Abstract
Traditional approaches to estimating beta in finance often involve rigidassumptions and fail to adequately capture beta dynamics, limiting theireffectiveness in use cases like hedging. To address these limitations, we havedeveloped a novel method using neural networks called NeuralBeta, which iscapable of handling both univariate and multivariate scenarios and tracking thedynamic behavior of beta. To address the issue of interpretability, weintroduce a new output layer inspired by regularized weighted linearregression, which provides transparency into the model's decision-makingprocess. We conducted extensive experiments on both synthetic and market data,demonstrating NeuralBeta's superior performance compared to benchmark methodsacross various scenarios, especially instances where beta is highlytime-varying, e.g., during regime shifts in the market. This model not onlyrepresents an advancement in the field of beta estimation, but also showspotential for applications in other financial contexts that assume linearrelationships.