The first two chapters develop ML-based models that generate conditional SDF estimates efficiently using conditioning information sets, resulting in minimal pricing errors and superior Sharpe ratios compared to traditional conditional asset pricing models. Key aspects of our approach include enabling the sharing of information between assets, applying NN-based spatial recognition techniques to IV surfaces, using a custom non-MSE loss function that enables the model to estimate conditional SDF directly, bypassing the next-period return prediction step, and employing the new weight updating scheme that controls the memory time of the neural network after the initial estimation on the burn-in sample.

The third chapter builds an asset pricing theory around ML-based asset pricing models, associating them with traditional linear factor models from the asset pricing literature. We show that sufficiently wide neural networks trained to maximize the SDF Sharpe ratios are equivalent to a large factor model â a linear factor pricing model that uses many nonlinear characteristics. We demonstrate how neural networks identify these characteristics through gradient descent. Our analytical solution also allows us to filter out noise from the neural network output.