←BACK TO BLOG
>

Polymarket's Trade-Flow Data Is Barely Better Than a Coin Flip

Prediction MarketsMarket MicrostructurePolymarket
JUL 2026·5 MIN READ

#THE LABEL EVERYONE TRUSTS

I read a lot of Polymarket commentary that leans on a specific move: pull the public trade feed, tag each print as a "buy" or a "sell" based on which side of the book it hit, and then build a story on top: smart money is accumulating YES on this Fed contract, retail is dumping NO on that election market. I've done versions of this myself. It feels like free alpha, because Polymarket is fully on-chain and the data is just sitting there in the API. The problem is that the label underneath the story is close to a coin flip, and almost nobody checks it against the one thing that would actually confirm it: the settlement record on Polygon.

#THE NUMBER: 59%

Philipp Dubach's paper, "The Anatomy of a Decentralized Prediction Market" (arXiv 2604.24366), builds a continuous, tick-level archive of Polymarket's public order-book feed (about 30 billion events over 52 days) and joins it to the authoritative on-chain trade record, the OrderFilled events that actually settle each trade. Across a pre-registered panel of 600 markets, feed-inferred trade direction agrees with on-chain settlement truth on roughly 59% of buckets, a panel mean of 0.615, 95% CI [0.58, 0.65].

Compare that to the Lee-Ready algorithm, the standard trade-classification method for equities, which hits about 80% accuracy on Nasdaq data. Nasdaq doesn't even give you ground truth to check against; the tape is the closest thing to it. Polymarket does, and the public-feed inference still undershoots it by more than twenty points.

#WHY THE FEED LIES

The root cause isn't noise, it's a field-semantics mismatch. Polymarket's public feed has a field called change_side that a lot of people assume tells you who initiated the trade. It doesn't; it marks which side of the order book moved, not which side was the aggressor. A resting limit order getting hit and the order that hits it sit on opposite sides of "who moved," and if you're inferring buyer-initiated versus seller-initiated flow from book-side movement instead of from the actual taker, you will misclassify maker and taker direction constantly. It's a subtle bug to fall into, and an easy one, because the field is right there in the WebSocket payload looking like exactly the thing you want.

Trade-Direction Accuracy: Feed-Inferred vs. Ground Truth

50Random guess59Polymarket feed-inferred (CI 58-65%)80Lee-Ready, Nasdaq equities

Polymarket's feed-inferred trade direction (59%, 95% CI 58-65%) sits only 9 points above a coin flip and trails the Lee-Ready equities benchmark by 21 points, even though Polymarket alone has on-chain settlement data to check the answer against.

#WHAT BREAKS WHEN YOU TRUST IT

This isn't a rounding error that averages out over a large enough sample. Dubach shows that swapping feed-inferred direction for the on-chain record flips the sign of standard microstructure metrics on a large share of markets, not just their magnitude:

A sign flip on Kyle's lambda means a market you'd call "informed trading dominant" is, on the correct data, the opposite. That's not measurement error; that's your conclusion turning into its own rebuttal.

  • ▸Effective half-spread flips sign on 67% of markets in one 7-day window, falling to 50% in a later, non-overlapping window
  • ▸Kyle's lambda (the standard measure of price impact and informed trading) flips sign on 60% of markets in the same comparison, falling to 43% in the later window
  • ▸Median self-counterparty (wash-trading) share sits at 1%, but has a 22% upper tail worth knowing about if you're weighting flow by volume
  • ▸Median archive-ingestion delay is sub-50ms, with a multi-second tail, so even the timing of a "print" you're reacting to can lag true settlement

#THE FIX ISN'T STATISTICAL

The honest reading of this isn't "add a correction factor." It's that the public feed is the wrong data source for the question, full stop. Dubach's actual remedy is architectural: source aggressor and trade-direction from the on-chain OrderFilled events directly, since that's the record that settles money, rather than reconstructing intent from a book-side field that was never designed to encode it. He's also published the replication package on GitHub (philippdubach/polymarket-microstructure) with the full 600-market panel and code, so the 59%-versus-80% number isn't a black box; anyone making claims about Polymarket flow can rerun it against the same ground truth.

#THE PART THAT ACTUALLY BOTHERS ME

In equities, you can forgive sloppy trade classification because there's no better ground truth available; Lee-Ready's 80% is close to a ceiling, not a floor someone chose not to clear. Polymarket has no such excuse. Every trade settles on-chain, publicly, permanently. The "smart money" threads, the flow dashboards, the informed-trading writeups: all of it is checkable against a record anyone can query for free, and mostly nobody does. If you're going to build a narrative on order flow in a market this transparent, the bar isn't "does the feed look directional," it's "did you check it against settlement." Right now, most of what circulates hasn't, and it's built on a label that's barely better than guessing.

Sign Flips: On-Chain Truth Overturns Feed-Inferred Metrics

67Half-spread: Window 1 (7-day)50Half-spread: Window 2 (later)60Kyle's lambda: Window 1 (7-day)43Kyle's lambda: Window 2 (later)

Swapping feed-inferred direction for on-chain truth flips the sign of effective half-spread on 67% of markets and Kyle's lambda on 60% in the first 7-day window, more than a coin flip's worth of markets reversing conclusion, with the flip rate easing but still substantial (50% and 43%) in the later, non-overlapping window.

#Further Reading