Enhanced Forecasting Methods, Fat Tails, and their Applications in Finance

This dissertation consists of two parts. The first part introduces a neural network approach that is used to forecast the conditional probability density function of asset returns. The model is unified with the idea of Arbitrage Pricing Theory (APT). In the second part, an algorithm called "individualized semi-linear regression" is discussed. The algorithm is an improvement of a linear regression, but slope and intercept may depend non-linearly on an arbitrary amount of exogenous variables.