Kalshi AI Trading Bot Speedrunning - Can We Profit In 12 Hours?

Summary

The video documents a speedrun experiment designed to test how quickly an automated trading bot can be built and deployed on Kalshi prediction markets using OpenAI Codex running GPT-5.6. Starting from an empty environment, the presenter sets up historical market data for Kalshi 15-minute Bitcoin contracts by feeding documentation endpoints directly into the Codex CLI environment.

To identify a viable edge, Codex uses a pre-configured quantitative hypothesis research skill. This workflow inspects historical contract definition endpoints, verifies timestamps, and formulates testable market hypotheses. Five specific hypotheses are evaluated: volatility persistence, weekend volatility, outcome alternation, minute-five under-dispersion, and activity cycles. Backtesting rejects outcome alternation and minute-five under-dispersion. Volatility persistence, designated as hypothesis H-001, demonstrates positive out-of-sample forecast improvements across multiple volatility thresholds, producing an estimated frequency of roughly thirty-two trade setups per day when restricted to the top third volatility filter.

Codex is then tasked with writing production execution code based on the validated volatility persistence specification. The logic requires checking the immediately preceding fifteen-minute Bitcoin contract to ensure an absolute BRTI move of at least 15.3975 basis points, or approximately 0.154 percent. If this condition passes, the algorithm evaluates the contract's YES midpoint price at minute five, choosing YES if the price is above fifty cents, NO if below fifty cents, or skipping if exactly fifty cents. After a one-minute delay, at minute six, the bot dispatches a five-dollar fill-or-kill order using Kalshi WebSocket market data rather than REST API polling.

The algorithm is deployed into a live testing environment using credentials tied to a funded Kalshi balance of approximately $125. The presenter confirms that WebSocket lifecycle messaging and order-state databases are operating cleanly before leaving the system active overnight.

The following morning, performance logs reveal eighteen completed trades consisting of fourteen wins and four losses. The final strategy profit and loss reached $15.44 with an account balance rising to $137.11, representing a 15.73 percent gain over the twenty-four hour window. Analysis of trade history indicates a maximum realized drawdown of $12.73 after trade five before recovering into positive equity.