Predicting the memorability of magic tricks: a machine learning analysis of curiosity, surprise and recall

This project investigates whether measurable characteristics of magic tricks can help predict whether they will be remembered by an audience one week later. Two publicly available datasets were combined: MagicCATs (Ozono et al., 2020), a validated collection of 166 magic-trick videos rated on curiosity, surprise, interest, confidence, and other characteristics by 451 participants, and the Magic, Memory, and Curiosity (MMC) dataset (Meliss et al., 2024), in which 50 participants watched 36 of these tricks during an fMRI experiment and completed a surprise memory test one week later. Combining the datasets produced 1,800 participant-by-trick observations. Four machine-learning models (Logistic Regression, Decision Tree, Random Forest, and K-Nearest Neighbors) were trained to predict whether each trick would later be recalled. Participants were separated between the training and test sets so that no participant appeared in both. Random Forest performed best on the held-out test set, with an accuracy of 0.695 and an F1 score of 0.616. Video duration and independently measured interest ratings were among the strongest predictors. Two additional features inspired by a recent psychological taxonomy of magic (Kuhn, Thomas, & Griffiths, 2026) were also explored: whether a trick involved a mental or physical effect and whether it involved a volunteer. Across the complete 166-trick MagicCATs dataset, tricks involving a volunteer received significantly higher curiosity ratings than magician-only tricks (p = 0.034). However, adding this feature produced only a small change in memorability prediction within the smaller 36-trick MMC sample. Overall, the results suggest that measurable properties of magic tricks contain useful information about later memory, while also showing the importance of sample size, feature design, and careful validation when applying machine learning to behavioral data.