Our $106k Kalshi "Bailout," Explained

The game

Western Michigan at Michigan, Saturday, September 5, 2026. WMU led 12–7 with the clock at 0:00 after a Michigan Hail Mary was intercepted. Officials ruled a WMU defender touched the ball out of bounds with one second left, put 0:01 back on the clock, and Michigan scored a 47-yard touchdown on the final play. Final: Michigan 13, WMU 12. Press reports put Kalshi volume on the game at $18.6M.

Kalshi settled the market when the clock hit zero, before the review.

X Post

On September 7th, I posted this on X:

Unbeknownst to me at the time, Kalshi had already published a notice identifying the affected combo trades for review and cancellation:

Kalshi's exchange notice, dated September 6, 2026:

Pursuant to its Rule 5.11 ("Trade Cancellations"), KalshiEX LLC hereby gives notice that trades meeting the below description made after 10:50:56 PM ET, September 5, 2026 are under Rule 5.11 review and shall be subject to cancellation: Trades in multivariate events containing the component event KXNCAAFGAME-26SEP05WMUMICH in which one component outcome was WMU Yes or Michigan No.

On September 10th, the credit landed in our accounts, giving us confidence for continued operations and scaling.

Over the week, the post drew many colorful replies:

What stood out was how widely three things were misunderstood: how settlements and combos interact, why market makers keep quoting combos after settlement, and why we think unwinding these trades supports a more reliable and liquid exchange. The rest of this post takes them in turn.

On Market Making Combos

There are at least two conceptual simplicities that combo market makers rely on to quote efficiently.

1. Combo MMs specialize in pricing the copula, or inter-dependency, of the legs, not the legs themselves.
2. Combo MMs rely on the finality of settlement to residualize (i.e. simplify) combos, permitting tighter spreads and a broader quotable universe.

Take a toy example: a combo $C$ with legs $X$, $Y$, and $Z$, each with a fair value of 50c (50%) and completely independent of one another. $C$ is worth $0.5 \times 0.5 \times 0.5 = 0.125$ (12.5c).

Now suppose the MM believes $Z$ is mispriced and its fair is really 90c. It could offer $C$ at 23c, but it's far more efficient to buy up liquidity on $Z$ directly, shift that book to 90c, and keep quoting $C$ at $X \times Y \times Z$. The combo marketplace is a bad place to capitalize on directional alpha in the legs, for takers and makers alike. Takers pay two MMs for the pleasure of acquiring exposure. Makers sit around waiting for combos that happen to include $Z$. Both end up with large directional exposure to events they are entirely uninterested in.

The two layers, and where the alpha goes:

Now game night concludes and $Z$ settles in favor of the combo. An MM trusting the finality of settlement can ignore $Z$ and price off $X$ and $Y$ alone: the new fair of $C$ is $0.5 \times 0.5 = 0.25$ (25c). Retail holders can liquidate early to take profit, at spreads tighter than when they bought, since the MM is compounding uncertainty twice instead of thrice. If $Z$ was a dependence curveball such as a same-game prop, many more MMs may emerge to quote $C$ post-settlement, driving costs for retail traders down further.

The toy example in one picture:

Panel 3 is what happened to us. The exchange said $Z$ (WMU) settled YES, we residualized and kept quoting $C$ off of the remaining legs, and then $Z$ became NO.

The implication for exchange participants: event MMs and combo MMs can run disjoint, independent strategies, provided the underlying exchange data is available and deterministic. Providing liquidity gets conceptually simpler, more MMs show up, retail flow gets tighter spreads, and everyone is happier.

What we quoted

All 36 finalized NO on September 10 between 16:15:23 and 16:16:13 ET, with revenue = 0 on every one.

Raw outcome: $9.20 (released when the 9.20 NO contracts netted against YES) − $106,260.84 − $442.88 = −$106,694.53. Fills settle in fractional cents and totals are summed before rounding, so displayed figures can be off by a cent.

The big ones

All tickers are prefixed KXMVECROSSCATEGORY-SHARD1-S2026…. "Residual leg(s)" are the legs still undetermined when our RFQ responder quoted; every other leg had already resolved YES, and WMU was being treated as resolved YES too. The price of the combo is roughly the product of the residual legs' odds, which is why the three-residual combo priced at 0.969 and most single-residual ones at ~0.99.

We were not trying to exploit the bad settlement. Our weighted average expected edge across these executions was under 1%; in normal conditions this volume earns about $1k. Fills at ~99c can only ever make 1c on settlement. The doctrine at Greed is capital scale and regime-independence. We're looking for a long buck, not a quick one.

What Kalshi credited

Source: Kalshi account activity export

At 16:36 ET on September 10, Kalshi credited $106,680.24, corresponding to the post-cutoff fills and fees.

So the refund covers the post-cutoff fills, including fees, to within an unreconciled $1.40. The two pre-cutoff fills ($12.89) were excluded, consistent with the notice.

The $9.20 column matters more than its size suggests. Those two NO fills were our only trades in the cancellation window on the winning side of the bug: they closed 9.20 YES at 63.5c on the WCU combo, a $5.84 gain. The credit nets that gain out. Had Kalshi simply reimbursed losses, the residual would be $7.16, not $1.40.

So our reconciliation fits a cancellation of the affected trades, winning executions included, rather than a reimbursement of losing ones. That is what a Rule 5.11 cancellation means: both sides of each trade are unwound, so by construction the exchange is not funding one side's losses out of its own pocket. We can only see our own account and cannot verify the counterparties' debits or the exchange's total cost. But on our side of the ledger, this was not a bailout in the traditional sense. The trades were put back to the state before the exchange's own bad data.

The two models side by side:

How big was the unwind exchange-wide? From the public trade tape, joined to combo legs from the RFQ broadcast feed: after the 22:50:56 cutoff, combos containing WMU yes or Michigan no traded ~4,000 times across ~1,350 combos, 18.9M contracts, roughly $19M of cash between the two sides. Our $106K was about 0.6% of it. Nearly all of it printed in the hour after the bad settlement. This is an estimate: combos whose RFQ we did not observe are missing, and we cannot see which trades Kalshi ultimately cancelled.

Conclusion

I hope this post clarified exactly how these losses materialized, and why an exchange that wants to be diverse and liquid would reverse them.

Appendix

The bad settlement, as seen through the Kalshi API

KXNCAAFGAME-26SEP05WMUMICH has two markets: -WMU and -MICH. Here is what the exchange's own interfaces said about them, in order. All times in this post are ET (EDT, UTC−4). Times are when we observed each message or response; Kalshi's internal transition times may be earlier.

The same sequence, row by row: