- Spread +
depthProxy - Scheduled eventsAround
- High-frequencyResolution
- Event studyApproach
Adverse selection is the market maker’s nightmare: trading against someone who knows more than you do. Makers cannot see who is informed, so they defend the only way they can, wider quotes and thinner books when the risk is high. Around scheduled events, that defensive response is observable, and the order book is where you read it.
This is a microstructure research piece, so two caveats up front. First, we measure the book’s response, spread and depth, as an observable proxy for adverse-selection pressure; it is a window onto the phenomenon, not a direct ledger of who traded on what. Second, the value is in the method, which you can reproduce on any event in the data.
What the book does when risk rises
Spreads widen
Facing possible informed flow, makers quote further apart to be compensated for the risk of being picked off.
- Defensive quoting
- Wider into the event
- A risk premium you can see
Depth thins
Makers also shrink size, pulling resting orders so less can be taken at any one level.
- Smaller resting size
- Less to pick off
- Fragile top of book
Then it normalises
After the information is public and symmetric again, quotes tighten and depth returns toward baseline.
- Post-event recovery
- Information symmetric
- Back to calm
Running it as an event study
- 1Pick events with known times, an FOMC decision, a payroll print, a game’s kickoff, a crypto market’s bell.
- 2Pull the snapshots across a window that brackets each event at high frequency.
- 3Track spread and depth relative to each market’s own pre-event baseline, not an absolute number.
- 4Aggregate across many events to separate a systematic pattern from one noisy instance.
The fields the study reads
The measurement leans on a handful of fields that ride on every snapshot. Bracket the event on Polymarket’s own emit time, not your processing clock, so the same event aligns across markets and across the many instances you aggregate.
- spreadThe maker’s risk premium, pre-computed
bid_depth_totalask_depth_totalHow much size is exposed to be picked offevent_timestampPolymarket emit time, the alignment axis- 700M +
snapshotsArchive deep enough to aggregate many events
A window, not a ledger
We are clear about what this is: the spread-and-depth response is an observable proxy for adverse-selection pressure, built from the order book we capture. It is a rigorous, reproducible lens, and it is not a claim to see every trader’s private information. Saying exactly what the data can and cannot show is what makes the study trustworthy.
Why reproducibility is the whole point
A microstructure claim is only as good as the next person’s ability to re-run it. Because the snapshots are the same fields for every event, anyone can take your window definition and baseline rule and get your numbers back. That is what separates a measured finding from an anecdote, and it is why the method, not a single chart, is the deliverable.
Same fields, every event
Spread and depth are recorded identically across categories, so one event-study recipe runs unchanged from crypto to weather.
- Uniform schema
- Cross-category
- No bespoke parsing
Baseline, not absolutes
Every response is scaled to the market’s own quiet pre-event window, so thin and deep books are compared honestly.
- Per-market baseline
- Relative response
- Comparable across events
Aggregate to find signal
One event is noise. Stacking many on a common emit-time axis is what makes a systematic response visible.
- Many events
- Common axis
- Signal over noise
Makers cannot see who is informed, so they defend the only way they can, wider quotes, thinner books. The order book is where that defence is written down.
Run your own event study
Pull bracketed event windows with the historical guide, or replay an event tick by tick to see the response form.
Frequently asked questions
What is adverse selection in a market-making context?
It is the risk of trading with a counterparty who has better information, leaving the market maker on the wrong side. Makers cannot identify informed traders, so they defend by widening spreads and reducing depth when adverse-selection risk is high, such as around scheduled events.
Can Resolved Markets show me who traded on inside information?
No, and we are careful not to claim that. What the order-book snapshots let you measure is the makers’ response, how spread and depth move around events, as an observable proxy for adverse-selection pressure. It is a rigorous lens on the phenomenon, not a record of anyone’s private information.
How do I run an event study on this?
Choose events with known times, pull high-frequency snapshots across windows that bracket them, measure spread and depth against each market’s own pre-event baseline, and aggregate across many events so a systematic pattern stands out from noise.
Should I bracket events on event_timestamp or capture_timestamp?
On event_timestamp, Polymarket’s own emit time, so the same event aligns consistently across markets and across the many instances you stack together. capture_timestamp records when we processed the update and is useful for gauging latency, but using it as the alignment axis would let processing jitter blur the response you are trying to measure.