Bershadskyy, D., Dinges, L., Fiedler, M. A., Al-Hamadi, A., Ostermaier, N., & Weimann, J. (2024). Experimental economics for machine learning—a methodological contribution on lie detection. PloS one, 19(12), e0314806.
And work in progress (contact me for working paper currently being rotated)
In this paper, we explore the intersection of experimental economics and machine learning (ML), demonstrating how ML can enhance experimental methodologies while simultaneously benefiting from the rigor of controlled economic experiments. We focus on lie detection as a case study, addressing the shortcomings of existing datasets often used to train ML algorithms. These datasets typically lack objectivity, proper incentivization, and quality control. To address this, we modified the well-known "lies in disguise" experiment, creating a high-quality dataset where participants, incentivized to lie or tell the truth, were recorded on video. Using this dataset, we trained an ML-based lie detector that achieved a 67% accuracy rate.
Our work illustrates how experimental economics can contribute to ML by generating datasets with controlled variables, high reproducibility, and clear ethical guidelines. Simultaneously, we demonstrate how ML algorithms can enrich experimental research by enabling deeper insights into individual behaviors, such as lying. We emphasize the importance of creating diverse and replicable datasets to improve algorithm performance and address broader societal challenges.
We advocate for closer collaboration between experimental economists and ML researchers to further enhance algorithm quality and expand the scope of economic research.