REFUTED: the Kalshi weather maker edge, and the liquidity-provision mechanism behind it

The claim that died

A maker (liquidity-provision) edge in Kalshi weather markets: +1.37c/contract net of full fee on KXHIGHNY, 79.7% of markets profitable, day-clustered t = 2.02 — with a stated mechanism of thin volume -> less market-maker competition -> wider spreads -> liquidity provision pays.

Two pre-registered tests, one narrow and one broad. Both refuted it.

Test A — out-of-sample weather cities: REFUTED

Six fresh cities, gates fixed in advance, strict parameter lock (identical fee model, P&L arithmetic, estimator):

Pooled: -1.61c/contract (60 city-day units), t = -3.92, p = 0.0001. Zero of six cities positive. Every gate failed — W1 (significance), W4 (majority + veto), W5 (decay), W6 (fill realism, both readings).

The sign did not merely fail to replicate, it inverted: in-sample +1.15c became out-of-sample -1.61c, significantly negative. Makers in weather markets lose money.

Test B — the mechanism acrossseries: REFUTED

If the edge came from thin liquidity, maker edge must fall with volume everywhere, not just in weather. Tested across 42 series spanningcategories and ~4 orders of magnitude of volume (558 to 2.58 M median contracts):

OLS   maker_net ~ log10(median volume)

      slope = -0.105c per decade    t = -0.15    p = 0.88    R^2 = 0.001    n = 42

No relationship whatsoever. R^2 = 0.001. Leave-one-category-out is unstable in sign (dropping Commodities, Economics or Sports flips the slope positive), failing M2 as well.

Per-series edges range from +9.72c to -13.54c with no volume pattern — the spread is noise, not structure. That is itself informative: if series-level maker edge estimates are this noisy at 30-60 markets each, the original +1.37c weather reading was almost certainly noise too, which is exactly what Test A found.

What this kills, and what survives

Dead: the weather maker edge; the liquidity-provision explanation; and the earlier suggestion that "we have been trading the wrong side of the weather market." That inference was built on the in-sample number and does not survive.

Survives (and is now better evidenced): takers lose. Measured with settlement ground truth on 1.49 M real executed trades: MLB -1.41c, crypto -0.77c, weather -2.84c per contract. And makers earn positive gross in the original three series while still losing net in two of three — the spread they capture generally does not cover the fee. The venue's fee is the binding constraint, on both sides.

Why the in-sample result looked real

Three compounding effects, all now documented:

1. Selection. Weather was the best of three series — Pav (2606.01650)

shows the selected max is an upward-biased estimate.

1. Clustering.markets were onlyindependent days (strike ladders share one temperature

outcome). Correcting the unit left t = 2.02, which atdf is p ~ 0.07 — never significant.

1. Small effective n. Ten days is not enough to distinguish +1.15c from zero given a per-day

sd of 1.79c.

The pre-registration is what caught it. The gates were written before the out-of-sample data was scored, committed with a sha, and not rewritten when they failed.

Process note

An earlier draft of these gates invented every threshold (t >= 2.5, 65% of markets, +0.5c). They were replaced with protocol elements from the literature — Pav for post-selection shrinkage, the AlgoXpert IS-WFA-OOS protocol (2603.09219) for staging, parameter lock, majority-pass and catastrophic-veto. The refutation is cleaner for it: the gates were not mine to bend.