Adam Szymański

Problem 05 · Volume integrity

Most of the volume is decoration.

Estimated wash trading went from around 15% to around 60% in a year, and Columbia put it at 90 to 95% at category peaks. Every headline number is built on this.

Data Unglamorous Becomes the reference

Simple assumes nothing

Volume is the number everyone quotes. It is also the easiest number to fake. Trade with yourself across two wallets and you have produced volume without producing an opinion.

People do this because platforms reward activity with tokens and airdrops, so fake trading has a real payout. Estimates of how much prediction market volume is wash trading run from about 15% a year ago to around 60% now, and one Columbia study found 90 to 95% during the peaks of popular categories.

This matters beyond honesty. Any claim that these markets are accurate, liquid, or growing is measured in volume. Nobody has a clean number, which means nobody currently knows the size of the industry they are describing.

Moderate assumes you know what a market is

Wash trading has a straightforward detection surface on-chain. Circular flows, matched sizes, wallet pairs with reciprocal fills, activity clustered around incentive snapshots. Existing estimates use these signatures and disagree by a factor of six, which tells you the methodology is unsettled rather than that the truth is unknowable.

The consequence for research is worse than the consequence for marketing. Every study of prediction market accuracy, reflexivity, or liquidity uses volume as a control or a weight. If contamination is around 60% and uneven across categories, results computed on raw volume are partly measuring the incentive programme rather than the market.

There is no shared benchmark dataset with labelled wash trades. Building one is unglamorous work that would immediately become the thing everyone cites, which is a good trade for whoever wants a foothold in this field.

Technical state of the art and the gap

Frame it as classification on the trade graph. Nodes are wallets, edges are fills, and the target is subgraphs whose net position change is near zero while gross volume is large. Self-trading through an intermediary makes this cycle detection with a tolerance parameter, and the hard cases are collusive rings large enough to resemble organic flow.

The methodological gap is ground truth. Detection work validates against heuristics rather than confirmed labels, which is precisely why estimates span 15 to 95%. Two routes to labels exist. Incentive programme timing gives quasi-experimental variation, since wash volume should collapse when a reward ends while organic volume should not. And venues with no incentive programme give a clean control for what natural circularity looks like in the absence of a payout.

The target worth aiming at is a per-category, per-venue contamination estimate with an error bar, published as a dataset. Then re-run one existing accuracy result on cleaned volume and report whether it survives. That second step is what turns plumbing into research.

Where I would start

  1. Pull Polymarket trade history and build the wallet trade graph. Public, on-chain, no gatekeeper, and a weekend of work.
  2. Detect near-zero net position cycles carrying large gross volume. Start with pairs, then triangles, then larger rings.
  3. Use incentive programme start and stop dates as your natural experiment. The discontinuity at the stop date is your best available label.
  4. Publish the labelled dataset before publishing any conclusions. The dataset is the contribution that gets cited for years.

What counts as a result

Contamination estimates per category with error bars, plus one existing accuracy result re-run on cleaned data. If the result survives, the field gets more confident. If it does not, that is more important.

Related