Multi-leg pricing analysis · Kalshi
Most of the time, the answer is no.
ComboCrusher reads Kalshi's public order books, prices every multi-leg combination against the product of its individual legs, and measures the gap. On a typical day it finds nothing worth acting on. That is the honest result, and it gets published either way.
Illustrative sample using representative market structures. Live output is published on the record.
Correlation is priced unevenly.
When you combine legs into one contract, the fair price is not the product of the legs unless those legs are independent. They rarely are. A moneyline and a game total move together. Two scorers in the same match share the same game script.
Kalshi's pricing charges for the obvious relationships — spread with total, a side with its own total — and charges heavily. What it handles less consistently is the subtle stuff: player-level correlation, legs that share a game script without sharing a market, marginal pricing that shifts as legs are added to a ticket.
ComboCrusher does one thing: it measures that inconsistency continuously, at a scale no one is going to match by hand, and records what it finds before the outcome is known.
There are no picks here.
Nobody needs another account posting screenshots of winners and quietly deleting the losers. What is actually scarce is a pre-committed, timestamped record of estimates made before resolution, published in full including the bad ones.
- No tips, no plays, no units. The output is a pricing discrepancy and the reasoning behind it. What you do with that is your call.
- No win rate marketing. Win rate is close to meaningless on binary contracts priced across the whole 1–99¢ range. The metrics that matter are calibration, net ROI after fees, and realized versus modeled edge.
- No claim of a proven edge yet. The sample is too small. It says so on the record page, in the same size type as everything else.
- No affiliation with Kalshi. Independent project, public API, public order books.
The pipeline runs whether or not it finds anything.
Ingest the day's markets. Normalize prices and unify markets that belong to the same game. Score candidate combinations for genuine correlation, separating it from mere redundancy. Verify against the live order book. Alert only if it clears the threshold and the price hasn't moved since scoring.
Every run writes a row, including the runs that find zero candidates. A pipeline that only logs its wins is not a record, it's a highlight reel.
Pre-track-record.
The pipeline is built and running daily. The sample of graded, pre-recorded estimates is not yet large enough to separate skill from variance, and it will not be for a while. Several hundred resolved positions is the floor for saying anything honest about calibration.
Until then the useful thing on offer is the build itself: what the scanner finds, what it rejects and why, and the running log of what breaks. If that's interesting to you, the list is the place to follow it.