Kristian Fagerlie reviews the first five days of an automated GPT-6 trading bot that uses a Google DeepMind weather model to estimate New York City temperature outcomes on Kalshi. The system runs continuously on a virtual private server and chooses a position only when market prices meet its configured threshold for a conservative expected edge.
The self-reported results include a loss on the first day, gains on later contracts and one position described as effectively settled but not yet formally resolved. Kristian Fagerlie estimates the account was approximately $34 positive when unrealized positions were included, but the transcript does not establish the capital at risk, a percentage return or a fully settled profit after all fees.
For the live example, the bot freezes its weather forecast shortly before the market opens and compares both sides of each temperature range. It requires at least a five-cent modeled edge. When no offer met that threshold, the automation completed successfully without placing a trade, showing that its risk rule can override the pressure to participate.
Kristian Fagerlie explicitly cautions that five days cannot demonstrate a durable advantage and that the early result may be luck. The experiment shows that the automation is operating and following its configured rules, but it does not yet prove that the forecast model has a genuine predictive edge or that the strategy will remain profitable after more trades, fees and drawdowns.
Watch on YouTube
