Real-time state of the four Polymarket signals

Each card displays the current standardised z-score of one Polymarket signal, measured against its rolling 20-hour distribution. The z-score isolates how anomalous the most recent move is relative to its own recent activity, regardless of the absolute magnitude of the signal.

How to read each card

• Z-score — number of standard deviations away from the rolling mean. Anything above ±2 is flagged as a SHOCK, the threshold our shock-panel framework uses to identify information arrival.
• Δ — raw signal change in the last hour.
• Threshold — current value of μ + 2σ. Any |Δ| above this line triggers a shock.
• Status badges: SHOCK means information arrival is currently in progress; CALM means quiet trading; ILLIQUID means the contract has no recent trades and the signal is unreliable.

How to use it

When a card flips to SHOCK, the live execution panel above expands with the asset–horizon combinations that historically respond to that signal, with recommended trade direction and holding period. Click any signal card to drill down into its history.

Shock detection formula: |ΔPM_t| > μ_t + 2σ_t where μ and σ are rolling statistics over the previous 20 liquid hours. Missing hours are NaN-masked before computing the rolling window to avoid false positives during illiquid periods.

Backtest performance for every PM-asset pair

Each row represents a strategy that opens a position whenever the relevant Polymarket signal enters a shock state, takes the direction implied by the sign of the shock, holds for h hours (the optimal lag from the shock-panel IRF), and closes the position. No overlapping trades are permitted.

Key metrics

• Gross SR — annualised Sharpe before transaction costs. Captures the raw informational edge of the signal.
• Net SR — Sharpe after applying realistic round-trip costs (2 bps SPY/QQQ/Oil, 1 bps BTC/Gold/SP500fut). Captures what is actually exploitable.
• Max DD — largest peak-to-trough drawdown on cumulative trade returns.
• Tier badge: PRIME (Net SR > 0.5) — standalone alpha; WATCH (Gross > 1.0 but Net negative) — overlay candidate; AVOID — not trade-worthy.

How to use it

Read this table as a capital allocation guide. PRIME pairs go directly into a systematic portfolio; WATCH pairs become signal amplifiers on top of existing factor strategies (momentum, carry, volatility); AVOID pairs are filtered out.

Click any row to see the full trade statistics, holding period, and the IRF coefficient that drives the holding decision.

Pattern of positive Gross / negative Net is the empirical signature of a Grossman-Stiglitz equilibrium: information exists in the signal, but the margin available to traders is consumed by realistic frictions.

Heatmap of impulse response coefficients

Each cell is the coefficient β from a bivariate OLS regression r(t+h) = α + β · ΔPM(t) + ε estimated only on shock hours, with HC3 heteroskedasticity-robust standard errors. The p-value below each β indicates statistical significance.

Cell colour code

• Solid amber (filled) — optimal transmission lag for that pair, as selected by the IRF. This is the holding period the backtest uses.
• Bordered amber (medium) — significant at p < 0.05.
• Light amber (faint) — significant at p < 0.10.
• Empty — no significant transmission detected.

How to use it

Use this as an execution timing guide. When a shock fires, the heatmap tells you which assets should react and how many hours to wait before unwinding the trade. For example, FED→SPY transmits in 1 hour while UNEMP→SPY takes 6 hours — execution timing matters.

Click any highlighted cell for a full breakdown including sample size, correlation, and directional t-test.

Following Jordà (2005), these horizon-specific coefficients are estimated via direct projection rather than inverting a VAR — more robust to lag misspecification in sparse panels.

Gradient-boosted directional forecasts

LightGBM is trained on the shock panel using PM features (pm_delta, pm_z, pm_vol), calendar features, announcement proximity, and lagged macro conditioning variables (DXY_chg_L1, US2Y_chg_L1) and target asset rolling volatility. Validation uses 5-fold walk-forward expanding window.

Metrics on the left

• DA LGBM — out-of-sample directional accuracy of the gradient-boosted model. 50% is random; above 55% is meaningful.
• DA OLS — directional accuracy of a bivariate OLS using only pm_delta, evaluated on the same walk-forward folds. The benchmark.
• p-binom — one-sided binomial test against a 50% random walk. p < 0.10 means LGBM is statistically better than chance.
• p-McN — McNemar test on discordant pairs. p < 0.10 means LGBM is statistically better than the linear OLS baseline.
• Verdict: STRONG (both tests significant); EDGE (DA > 50% with one test marginal); FLAT (no edge); BAD (DA < 50%).

Feature importance on the right

Switch between PM-asset pairs to see which features drive the model. The systematic dominance of DXY and asset volatility above the raw PM signals confirms that PM-to-asset pass-through is conditional on the prevailing macro regime, not unconditional.

SPY is excluded from LightGBM because its shock panel contains fewer than 50 training observations in early CV folds — insufficient for reliable model fitting. SPY remains in the OLS shock-panel analysis where small samples are tractable.

Are PM participants forward-looking?

For each of the four scheduled macro releases, this panel measures the average absolute PM signal change in the 6 hours before the announcement, divided by the average change during all other hours. A ratio > 1 means traders are systematically more active before the release than at other times.

How to read

• Activity ratio — large number means strong pre-event positioning. Stars indicate statistical significance against the null of random activity (Welch t-test).
• Best asset — for each event, the asset whose post-announcement return is most strongly predicted by the cumulative pre-announcement PM signal. β is the OLS slope coefficient with HC3 standard errors.

How to use it

This validates the premise that Polymarket genuinely anticipates news rather than just reacting to it. High activity ratios for CPI and Employment confirm that pre-release PM positioning is informative — useful as a signal in event-driven strategies. The most robust predictive link is INF→Oil with p < 0.001.

Sample sizes are small (9–13 events per type) due to the 13-month window and quarterly cadence of GDP. Treat the asset-level regressions as directionally informative rather than as definitive evidence.

Information transmission speed across the full sample

Unlike the shock-panel IRF (which conditions on shock hours), the DLM regresses next-hour returns on PM signals at lags 1 to 6 simultaneously, on the entire 9,673-hour sample. The peak lag identifies the horizon at which the PM signal carries the most remaining predictive content for next-hour returns.

How to read each row

• Peak h — horizon at which the PM signal has its strongest significant β coefficient.
• Bar visualisation — taller bar at the peak lag, lower bars elsewhere. Quick visual cue for absorption speed.
• β at peak — coefficient magnitude at the peak lag.
• p at peak — HAC-robust p-value at the peak lag.

How to use it

Cross-validates the shock-panel IRF results on a different sample. When DLM and IRF agree on a lag (e.g., FED→SPY at h=1–2), confidence is high. When they disagree, treat the IRF as primary because it isolates information arrival hours.

Specification: r(t+1) = α + φ·r(t) + γ'X(t) + Σ β·PM(t-k) + ε with k = 1…6. Newey-West HAC standard errors with 5 lags. Missingness dummies excluded for interpretability.

Do PM signals add anything beyond standard financial controls?

For each asset, two OLS specifications are estimated:

• Baseline: regresses next-hour return on standard financial controls (VIX, DXY, US2Y, US10Y), hour/day-of-week dummies, announcement dummies, and own lagged return.
• Augmented: adds the four PM signals plus their missingness dummies.

The incremental R² (ΔR²) measures how much extra explanatory power the PM block contributes. The joint F-test asks whether the four PM coefficients are collectively non-zero.

How to interpret

Small but consistently positive ΔR² across all assets confirms that PM signals carry incremental information. The fact that the joint F-test is significant only for Oil (p < 0.05) confirms that the unconditional signal is weak — most of the predictive content is concentrated in the rare shock hours analysed in §02–§04, not in the average hour.

This is the empirical motivation for the entire shock-panel framework: averaging across all 9,673 hours dilutes the genuine signal because most hours contain no information arrival.