Big Data in Financial Economics

Abstract

This dissertation focuses on understanding several finance and economics problems using big data. We first process and analyze financial textual data to investigate how sentiments can be learned directly from news data. We present a novel framework for textual analysis based on the factor model and sparsity regularization, called FarmPredict, to let machines learn financial returns from news data automatically without the help of prior knowledge like sentiment dictionaries. We validate our method using the articles in the Chinese stock market and analyze the magnitude, sources, and durations of effects caused by the positive or negative sentiments scored. Then we focus on high-frequency financial data and study the question of how and when returns are predictable at a high-frequency level. We quantify and document the predictability and its universality in very short horizons of a few seconds or transactions. We discover the inherent sources of this predictability and extrinsic cross-sectional and time-series factors and market environments affecting the predictability. We also examine how the predictability is affected by the timeliness of data. Finally, we investigate the problem of measuring housing activity in real estate economics. Timely measuring housing activeness at high-resolutions is critical for policy-making and urban design but hard, if not impossible. By using energy consumption data to infer housing activeness and modeling several disjoint datasets including nightlight satellite images and land-use data, we build an approach that can explain the majority of the spatial and temporal variation in monthly housing usages.