Supervised classification-based stock prediction and portfolio optimization

As the number of publicly traded companies as well as the amount of their financial data grows rapidly and improvements in hardware infrastructure and information processing technologies enable high-speed processing of large amounts of data, it is highly desired to have tracking, analysis, and eventually stock selections automated. Machine learning has already attained an important place in trading and finance. One currently major area is high-frequency trading. There are many techniques in the literature and applications to predict shortterm movements based on different stochastic models of temporal variations of stock prices [1], [2], [3] . Such approaches generally rely on treating individual stock data as a time series without analyzing correlations and patterns between different companies, mainly because of the limitations of processing very large data sets at very high-speeds. Another major application is valuation of the market based on economic parameters [4], [5], [6], [7], [8], [9]. For most of these applications the amount of data processed is limited and most do not drill down to the granularity of individual companies [4], [5], [6], [7], [8], [9]. Few works which studied portfolio optimization on individual company level focused on very small number of financial parameters [10]. A very large portion of the finance industry is comprised of midto long-term portfolio construction and is still mostly performed by hedge fund managers and financial analysts based on analysis and decision processes using company fundamentals. Considering the entire New York Stock Exchange (NYSE) stock market between the years 1993 and 2013 there are more than 27189 stocks (and that number grows almost everyday with new initial public offerings). Since this number is far larger than a human could manually handle, automation of financial analysis and investment decisions by modeling company fundamentals is a clear need. In this project, we use machine learning techniques to address portfolio optimization. Our approaches are based on the supervision of prediction parameters using company fundamentals, time-series properties, and correlation information between different stocks. Rather