15.097 Student Project: Security Analysis using Machine Learning

Four times per year, companies release earnings reports. Often the day following such reports, the stocks of these companies make significant moves either up or down from the resulting information. Such moves are often attributed to a large surprise between the analyst estimates for the stock and the actual earnings amounts. Unfortunately for the average trader, these jumps usually occur in a very short window of time following the earnings release, making it difficult to profit on either bullish or bearish earnings. Since earnings reports are not usually released during market hours, a method for determining which stocks will continue to move either up or down throughout the day would allow a trader to create a strategy that takes actual earnings information into account without being limited by this extremely short window of opportunity. Further, with the hundreds of earnings reports that are released each day, this type of method could parse and interpret an enormous amount of information—far beyond the capability of a single person. This study endeavors to create such a method that could utilize publicly available stock data in order to make trading recommendations for the entire day following the release of earnings reports. The traditional method that is used to measure the strength of earnings reports is earnings per share (EPS)—or the ratio of a company's profit to the number of outstanding shares of its common stock [5]. Outstanding shares of common stock simply represent the number of shares of a company that are held publicly [5]. Many fundamental financial analysts often give their own predications for what a given company's EPS will be in any given quarter, and as such, data concerning EPS estimates and actual realizations are much less sparse than other measurements of earnings strength (one such possibility could be company revenue). In addition, EPS has an intuitive interpretation which makes its use easy—EPS simply represents the average dollar 2 amount of total gain or loss of publicly owned stock during a specified time period. Thus, this project will attempt to learn patterns of market behavior following EPS news reports by examining behavior of past data. Further, this newly learned information will be used to create a machine-learning based algorithm for trading following earnings reports. In turn, this discovery will provide a foundation for algorithms that can process the information of hundreds of earnings reports each day in a meaningful manner. The first …