Files
Abstract
Using machine learning techniques to assist financial decision making surged in several areas in the past decade. The availability of high-frequency data enriches the forecasting models with features from market microstructure. Text mining introduces count, tonality and sentiments of financial buzz into machine learned equity price models. In this research, we first conduct a comprehensive survey of the most recent developments of financial text mining techniques. We organize and summarize financial text mining techniques in six aspects: news source selection, text preprocessing, document alignment and labeling, time series preprocessing, forecasting algorithm, and performance evaluation. We list available configuration choices in tables for each design aspect and highlight the performance comparison of different alternatives available in the literature. The survey is finished with a summary of most recent developments in this area and some suggestions on possible future research directions. In Chapter 3, we demonstrate a new stock forecasting perspective that directly learns the stocks relative performance with a ranking algorithm. We argue that the traditional regress-then-rank approach casts the portfolio selection practice into an unnecessarily hard problem and show that ranking algorithms outperform the neural network regressor significantly in terms of both ranking quality and simulated profit on out-of-sample testing data. More specifically, with testing data gathered from the ShenZhen stock exchange, LambdaMART scored an NDCG of 82.725 and 0.054% in return per position, while neural network return regressor can only get 10.998 in NDCG and its averaged return per position is -0.237%. With simulated trading under rigorous constraints of transaction costs and order execution price, we demonstrate that the ranker can be used to build highly profitable portfolios.