Snapshots
Polymarket identifiers, questions, rules, prices, and time-stamped snapshots flow into the database on every cycle.
Partisan Premium Index
An independent forecasting system for identifying where political prediction markets diverge from model fair value.
PPI runs on a twice-daily schedule: it captures market snapshots, collects structured evidence, builds independent probability estimates, and publishes the spread between model and market. The index is live.
The live index.
The spread.
PPI = Pmarket − Pmodel
The partisan premium is the market-implied probability minus PPI's independent fair value. Positive means the market prices the outcome higher than the model does.
Largest current dislocations.
The five widest gaps between market price and model fair value, from the live terminal. Across all twelve tracked markets the average absolute premium is 23.4 points, and the market sits above model fair value in ten of twelve.
| Market | Market | Model | Premium |
|---|---|---|---|
| Democrats win the NC Senate race · RSO-0042 | 91.5% | 42% | +49.5 |
| Democrats win the MI governor race · RSO-0052 | 90.5% | 42% | +48.5 |
| Democrats win the GA Senate race · RSO-0044 | 93.5% | 55% | +38.5 |
| Democrats win the ME Senate race · RSO-0026 | 70.5% | 35% | +35.5 |
| Republicans win the OH governor race · RSO-0061 | 45.5% | 68% | −22.5 |
Premium in percentage points. All five markets resolve at the 2026 U.S. midterms and are unresolved as of this snapshot.
One market, end to end.
Will the Democrats win the North Carolina Senate race in 2026? — the widest dislocation in the current index. Every PPI value has this much provenance behind it.
Market probability91.5%
Model fair value42%
Premium+49.5 pts
Evidence11 public evidence items collected this run — polling, race ratings, and primary reporting.
Model & run
deepseek-v4-flash · run ppi-daily:2026-08-26:primaryRationale“North Carolina is a competitive state with a slight Republican lean. The Democratic incumbent is retiring, and the GOP has a strong bench… the evidence is limited and mixed, so I set the probability below 50%.”
ReadingThe market is near-certain of a Democratic hold; the model sees a genuine toss-up and flags its own evidence as thin. The value is a disagreement with provenance, not a verdict.
Track record.
Resolved markets0
Live tracked markets12
EvaluationBegins as tracked markets resolve
Real results
No tracked markets have resolved yet. This section will populate automatically as tracked markets reach resolution — from the live terminal's own record ↗ opens in a new tab, not from anything on this page.
How resolved markets will be scored
Illustrative evaluation preview — synthetic examples, not PPI results.
| Market | Model | Market | Outcome | Model error | Market error |
|---|---|---|---|---|---|
| Example race A | 62% | 55% | 1 | 38 pp | 45 pp |
| Example race B | 31% | 44% | 0 | 31 pp | 44 pp |
| Example race C | 78% | 71% | 1 | 22 pp | 29 pp |
| Example race D | 40% | 58% | 1 | 60 pp | 42 pp |
Outcome is 1 if the event happened, 0 if it did not. These four rows live only in demo/evaluation-fixtures.json, flagged synthetic: true — they are not in any PPI dataset, database, or run.
What real resolutions will produce
- Brier score
(forecast probability − outcome)², averaged over resolved markets — lower is better. - Absolute forecast error
|forecast probability − outcome|per market, in percentage points. - Model vs. market
The model's error against the contemporaneous prediction-market probability captured at the same time — which was closer, and by how much.
- Calibration
Group forecasts into probability bands and compare the predicted frequency with the realized frequency, once the sample is large enough to mean anything.
Every market snapshot, evidence item, and fair-value estimate is stored prospectively rather than overwritten, so these run against a genuine forecast-of-record the moment a market resolves — no back-fitting.
How it works.
Provenance
Every forecast preserves the evidence and source-run provenance behind it — a record of what the model knew at prediction time.
Versioned
Every estimate is versioned alongside the evidence and inputs that produced it.
Twice daily
PPI captures market and evidence snapshots on a twice-daily schedule; the published index updates as new runs complete.
Market ingest→Evidence→Fair-value estimate→Spread→Publish
Method.
Each estimate is an independent probability built from LLM-assisted analysis of polling, fundamentals, and primary-source evidence, versioned with the inputs that produced it.
- Source repositoryGitHub ↗ opens in a new tab
- Market provider in the ingestion layerPolymarket ↗ opens in a new tab