Predicting IPO Performance from Nearest Neighbors Using TF-IDF Weighted Word Count Vectors

We introduce a novel approach to mining and leveraging data concerning stocks in order to predict the performance of new stocks following their initial public offering, a traditionally difficult task due to the lack of information and historical performance data. We collect a large corpus of articles for every existing stock between March 1st, 2014 and March 1st, 2015. We create weighted feature vectors by calculating the TF-IDF values for every word that appears in a document about a given stock. We then perform locality-sensitive hashing on these feature vectors and bucket similar stocks into a neighborhood. Using a variety of monotonically decreasing functions, we examine the relationship between 30-day price fluctuations of each IPOs neighbor by propagating each neighbor’s 30-day change to the IPO relative to its distance. Locality-sensitive hashing uncovers interesting relationships between IPOs and nearby neighbors. We observe a strong correlation between neighbor stock distance and respective 30-day performance.