DV-PMI Surveillance · Spreads
Adverse Selection / Spread Pattern
Cumulative snapshot through 2026-02-28
Effective spreads on Polymarket are measurably tighter (-0.173 units, t = -176) in markets where flagged wallets trade. Of 204,772 markets with sufficient data, 65,910 (32.2%) contain at least one wallet flagged by the orthogonality test. The textbook Glosten-Milgrom prediction is widening, not tightening; the realized direction inverts that prediction and is consistent with the Paper 1 finding that algorithmic market makers concentrate in information-rich markets rather than retreating from them.
| Mean spread | Markets | |
|---|---|---|
| Markets with flagged wallets | 0.471 | 65,910 |
| Markets without flagged wallets | 0.643 | 138,862 |
| Gap | -0.173 | (tighter) |
Effective spread is measured per trade as the gap between price paid and the immediate post-trade mid; aggregated to the market level. The direction reverses the canonical Glosten-Milgrom prediction.
Regression summary
| Model | n | β (flagged indicator) | t | R² |
|---|---|---|---|---|
| (1) has_sig_wallet | 204,772 | -0.173 | -175.7 | 0.100 |
| (2) sig_fraction | 204,772 | -1.674 | -21.7 | 0.016 |
| (3) + Controls | 204,772 | -1.354 | -20.3 | 0.126 |
Interpretation
The naive adverse-selection prediction (Glosten-Milgrom 1985) is that markets with private information should exhibit wider spreads, because market makers widen to recoup expected losses to informed traders. On Polymarket the pattern is reversed: spreads are tighter where flagged wallets trade. The most likely explanation, drawn from Paper 1, is that algorithmic market makers do not retreat from information-rich markets, they crowd into them. The information flow that should widen spreads instead attracts compete among market makers, and the net effect is tighter spreads.
This is not a refutation of Glosten-Milgrom; it is a different equilibrium under different market-maker incentives. Polymarket bots do not face the same inventory and capital constraints as equity dealers, so the standard widening response is muted or reversed.
Methodology
For each market with at least 50 trades, we compute mean effective spread across all trades, the presence of a flagged wallet (one-sided binomial z-test at p < 0.01), and the fraction of trades placed by flagged wallets. The reported regression is OLS of market-level mean spread on the flagged indicator, with no fixed effects in (1), with controls (volume, n_trades) in (3).
Limitations
The effective-spread measure is a single-trade approximation that does not account for changing quote depth or maker-taker fee schedules. The on-chain trade record does not include quote-level data, so a tighter (Lee-Ready style) measure is not feasible.
Source paper
Della Vedova, J. (2026). Detecting Informed Trading in Prediction Markets: An Orthogonality Test. SSRN. Section: Adverse Selection Test.
Data
surveillance_spread_latest.json
Other surveillance indices
← All surveillance indices · Informed Trading (PII) · Insider Timing · Resolution Surprise · Wash Trading
Cite this index
@misc{dellavedova2026spreads,
title = {Adverse Selection / Spread Pattern Index},
author = {Della Vedova, Joshua},
year = {2026},
url = {https://jdellavedova.com/surveillance/adverse-selection}
}