Conditional Betas: a Non-Standard Approach

The exposure of stock returns to risk factors is known to vary over time. Traditional methods, such as rolling windows, have been widely used to capture this variation. However, these methods are highly dependent on the choice of window length and may fail to capture nonlinearities in the data adequately. We propose a novel approach that employs a neural network to estimate stock exposure to market risk on a time-varying basis. Our findings demonstrate that this neural network-based estimator outperforms regression-based estimators in out-of-sample predictive performance and exhibits no systematic bias across model-implied expected beta quintile portfolios. Additionally, the proposed estimator effectively classifies stocks into quintiles based on forecasted betas.