Kristian Fagerlie trains a machine-learning model on 5,335 resolved Bitcoin markets from Kalshi. The model observes the first five minutes of each 15-minute window, then predicts whether the final price will finish above or below the opening level. A separate 600-market holdout set keeps the initial evaluation outside the training data.
A 16-window live test produces 12 wins and four losses, including a striking streak after the entry size rises from one dollar to five dollars. Kristian Fagerlie treats that result cautiously instead of presenting it as proof. The model only nudges the estimated probability slightly beyond the market price, so a short lucky sequence can look much stronger than the underlying signal.
A larger quant-style simulation removes the excitement from the small live sample. Across roughly 1,000 windows, the strategy is expected to lose about 12 dollars after fees because its edge is too weak and inconsistent. Kristian Fagerlie publishes the code and data as an educational experiment and explicitly warns viewers not to stake money on the result.
Watch on YouTube
