What question did this study set out to answer?

The aim was to test the artificial intuition hypothesis in financial forecasting using a pre-registered protocol.

June 10, 2026Open Access

Artificial Intuition for Forecasting the Future: Confirmatory and exploratory analyses of 50 pre-registered financial forecasts

Key Points

The aim was to test the artificial intuition hypothesis in financial forecasting using a pre-registered protocol.
Single-subject design with N=1 participant for 50 sessions of 400 intuition trials each.
Binary forecasts generated for 50 future closing prices of major US technology stocks.
Utilized both pre-registered endpoints (directional hit rate, Brier score) and post-hoc exploratory analyses.
Directional hit rate was 60.0% (30/50), not meeting the success threshold (k ≥ 32, p = 0.101).
Bayes factors indicated insufficient evidence for the artificial intuition hypothesis (range 0.47-0.86).
Post-hoc analyses identified phrase coherence as a promising moderator for future testing.

Abstract

Direct prediction of future financial events is notoriously difficult and rarely outperforms chance. The present study pre-registered a single-subject protocol designed to detect implicit forecasting signals through an opaque-panel intuition task combined with a supervised model trained on per-trial reliability. A single adult male participant (N = 1) completed 50 pre-registered sessions of 400 forced-choice intuition trials each, generating binary forecasts on 50 future closing-price events for eight major US technology stocks (March-April 2026). As specified in the pre-registration, each forecast was time-stamped on Figshare before the earliest date contained in the event statement. The pre-registered primary endpoint failed: directional hit rate was 30/50 = 60.0% (one-sided binomial p = .101), below the pre-registered success threshold of k ≥ 32. Bayes factors under multiple default priors were in the range 0.47-0.86 ("not worth more than a bare mention"; Kass all of these session-level variables are formally defined and characterised in the subsequent sections of the paper. These exploratory observations are reported as hypothesis-generating for a future pre-registered replication and not as confirmed effects. The closed pre-registered dataset therefore does not provide evidence in favour of the “artificial-intuition” hypothesis, but the exploratory profile produces a falsifiable working hypothesis (phrase coherence as primary inclusion criterion) to be tested out-of-sample in a subsequent round. The principal exploratory findings are described in §3.5.1 (phrase coherence between the target and opposite-control phrasings as the strongest single moderator of HIT), §3.5.6-3.5.7 (the hierarchical decision rule that combines phrase coherence with within-session model fit and verifiable trial accuracy), §3.5.2 (the inverse relationship between verifiable-trial accuracy and forecast HIT), and §3.6.4 (the dynamic-rules interpretation of why the per-session model fits locally but does not transfer across sessions).

Perguntar à IA

Bookmark

View Full Paper