Artificial Intuition for Forecasting the Future: Can a Model Trained on Decision Reliability Predict Real Financial Events?

This study investigates whether a machine learning model trained on patterns of "decision reliability" — measurable regularities in a participant's binary response sequences — can forecast real future financial events above chance. A single participant completes 50 sessions of a web-based binary decision test (https://prova.riff.works/cognito-predict). In each session, the participant responds True or False to approximately 400 statements about stock price comparisons. Critically, the statements are concealed behind opaque panels: the participant cannot read them and responds based solely on intuition. Stimuli are a pseudo-randomized mix of known-outcome trials (verifiable past financial facts) and unknown-outcome trials (referring to one of 50 pre-registered future financial events involving major US technology stocks, spanning February–April 2026). For each session, a Gradient Boosting model is trained exclusively on the known-outcome trials to learn whether the participant's response on each trial is correct or incorrect, based on sequential response patterns and reaction time features. The model is then applied to the unknown-outcome trials: responses estimated as correct are kept, responses estimated as incorrect are inverted. The corrected responses are aggregated into a vote count that produces a single probabilistic forecast for the target event. Each prediction is pre-published on Figshare before the outcome becomes observable, ensuring full temporal integrity. The primary outcome is binary forecasting accuracy across the 50 events, tested against chance (50%) with a one-sided binomial test. The secondary outcome is the mean Brier score, tested against the non-informative baseline (0.25) with a Wilcoxon signed-rank test. The expected outcome under the null hypothesis is 50% accuracy and a Brier score of 0.25. If the model captures transferable predictive information from decision reliability patterns, accuracy should exceed chance and the Brier score should fall below 0.25. All results — including negative findings — will be reported in full.