Algorithmic Footprints in Prediction Markets: How to Spot Execution Algorithms in the Public Tape

# Algorithmic Footprints in Prediction Markets: How to Spot Execution Algorithms in the Public Tape

Kresmion Research

When a trader works a large order carefully, they leave a geometric trace in the public record of executed trades. Three shapes recur across order-driven markets, and prediction markets are no exception. The first is TWAP-style slicing: near-identical clip sizes released at regular time intervals. The second is absorption: a stream of one-sided orders soaked up by resting size while the price barely moves. The third is a bid ladder: resting orders parked at evenly spaced price levels. Kresmion runs detectors for these three shapes on the most active prediction markets and flags them descriptively on the Polymarket intel pages.

Here is the single idea to carry away. A flag describes the structure of order flow. It never tells you which way the price will go. Spotting careful execution reveals how someone is trading (in slices, passively, patiently), not which direction they expect. This paper walks through what the detectors saw in their first day of operation and, just as importantly, why a pattern that looks like an algorithm is very often a crowd of small bettors or a plain coincidence. Everything below is a historical observation from a narrow window, past patterns need not repeat, and none of it is advice or a forecast.

Why careful traders slice their orders

A large order that hits the book all at once eats through the resting liquidity and moves the price against the person placing it. That cost is called market impact, and it is real but sub-linear: empirically, the impact of a large parent order (a metaorder) grows roughly with the square root of its size relative to available volume. To avoid paying it, institutions cut a big parent order into many small child orders spread over time. The classic framing of this tradeoff, slice too fast and you pay impact, slice too slow and you carry timing risk, is the optimal-execution problem set out by Almgren and Chriss. A constant-rate schedule (equal slices at equal intervals) is exactly TWAP.

The same incentive exists on a prediction-market central limit order book. Polymarket runs a genuine order book where, in the venue's own words, "your order sits in the book until someone trades against it" (Polymarket documentation). Prices read as cents on the dollar, and YES plus NO sum to $1.00. Because prediction markets are real order-driven markets whose prices aggregate information reasonably well, the execution behavior that appears in stocks appears here too. You can watch live prices on the odds board.

Three fingerprints in the tape

TWAP: even clips at an even beat

Time-Weighted Average Price execution releases near-equal clips at roughly even time intervals, regardless of how busy the rest of the market is. Its fingerprint is near-identical clip sizes at a regular cadence. It is not the same as VWAP, which deliberately trades more when volume is high, so VWAP clips vary with the volume profile. Kresmion's cadence-regularity detector looks for the even-time-interval signature specifically.

Absorption: one-sided flow that the price shrugs off

Absorption describes resting liquidity soaking up a stream of one-sided aggressive orders without the price moving much. A common mechanism is an iceberg order, where only a small slice of the true size is displayed and it silently refills each time it fills. Hidden resting liquidity is not exotic. In traditional equity and electronic venues, studies have put icebergs at roughly 16% of executed shares on one European book (Frey and Sandas, 2017) and hidden orders at almost 12% of executions on a US electronic venue (Hasbrouck and Saar, 2002). Those figures are from stock markets, not prediction markets, but they show iceberg-style resting size is a normal feature of order books. The correct read of an absorption flag is descriptive: a lot of pressure met a wall of resting size and the price held here for a while. It is not a claim that the price will keep holding, because the resting size can be pulled at any moment.

Bid ladders: a staircase of resting orders

A ladder is a set of resting orders placed at evenly spaced price levels, often with similar sizes, so one participant is quoting a staircase of passive interest. This is an ordinary way to work a position or provide liquidity across a band of prices. The detector sees geometry (even rung spacing), not intent.

What the detectors saw on their first day

The tables below cover the detectors' first day of operation, 2026-07-15, from 08:31 to 21:37 UTC. Every flag in the snapshot was written that day. Trade ingestion itself reaches back to 2026-07-04, but no flags predate 2026-07-15, so this is roughly a 13-hour first run, not a two-week base rate. Treat the split as illustrative of one run, not a stable long-term rate.

The headline surprise is that clean TWAP is nearly absent. Despite being the textbook execution pattern, the cadence detector produced exactly one hit. The tape in this window carried far more one-sided absorption and resting-ladder structure than tidy clip sequences.

Three concrete examples, described by public market title only (Kresmion keeps wallet addresses out of published work):

The Fed absorption example sits on exactly the kind of rate market covered in our explainer on how prediction markets price the Fed.

Why a flag is not a trading signal

The most important statistical point is the base rate. When the thing you are detecting is rare, even a good detector produces mostly false positives. As an illustrative calculation from Bayes' theorem, a detector with 90% sensitivity and a 5% false-positive rate, run over windows where only 2% genuinely contain an algorithm, has a precision near 27%, meaning roughly 73% of its flags would be false alarms. Those inputs are assumed for illustration, not measured from Kresmion's detectors, but the lesson holds: a rare-event detector run over many quiet windows is like a sensitive smoke alarm, and most beeps are burnt toast.

The data shows why the wallet count matters. The wallet numbers a flag records are on the aggressing side, the flow that hits the resting size, and they are large: absorption flags averaged about 51 distinct aggressing wallets, with one flag reaching 630 and many sitting on sports longshots (for example an exact-score soccer market at 7 cents with 205 wallets). A diffuse crowd of order senders is not one carefully worked institutional order; on these markets it often looks like ordinary retail betting meeting a price that happened to hold. The count also says nothing about the resting side, because the detector never resolves who provides the absorbing liquidity. So an absorption flag points to no single actor on either side. It records one-sided flow meeting a price that did not move, and nothing more. You can explore the wallet-level view on the smart-money wallet pages.

Intent cannot be read off geometry either. The statutory definition of spoofing is "bidding or offering with the intent to cancel the bid or offer before execution" (CFTC fact sheet), and intent is exactly what separates manipulation from legitimate resting orders. Evenly spaced ladders are overwhelmingly ordinary liquidity provision. Kresmion presents flags as descriptions of order-flow structure, never as allegations. (Polymarket is not a CFTC-registered venue, so this source defines the intent principle, not any rule that governs the venue.)

Methodology and sources

Kresmion ingests the public trade tape and resting-order snapshots for the top 100 prediction markets by volume from Polymarket's public data API, polled on a roughly 30-minute cadence. Three detectors then run over that data: cadence regularity (TWAP), one-sided aggressive flow meeting a stable price level (absorption), and evenly spaced resting rungs (ladder). Each flag records the pattern, side, price, fills, notional, an evidence blob, and a detector-assigned confidence score, and writes to the `pm_algo_flags` table; a subset also surfaces as ALGO_FLOW rows in the user-facing signal feed. In this snapshot only 67 rows reached that feed across 5 asset symbols, so what users see is a small, symbol-gated fraction of the 266 raw flags. Wallet addresses are excluded from all published output.

Data window: all 266 flags were detected on 2026-07-15 between 08:31 and 21:37 UTC. Trade ingestion spans 2026-07-04 to 2026-07-15 (88,846 trades across 167 distinct markets). Coverage is 100 markets at any instant (confirmed on the latest holder-snapshot cycle) and 167 markets seen across the full ingestion history, because the top-100 set rotates as volume rankings shift. Numbers here are attributed to Kresmion's production database as of 2026-07-15. The full method framing lives on our methodology page.

Limitations

• One day, small sample. 266 flags, 74 markets, about 13 hours. Every rate and average is a first-run illustration, not a stable base rate. The 206/59/1 split should not be extrapolated. The flag table is live and has kept adding flags since this snapshot, which is itself a reminder that these are moving, provisional counts.
• No ground truth. Flags are unverifiable inferences about execution intent. The confidence scores (about 63 to 92 in this run) are the detector's own weightings, not a measured accuracy rate.
• Absorption names no single actor. The wallet counts (average 51, maximum 630) are on the aggressing side, so the flow hitting the resting size is typically a diffuse crowd, not one worked order. The detector does not resolve the resting side at all, so a flag never implies a single hidden institution on either side of the trade.
• Completeness, not cadence precision. The trade tape carries per-trade timestamps that are real to the second, so cadence within a single burst is measured precisely (the TWAP example's 24-second interval is not a guess). The real weakness is completeness across time: because ingestion polls roughly every 30 minutes rather than streaming, executions that straddle a poll boundary or exceed the data API's per-poll trade cap can be split or truncated, and a resting ladder that appears and then vanishes between two snapshots can be missed entirely. The limitation is coverage across poll boundaries, not the resolution of any single measured interval.
• The signal feed is sparser than the raw table. The 67 surfaced rows across 5 symbols are not one-to-one with the 266 flags across 74 markets.

Frequently asked questions

Q: Does a TWAP flag mean a big buyer is coming, so the price will rise?

No. Execution style is direction-agnostic and, by design, direction-hiding. The whole purpose of TWAP and iceberg execution is to minimize information leakage. From aggregated public tape you often cannot even cleanly label a print as buyer- or seller-initiated, so a flag is a structural observation, full stop.

Q: If a market shows absorption, is that price level "support"?

No. Absorption means one-sided flow met resting size and the price held there for a while. That resting size can be pulled at any instant, and an iceberg can vanish the moment it stops refilling. It is a description of what already happened, not a floor under future prices.

Q: Are these flags accusing anyone of manipulation?

No. The detectors see geometry, not intent, and intent is what the law uses to define manipulation. Evenly spaced ladders and layered resting orders are overwhelmingly legitimate ways to work a position. Kresmion labels structure and stops there.

Q: Does a flag tell me whether the market's price is accurate?

No, and the two questions are unrelated. Whether a market priced at 70% is well-calibrated is a property of many forecasts judged in aggregate, not of a single price, and it has nothing to do with whether an execution algorithm is present in the tape. This paper is about microstructure (how orders are worked), not forecast quality.

Sources

• Almgren, R. and Chriss, N. (2001). Optimal Execution of Portfolio Transactions. Journal of Risk 3(2). https://www.smallake.kr/wp-content/uploads/2016/03/optliq.pdf
• Wolfers, J. and Zitzewitz, E. (2004). Prediction Markets. Journal of Economic Perspectives 18(2). https://www.aeaweb.org/articles?id=10.1257/0895330041371321
• Polymarket Documentation. Prices and Orderbook. https://docs.polymarket.com/concepts/prices-orderbook
• Frey, S. and Sandas, P. (2017). The Impact of Iceberg Orders in Limit Order Books. Quarterly Journal of Finance 7(3). https://ideas.repec.org/a/wsi/qjfxxx/v07y2017i03ns2010139217500070.html
• Hasbrouck, J. and Saar, G. (2002). Limit Orders and Volatility in a Hybrid Market: The Island ECN. Working paper, New York University Stern School of Business. (Original source for the almost-12% Island ECN hidden-order figure.)
• Snowberg, E. and Wolfers, J. (2010). Explaining the Favorite-Longshot Bias. NBER Working Paper 15923 (published Journal of Political Economy 118(4)). https://www.nber.org/papers/w15923
• CFTC Office of Public Affairs. Disruptive Trading Practices fact sheet (statutory spoofing definition, CEA 4c(a)(5)(C)). https://www.cftc.gov/sites/default/files/idc/groups/public/@newsroom/documents/file/dtp_factsheet.pdf