Actionable Signals in Prediction Markets: Public vs. Private Information on Polymarket
TL;DR: We checked whether two signals anyone can see onchain, “the market has effectively resolved” and “someone is building a huge position under uncertainty,” are reliable enough for an automated agent to trade on. Once you count the false signals as well as the wins, neither one is.
Paper: Actionable Signals in Prediction Markets (ICAIF)
Authors: Oriol Saguillo (IMDEA Networks / UC3M), Fabio Giacomelli (Tor Vergata), Lucianna Kiffer (IMDEA Networks)
Setup
We reconstructed every matched trade, split/merge and transfer on Polymarket from Polygon, covering Jan 2024 to Jan 2026 (280M trades, $28B in volume). After liquidity filters, we kept 26,964 markets. We then evaluated two strategies, counting both their wins and their failures:
- Risk Arbitrage (public signal): buy the winning outcome once it trades at ≥ 0.94 and hold it until settlement.
- Information Asymmetry (private signal): flag outlier-sized positions (Tukey IQR) built at a VWAP below 0.70 and held for at least 48h.
Findings
- Risk Arbitrage loses money overall: $15.3M gross profit against $38.7M lost on false convergence, for a net of −$23.4M. It is net-negative in every topic. Most accounts (85.5%) come out ahead, but the median account makes $0.07, and the losses are 18× larger per failed market than the gains per successful one.
- The best arbitrageurs don’t win bigger, they lose less. The top account netted about $468K with under $15K in false-positive losses.
- Information Asymmetry is roughly break-even: $139.3M in profit against $131.5M in losses, for +$7.8M net. Large losing positions (2,018) outnumber large winning ones (1,522). About $125M of the profit comes from the 2024 US election.
- Profits are very concentrated in both cases, and they track position size more than predictive skill.
Takeaway
Analyzing only profitable wallets overstates both how reliable the public consensus is and how informative “whale” positions are. Concentrated profits alone are not evidence of an informational edge.
Feedback very welcome!