Graveyard / Prediction markets / Prediction markets
Trading Kalshi prediction markets
Can you beat Kalshi's hourly Bitcoin, economic data or weather markets?
The idea
Prediction markets pay $1 if an event happens. Three ideas: compare hourly Bitcoin contracts with a fair-value model from the Binance price, trade economic data contracts right after the release, and buy weather contracts for outcomes that are already decided.
How we tested it
| Data | Test 1: the Kalshi trade archive, about 6,500 trades at 14 per day, taker orders held to the hour or until the gap closes. Test 2: 372 Kalshi economic markets (CPI and jobs). Test 3: 5,782 decided weather markets over 15 months. |
|---|---|
| Period | Test 3: 15 months. Others not stated. |
| Costs | Taker prices. |
| Holdout | Two halves. |
| Control | The market's own probability (Brier score). |
Result
| Test 1: market Brier score | 0.176 |
|---|---|
| Test 1: model Brier score | 0.187 |
| Test 1: per trade, every cell | −3.8% to −6.2% |
| Test 2: markets that closed before the release | all (close at 08:25 ET) |
| Test 3: decided markets that traded again | 63 (at 98-99c) |
| Test 3: buyable at 97c or less | 16 trades; halves −8.1% / +34.1% |
Why it failed
The market's own forecast was better than the model's (a lower Brier score is better), and both halves lost money. Economic data markets close before the release, so the idea is structurally impossible. The weather idea was close to a free lunch in theory, but only 63 of the decided markets traded again and a 'certain' outcome was wrong once because two weather data sources disagreed.
What would make it worth retesting
Only with a model that beats the market's Brier score out of sample.
Past results of a backtest say nothing certain about the future.