Prediction Market Data Intelligence was built to answer a single question: what do prediction markets actually know, and when do they know it?
Prediction markets are exchange-traded contracts that resolve to $1 if a specified event occurs and $0 if it does not. A contract trading at $0.67 represents the market's aggregate probability estimate — 67% — that the event will happen. Unlike opinion polls or pundit forecasts, prediction market prices are backed by real capital. Participants bear financial consequences for inaccuracy, which produces a powerful incentive for honest probability assessment.
The modern prediction market landscape is split across two primary venues: Polymarket, a decentralized, blockchain-settled CLOB exchange covering thousands of global events, and Kalshi, a CFTC-regulated US exchange offering regulated contracts on macro, political, and financial outcomes. Together they represent billions of dollars in aggregate open interest and the most liquid publicly observable probability data in existence.
Unlike traditional financial instruments, prediction market contracts terminate on discrete events with binary outcomes. This structure creates clean, interpretable probability time series that do not require the return-normalization steps needed for equity or commodity price analysis. The probability itself is the signal.
Prediction markets are among the most accurate probability estimators available for near-term discrete events. A substantial body of academic literature, originating with the Iowa Electronic Markets and continued through the Superforecasting research program, documents consistent outperformance of expert panels, polls, and institutional forecasts. The accuracy advantage is particularly pronounced in the 1–30 day window before an event, where market participants have strong incentives to incorporate all available information quickly.
For institutional participants, the value of this data is not the raw probability — it is the velocity and structure of probability changes. A prediction market probability that moves from 0.40 to 0.55 in 90 minutes on above-average volume carries qualitatively different information than the same move over three days on thin volume. The first pattern suggests that informed participants with new information are adjusting positions; the second is noise. Identifying the former in real time is what PMDI is built to do.
Cross-asset correlation analysis adds a second dimension of value. Prediction market probabilities on macro outcomes — Federal Reserve rate decisions, CPI print outcomes, election results — routinely lead corresponding adjustments in Treasury yields, equity volatility indices, and cryptocurrency prices by material time margins. This lead/lag structure, when documented with statistical rigor, constitutes actionable intelligence for participants in those correlated markets. PMDI computes and exposes these relationships continuously across 7-day, 30-day, and 90-day windows.
Polymarket operates a central limit order book (CLOB) settled on the Polygon blockchain. PMDI connects to Polymarket's real-time WebSocket API to capture every price tick, order book update, and volume event for all active markets. Market IDs are on-chain contract addresses, ensuring immutable resolution records. Coverage includes 1,000+ active markets across crypto, politics, economics, sports, and science.
Kalshi is a CFTC-designated contract market (DCM), the first federally regulated prediction market exchange in the United States. PMDI uses a single multiplexed WebSocket connection to subscribe to all active ticker channels simultaneously, capturing orderbook snapshots and delta updates. Coverage includes 500+ active contracts across macro economics, crypto prices, political events, and weather.
Cross-asset correlation analysis requires concurrent asset price data. PMDI ingests real-time cryptocurrency price data from Binance and Coinbase WebSocket streams, covering BTC, ETH, and major altcoin pairs. Traditional asset prices (equity indices, interest rate instruments) are integrated via secondary data feeds. All asset prices are stored in the asset_prices hypertable alongside prediction market data.
Every price tick is validated against three constraints before storage: (1) yes_price and no_price must each fall in [0, 1]; (2) their sum must not exceed 1.05 (a 5% float tolerance for market maker spread); (3) the tick must arrive within 30 seconds of its exchange-side timestamp. Ticks failing validation are logged to data_quality_log and discarded — never inserted with corrupted values.
Velocity is computed as the change in yes_price divided by elapsed seconds over four rolling windows: 60s, 5m, 15m, and 1h. A velocity score normalizes this change against the market's historical standard deviation of velocity, producing a Z-score-like measure that is comparable across markets with different base volatility levels. Signals are only generated from ticks where volume_usd > 0 — zero-volume ticks are treated as artifacts and excluded from signal computation.
A velocity score exceeding the 3-sigma threshold on any window sets the anomaly_flag on the probability_velocity record. Anomaly thresholds are computed per-market using a rolling 30-day window to adapt to market-specific activity levels. The system also flags volume surges defined as single-tick volume delta exceeding 2 standard deviations above the market's 24-hour average.
Pearson correlation coefficients are computed between probability time series and asset price series on 7-day, 30-day, and 90-day windows. Lead/lag analysis tests correlation at time offsets from -6 hours to +6 hours in 5-minute increments to identify whether the prediction market leads or lags the asset price. Results are filtered to p-value < 0.05 before storage. Correlations are recomputed every 4 hours.
High-confidence signals trigger a three-agent AI pipeline running in production autonomously. Claude (Anthropic) performs deep signal analysis: it receives the market question, detected signal, recent price history, and news context, then returns a narrative annotation, signal tags, and a confidence score (0–1). Gemini (Google) generates plain-language signal explanations and cross-platform divergence analysis, surfacing when Metaculus forecasters and Polymarket traders disagree by more than 15 percentage points. Perplexity retrieves live news context that both Claude and Gemini use to ground their analysis in real-world events. All AI calls are logged with token counts, latency metrics, and model version for cost monitoring and reproducibility.
Until September 26, 2026 this page reported a 71.0% “gated hit rate,” reproduced across three backtest runs. An internal audit found that figure was an artifact of how it was computed, and we have removed it.
The backtest assigned each signal a pattern_type
after its outcome was known: signals that turned out correct were labelled
sustained_divergence, while a subset of the incorrect ones were
labelled mean_reversion and then excluded as “noise.”
Filtering on that label meant filtering on the answer. It is also why
mean_reversion appeared to score 1.2% — not a weak pattern,
but a category that could only contain errors by definition.
Because the gate only ever removed errors, the resulting number reflects how often prediction markets snap back — a structural property of the markets — rather than whether our signal predicted anything. A signal with zero predictive power scores about 68% under that method. That is also why three runs agreed so closely: the metric barely responds to the underlying data, so reproducing it confirmed nothing.
The defect was in the gating, not in the underlying measurement. Raw directional accuracy — did the probability move in the direction the market ultimately resolved — was computed correctly and is unaffected. That is what we report below.
| Period | Source | Raw Directional | Scored Events | Verifiable Today |
|---|---|---|---|---|
| Dec 2022 → Apr 2026 | May 2026 backtest | 58.0% | 36,670 | No — source rows retired |
| Jun 1 – Jul 25, 2026 | Live capture | 54.7% | 2,810 | Yes — reproducible |
Two independent periods, different market mixes, landing in the same ~55–58% band. That consistency is the honest strength of the result. It is a real but modest edge, which is what a genuine edge in prediction markets looks like — a 71% directional hit rate would have been extraordinary.
These are volumes processed, not performance claims. Per-domain lead-time figures previously shown here were drawn from the gated event set and have been withdrawn pending recomputation on an ungated basis. We would rather publish less than publish a number we cannot stand behind.
PMDI is a data product. Quality is the product. The following operational targets are enforced by automated monitoring with alerting on violation:
Request a 14-day trial. No credit card required. Full API access within one business day of approval.