Political Prediction Markets Dataset

Data Glossary Updated Daily

Methodology

How we measure prediction accuracy and why the choice of when to sample prices is a first-order decision.

Brier Score Calculation

The Brier score measures forecast accuracy: (prediction − outcome)². A score of 0 is perfect; 0.25 is equivalent to random guessing. We calculate Brier scores at multiple time horizons (60, 30, 14, 7, 3, 1 days before the event) and, for electoral markets, using the election eve price.

Election Eve Price (Electoral Markets)

For electoral markets, we use the market price at 00:00 UTC on election day as our primary forecast evaluation price. We call this the election eve price. For U.S. elections on November 5, this corresponds to approximately 7–8 PM Eastern on November 4—before any polls have closed.

This is an event-anchored truncation rule, not a resolution-anchored one. The distinction matters because Polymarket and Kalshi settle contracts at different speeds after outcomes are known (24–48 hours for Polymarket, 3–12 hours for Kalshi). Any truncation rule defined relative to resolution time mixes genuine forecasts with post-outcome convergence and introduces a platform-specific bias into cross-platform comparisons.

The election eve price avoids this. The same election produces the same measurement timestamp on both platforms. The price reflects what the market believed before the outcome was known—which is what forecast accuracy should measure. For full discussion, see our methodological note.

Reference Dates by Market Type

Different market types use different reference dates, reflecting the information available to researchers and the settlement characteristics of each platform:

Market Type	Reference Date	Rationale
Electoral Markets	00:00 UTC on election day	Event-anchored. Uses verified election dates from our lookup table. Price is queried at exact UTC midnight via the Dome API. This is pre-outcome for all U.S. elections and unambiguous across platforms.
Non-Electoral (Polymarket)	trading_close_time − 2 days	Resolution-anchored with offset. Polymarket markets typically close 24–48 hours after event outcomes are known. The 2-day offset is a conservative pre-event estimate.
Non-Electoral (Kalshi)	trading_close_time − 1 day	Resolution-anchored with offset. Kalshi markets settle within 3–12 hours. The 1-day offset provides a conservative pre-event window.

Note: For non-electoral markets, we do not have event dates and must rely on resolution-relative offsets. These offsets are less precise than event-anchored dates and may include some post-outcome trading, particularly on Polymarket. We document this limitation transparently. Where possible, we use the electoral approach as the benchmark.

Platform Settlement Windows

Our methodology accounts for documented platform settlement procedures:

• Polymarket: Per their documentation, "Markets are settled within 24 to 48 hours after the event outcome becomes definitively known." [docs]
• Kalshi: Markets settle via automated processes typically within 3–12 hours of outcome determination. [settlement sources]

Price Data Sources

Election eve prices are queried from the Dome API at exact Unix timestamps (UTC midnight on election day). This provides point-in-time prices rather than daily candle approximations. For time-horizon Brier scores (1d, 7d, etc.), we use daily price observations at 00:00 UTC intervals.

Citation

If you use this dataset in your research, please cite:

BibTeX:

@misc{hall2026bellwether,
  author       = {Hall, Andrew B. and Paschal, Elliot J.},
  title        = {Bellwether Political Prediction Markets Dataset},
  year         = {2026},
  publisher    = {Stanford Graduate School of Business},
  howpublished = {\url{https://elliotjames-paschal.github.io/Bellwether/data.html}}
}