snapshot into historical state, the frozen forecast question, and YES/NO candidates. The blind track omits market probability; the market-aware track reveals it.
- Collect forecasts
For each selected provider, track, and valid resolved snapshot, send the rendered question and collect a normalized probability distribution.
- Simulate arena decisions
Run the selected arena mode over timestamp-ordered snapshots. The action mode sends a three-option Choice request; edge-based modes compare Jev's probability with market prices. Score resulting positions against the eventual outcome and assemble the leaderboard.
Why it is interesting
One leakage-controlled snapshot powers both blind and market-aware tracks, with the renderer deciding whether to reveal the contemporaneous market price. The same question contract works across providers, while the arena connects calibrated probabilities to sequential trading decisions and realized P&L.