Research methods
Polymarket Research Methods
Research on Polymarket succeeds when the method matches the record: settlement studies want 250ms final minutes, reprice studies want the tape and book on the same grid, and capacity studies want depth history. This page catalogues the methods that work.
Figures measured as of 2026-09-24 on the published PolyOrderbooks archive.
Methods
Method families that work
Design
Design rules
Every method shares three rules: define the sample before the analysis, report the flag treatment, and validate against the recorded grid — a study that cannot say when each row was captured has not designed its data.
The measured baseline makes effect sizes visible: a settlement study that finds one-sidedness in final minutes is confirming a 76.2% baseline, not discovering it, unless it isolates a subgroup.
Worked example
One study, end to end
A settlement-surprise study: define surprise as "winner side outweighed by loser side at t-minus 60 seconds", pull the corpus frames, compute the flag and depth distributions split by surprise/normal, and report with the window and sample stated.
The honest output is the distribution and the denominator — the methods page exists to keep the design from morphing into cherry-picking mid-analysis.
Honest
The honest framing
The best method on this record is the one that can be rerun: fixed sample, stored frames, published flags. This catalogue is organized around that so each family ends with a reproducible workflow, not a narrative.
FAQ
What are the proven research methods?
Settlement, reprice, one-sidedness, capacity, and fidelity studies — each with a corpus rule and a reproducible workflow.
What breaks a Polymarket study?
Undefined samples, unfiltered flag reporting, and conclusions that cannot be re-run against the recorded grid.
What is the baseline to check against?
One-sidedness at 16.9% of 5-minute snapshots and 76.2% of final minutes — effects must beat the baseline, not echo it.