The Graveyard. Forty experiments on Kalshi and the one that survived.
A month of automated trading experiments, almost all of which failed, and the single boring strategy that earned real money.
Over the past month this lab ran about forty automated trading experiments on Kalshi. Almost all of them failed. This is the graveyard, and the one thing that walked out of it.
Every strategy here was tested the same way. First against a recorded archive of roughly 18,000 settled fifteen minute markets. Then, if it survived, in forward paper trading against live order books with honest fill rules. Only after both gates does anything touch real money, and then only with single contracts until the live results earn more. Most ideas never make it past the first gate.
What died
Buying longshots. The cheap side of these markets is systematically overpriced. A contract priced at 17 cents settles true about 4 percent of the time. We tried several variants. All of them lost.
Taking favorites at the ask. An 84 percent win rate that still loses money. Crossing the spread costs about 6 cents per fill. We proved this the direct way, by running two bots on the same signals at the same time. The one resting orders made money. The one paying the spread lost 6 cents a fill over 277 fills. Win rate is not edge. Price is edge.
Following whales. We replayed more than 4,000 large trades at every reaction speed we could manage. Negative at all of them. A commenter later measured why. The order book reprices in well under a second, before the big trade even shows on the tape. By the time you can see the whale, you are the liquidity.
Momentum between BTC and the alts. Real in the data, and already harvested by someone faster. Our probe paid a fee and slippage toll on every attempt and never covered it.
Buying the dip on crashing favorites. Favorites that crash below 50 cents still recover and win about 70 percent of the time. Sounds like free money. We built a bot to buy exactly that dip at real asks. It lost 8.3 cents per fill over 662 trades. The panic price already knows. Everyone smart is selling to you at the moment of maximum fear.
Stop losses on favorites. Cutting a dipped favorite costs about 20 cents of expected value versus just holding it. Same lesson from the other side. The dip is the mispricing, and selling into it is the mistake.
Selling winners early. After a reader asked, we replayed 15,808 past favorite trades with early maker exits at 95, 96 and 97 cents. Every version underperformed holding to settlement. Exiting at 96 dodged 149 late collapses but gave up 4 cents on more than ten thousand winners. The boring answer keeps winning. Never sell. Just wait.
Esports. Six mechanical strategies on CS2 and Valorant markets. All six negative. A passive market maker there breaks exactly even after 440 fills. Those markets are sharper than they look.
What survived
One boring thing. Resting maker bids on heavy favorites in the final minutes of fifteen minute crypto markets.
The mechanism is simple. A confirmed favorite priced in the eighties wins about 87 percent of the time, which makes paying the ask roughly break even. The entire edge is the two or three cents you save by resting a bid instead of taking. It only exists because makers trade free on these markets, and it is small, a few cents per fill, harvested many times a day.
That strategy passed the archive, passed forward paper, went live with single contracts, and only after about 900 real fills stayed positive did it earn larger size. It has been trading real money for two weeks across roughly 1,300 fills, and every one of those fills is verified against the exchange's own records.
What the graveyard is for
Half of our failed ideas looked great in the archive and died the moment they traded forward. That is why nothing here gets promoted from a backtest, and why the failures get published next to the wins. The losers are where all the useful information turned out to be.
Three new experiments entered the lab this week, all suggested by readers. A fat finger net resting silly offers across sleepy low volume markets. A pair completion maker that rests both sides and locks the spread when both fill. And a match clock lane that reads live esports game state from an official API, chasing the one esports edge our data ever validated. All three are paper only, all three face the same bars everything else did, and all three will be reported here whether they live or die.
The lab publishes its full results every Sunday, including the losses.