Kalshi · MLB game winners · Sep 18 – Oct 6, 2026
Spinda Large vs Jev
Every two minutes of 143 baseball games, Spinda Large and Jev read the same live text: the score, the latest plays and the market’s order book. Each answered one question: what is the chance the home team wins? Neither model was given any number from our code.
One game, every two minutes
The chance the home team wins, from the market price and from each model. Click the chart to see the exact text both models read at that moment.
Input: what both models read
Output: chance the home team wins
Results over 143 games
Games where each model was closer, adding up
Each game is scored by how close each model’s chances were to the final result, across every decision in the game. The line steps up for whichever model was closer.
Every game
When a run scores
Average change in the scoring team’s chance of winning, from the last decision before the run.
Trade on these forecasts
A simple trader buys when a model’s chance beats the market price by at least the edge you set, after Kalshi’s fee. Temperature is Spinda’s setting for how bold its chances are: below 1 is bolder, above 1 is more cautious. Here it applies to both models.
Run your own test
Every decision from Sep 27 to Oct 7, with both models’ answers, the market price and the result.
Same input for both
- At each decision both models read the same text: the contract and its rules, the time, the live score and recent plays from Kalshi’s game feed, and the market’s order book and trades.
- No number computed by our code goes into either model’s input.
- Jev is jev-1.13.0 through TypeSafe’s API. Spinda Large is decision-model-large.
Locked before the test
- The input format and the trading edge were set on Sep 19–26. Sep 27 – Oct 7 was run once, after that.
- All games are after Spinda’s training data ends (Sep 12).
- Trades buy at the quoted price and size with Kalshi’s fee. No resting orders and no market impact.
Limits
- A replay of archived games, not live trading. No claim about future profit.
- One sport over three weeks. Small gaps are noise.
- On Bitcoin 15-minute markets, which are all numbers, Jev did better, on an input format first tuned on Jev.