A prediction market prices one thing extremely well: the probability of a discrete future event, updated continuously, with money at stake. Polymarket and Kalshi alone combined for roughly $44.5B in volume in 2025, which means real capital is now repricing tariff decisions, Fed meetings, and deal closes in real time.
The problem for an equity desk is vocabulary. A market resolves on "Will the Fed cut rates in March?"; a portfolio holds banks, REITs, and rate-sensitive names. The probabilities are public. The bridge between an event and the securities it moves is not, and building that bridge is the subject of Gerra's research on mapping prediction markets to securities and the resulting Event Probability Map.
This framework also runs outside our own pages: Integrity Research published our case study on the approach, Prediction Markets as an Event-Driven Signal, written with the alternative-data advisory Eagle Alpha.
From event to exposed securities
The framework maps 300+ active events to 500+ securities. Each event links to between two and eight tickers, and every link carries two attributes: an exposure type (positive if the security rises as the event probability rises, negative if it falls) and a sensitivity score from 0 to 1, derived from historical co-movement between probability changes and security price changes.
A live example from the published sample file, a geopolitical market on whether Iran closes the Strait of Hormuz, observed at 68% probability after an 11-point rise over seven days on $8.4M cumulative volume:
| Ticker | Name | Exposure | Sensitivity |
|---|---|---|---|
| CL=F | Crude Oil WTI Futures | Positive | 0.93 |
| XLE | Energy Select Sector SPDR | Positive | 0.87 |
| XOM | Exxon Mobil Corp | Positive | 0.82 |
| CVX | Chevron Corp | Positive | 0.80 |
| GLD | SPDR Gold Shares | Positive | 0.78 |
| SPY | SPDR S&P 500 ETF | Negative | 0.65 |
One event, four venues, one identifier
The four covered platforms identify markets in incompatible ways: Polymarket uses an on-chain condition ID and token ID (documented in its developer API), Kalshi a CFTC-registered ticker, Limitless a Base-chain contract, and Metaculus a numeric question ID. The feed resolves all four to one canonical event, so the same question traded on multiple venues collapses to a single series with multiple probability sources.
That unification produces a useful derived field: cross_platform_spread, the probability divergence when one event trades on more than one venue. In the sample, a Fed no-change market runs at 97% with a 0.01 spread between Polymarket and Kalshi. When that spread widens, one venue is mispricing the other, which is information in itself.
Why point-in-time history is the load-bearing property
Every probability observation is stamped at its trade-execution timestamp and never restated. History runs at 5-minute resolution back to June 2022, and resolved events stay in the panel, so there is no survivorship filter quietly removing embarrassing markets.
This is the property to interrogate hardest in any event-probability feed, because a restated series manufactures predictive power. If historical probabilities have been smoothed, gap-filled, or rebuilt from a later snapshot, a backtest sees odds no trader could have seen, and the resulting strategy is fiction. The test is simple: any analysis run against the series must see exactly what was knowable at each timestamp, the same standard that applies to sentiment data or any other point-in-time panel.
Three concrete workflows
Event-risk inventory. Join portfolio holdings against the mapped_securities arrays of all active events and aggregate sensitivity per event. The output is a ranked report of which live events the book is actually exposed to, in which direction. Around a policy decision, that turns "the Fed meets Wednesday" into a signed exposure statement: in the sample's Fed no-change market at 97%, the mapped exposures are TLT negative at 0.88 sensitivity, XLF positive at 0.72, the dollar index positive at 0.70, and gold negative at 0.62, with the residual 3% defining the tail scenario.
Probability-move alerts. The panel carries 24-hour and 7-day probability changes per event. The research backtested moves greater than 10% across 500+ resolved events spanning the 2024 US election cycle, 2025 tariff announcements, and Fed rate decisions, and found that probability changes tend to lead the mapped security moves by hours to days. Alerting on large moves in events mapped to held names creates a window to hedge or pre-position while the equity side is still absorbing the news.
Earnings-adjacent event separation. A name reporting this week may also face a regulatory ruling or a deal-close market resolving in the same window. Checking active mapped events before attributing a move to earnings separates the event component from the fundamental one, which matters for both pre-positioning and post-mortems.
What the evidence supports, and what it does not
The lead-lag finding is reported qualitatively, and deliberately so: the research declines to publish a single headline hit-rate because the relationship varies by event category, by liquidity, and by how cleanly an event maps to its securities. That restraint is consistent with two decades of literature. Wolfers and Zitzewitz (2004) established that prediction markets aggregate dispersed information into efficient forecasts of uncertain events, and also that market design and depth determine how far those forecasts can be trusted.
The honest limits, stated in the research and worth restating here:
- Liquidity varies enormously. Thin markets produce noisy probabilities. The feed carries
volume_total, 24-hour volume, and open interest per event precisely so users can filter to markets deep enough to mean something. - The mapping is a model. Which securities an event affects, and how strongly, is an estimate validated against history. A misclassified or incomplete link degrades the signal no matter how clean the probability series is.
- Resolution risk is real. Markets resolve on precise written criteria by a stated resolution date, and the resolved outcome can diverge from the economic outcome a portfolio actually cares about. The probability is an input to judgment, not a verdict.
The delivery shape
The Event Probability Map is a derived analytical product: no raw prediction-market data is redistributed, sourcing is from public APIs with no PII or MNPI, and securities are mapped to CUSIP and ISIN for direct joins against a book. Full history is roughly 15 GB, delivered over S3, REST, or WebSocket, real-time at 5-minute resolution with sub-second processing lag or as a daily batch.
Prediction markets now price a meaningful share of the event risk that moves portfolios. The dataset exists so that this pricing arrives structured: signed, scored, cross-venue, and safe to backtest.