An edge in five markets, and we can only collect it in one

By The Heron. Figures from Kalshi's own public trade tape,
re-graded on 20 September 2026.

Kalshi runs fifteen minute markets on natural gas, silver, oil, copper and
gold. Every one of them closes on a price print, and in the closing stretch
one side of the book is the favourite and the other is the longshot.

Whoever is resting on the favourite side late in those markets has been
getting paid. We measured it on two thousand settled markets, four hundred
per series, across six days and several million real prints. Every series
came out positive.

The number that makes this credible is not in the table. It is the mirror.
The traders taking the other side of those same contracts lost between
three and six and a half cents each, on zero or one day out of six, with
the same statistical weight. What one side gains the other loses, to the
cent. That is the signature of a real transfer between two groups of
people rather than a quirk of how we measured it.

So there is an edge worth three to six cents a contract sitting in five
public markets, in a pool carrying roughly three million contracts a day.

Then we tried to pick it up

We ran a resting order into those same five markets under live conditions
for four days, on paper, filling only when a real print came to our price
or through it. If the edge belongs to whoever rests on the favourite, a
resting order should collect it.

It collected in one market out of five.

Natural gas paid better than the tape suggested. The other four paid
nothing, or took money off us, in exactly the cell where the tape says
there is money to be made.

Two explanations, both killed by our own data

The first was that the edge lives in thin books, where a resting order is
alone at its price instead of queued behind professionals. It is a tidy
story and it is wrong. Copper has the thinnest book of the five and loses.
The apparent relationship between depth and edge rests on gold being both
the deepest book and among the worst, which is one data point wearing a
disguise.

The second was that our order sits one tick better than the best bid, and
that in a market quoted a penny wide, one tick better than the bid is
simply the offer, which is the losing side. Also wrong, and we could check
it directly: every one of the five enters at the same median spread,
because the rule will not trade a tighter one. Strip out natural gas and
the result is negative at every spread we traded.

So natural gas is one market out of five, with no mechanism behind it, on
a sample small enough to be luck. We are not going to trade it on that
basis, and saying so is the entire point of publishing this before the
verdict rather than after.

What we think is actually happening

The tape measures everyone resting on the favourite side, at every price,
for the whole closing stretch. Our order rests at one price, for a few
minutes, and only fills when the market comes to it. A resting bid gets
hit when the price falls toward it, which means our fills are selected for
having moved against us before they happened. The pool average includes
all the makers who got filled without that happening.

If that is right, the edge is real and ours to lose on execution, not on
the idea. That is a better problem to have than the reverse, and it is a
solvable one.

What each book quotes is public and how it quotes is not; the reasoning
is on the About page. The weekly board is in
State of the Lab, and the
sealed fill ledger is at /proof/.

The paid edition of this study contains: then we tried to pick it up; two explanations, both killed by our own data; what we think is actually happening; the capture numbers in full; what happens next. It is $15 a month, and the annual plan is two months free. Read the paid edition. Not sure yet? Read it free for seven days.