Explaining the Unconditional Value Premium using Neural Networks

We investigate whether a conditional capital asset pricing model (CAPM) can explain the value premium of the unconditional CAPM. By modelling the exposure to the expected market risk premium (βt), the excess return (αt) and the error volatility (σt) as the output of a neural network we circumvent functional form misspecification from which previous studies suffer. We show that βt covaries negatively with business conditions, hence going into the right direction to explain the value premium. Nevertheless, the variation in βt is not large enough to fully account for the pricing error. Since this is possibly due to our post 1949 sample including insufficiently many deep crises, we compare our estimates with the predictions from Bai et al. (2019) and find support for their model in which value is riskier than growth, but a value premium is found in samples without large economic downturns.