Do generative AI models, particularly large language models (LLMs), exhibit systematic behavioral biases in economic and financial decisions? If so, how can these biases be mitigated? Drawing on the cognitive psychology and experimental economics literatures, we conduct the most comprehensive set of experiments to date—originally designed to document human biases—on prominent LLM families across model versions and scales. We document systematic patterns in LLM behavior. In preference-based tasks, responses become more human-like as models become more advanced or larger, while in belief-based tasks, advanced large-scale models frequently generate rational responses. Prompting LLMs to make rational decisions reduces biases.

We propose a framework for rigorous inference in the evaluation of asset pricing anomalies that explicitly accounts for multiple methodological choices. We demonstrate that running multiple paths on the same dataset results in high correlation across outcomes, distorting inference. Alternatively, path-specific resampling reduces outcome correlations and tightens the confidence interval of the average return. Accounting for across- and within-path variability allows us to decompose the variance of the average return into a standard error, a nonstandard error, and a correlation term. We define the nonstandard Sharpe ratio as the ratio of the average return to the nonstandard error and show that this metric enables the identification of persistent anomalies. Empirically, we show that nonstandard errors dominate standard errors, and that 24\% of the anomalies in our sample (26 out of 107) are fully persistent.

Standard empirical asset-pricing models impose a common mapping from firm characteristics to expected returns. We develop the Lasso Fama–MacBeth Clustering Model, which jointly estimates characteristic-defined regions of the cross section and locally sparse average slopes within each region. In U.S. equities from 1980 to 2024, a pervasive core of short-term reversal and earnings surprise coexists with a broader periphery of characteristics retained only in particular regions. Cluster-based value-weighted portfolios deliver an out-of-sample Sharpe ratio of 2.04 and monthly alphas above 1.6% unspanned by standard factor models. They outperform pooled benchmarks and remain robust after transaction costs. The set of selected characteristics changes over the business cycle. State-specific fits for expansions and recessions produce distinct partitions and a broader set of selected characteristics in recessions.