[Paper Review] Predicting human decisions with behavioral theories and machine learning

Ori Plonsky, Reut Apel+10 more|arXiv (Cornell University)|Apr 15, 2019

Decision-Making and Behavioral Economics47 references59 citations

TL;DR

BEAST Gradient Boosting (BEAST-GB) hybrids behavioral theory with machine learning to predict risky choices, outperforming neural nets and many behavioral models and generalizing across contexts.

ABSTRACT

Predicting human decisions under risk and uncertainty remains a fundamental challenge across disciplines. Existing models often struggle even in highly stylized tasks like choice between lotteries. We introduce BEAST Gradient Boosting (BEAST-GB), a hybrid model integrating behavioral theory (BEAST) with machine learning. We first present CPC18, a competition for predicting risky choice, in which BEAST-GB won. Then, using two large datasets, we demonstrate BEAST-GB predicts more accurately than neural networks trained on extensive data and dozens of existing behavioral models. BEAST-GB also generalizes robustly across unseen experimental contexts, surpassing direct empirical generalization, and helps refine and improve the behavioral theory itself. Our analyses highlight the potential of anchoring predictions on behavioral theory even in data-rich settings and even when the theory alone falters. Our results underscore how integrating machine learning with theoretical frameworks, especially those-like BEAST-designed for prediction, can improve our ability to predict and understand human behavior.

Motivation & Objective

• Motivate accurate prediction of human decisions under risk and uncertainty across disciplines.
• Introduce BEAST Gradient Boosting (BEAST-GB) as a hybrid modeling approach.
• Show that anchoring predictions in behavioral theory improves performance in data-rich settings.
• Demonstrate generalization of BEAST-GB across unseen experimental contexts and its potential to refine theory.

Proposed method

• Introduce BEAST-GB, a hybrid model combining BEAST with gradient boosting.
• Describe CPC18, a competition for predicting risky choice, where BEAST-GB won.
• Compare BEAST-GB against neural networks trained on large datasets and multiple behavioral models.
• Evaluate generalization to unseen experimental contexts across two large datasets.
• Analyze how ML enhances or refines the underlying behavioral theory through data-driven adjustments.

Experimental results

Research questions

• RQ1 Can BEAST-GB outperform neural networks trained on extensive data in predicting risky choices?
• RQ2 Does grounding machine learning predictions in BEAST improve generalization to new experimental contexts?
• RQ3 How does BEAST-GB compare with existing behavioral models in predictive accuracy?
• RQ4 Can the integration of ML with BEAST refine and improve the behavioral theory itself?

Key findings

• BEAST-GB predicts more accurately than neural networks trained on extensive data.
• BEAST-GB outperforms dozens of existing behavioral models.
• BEAST-GB generalizes robustly across unseen experimental contexts.
• The approach helps refine and improve the underlying behavioral theory.
• Anchoring predictions on behavioral theory remains valuable in data-rich settings.