Work

What 1,236 settled prediction market bets taught me about building a trading bot

Year: 2026
Type: preprint
Abstract: I run an automated Kalshi prediction market trading system from my Mac lab in Honolulu. Since its October 11, 2026 restart, the settled record as of November 11 is 1,236 real bets, with 1,115 wins and 121 losses. The win rate is 90.2%, and the net result is +$250.93. Those figures sit beside 66 bets that lost half their stake or more, including a gold bet at a loss of $61.44 and a SOL bet at a loss of $39.26. On October 13, the system was winning 75% of its bets and was net negative. This article argues that a useful trading bot must be explainable in accounting and operational terms. A complete result needs several columns, including counts, dollars, and loss severity. An accuracy only dashboard is inadequate because profit depends on cost, charges, and how gains and losses were distributed. Market coverage across crypto, commodities, and foreign exchange creates more questions to separate, and parallel AI development needs a firm boundary around live trading code. The observation window and settled scope should travel with any shared figure.
Source: Figshare
Author: Jeremy Koski
Language: English
Subfield: Finance
Open Access status: green
OpenAlex ID: W7225891179
Cites:
Cited by:
Related to: 10