Knowledge-based recommender system for stocks using clustering and nearest neighbors

Abstract

Recommendation systems and algorithms are part of many services we use today. Online marketplaces, social media sites, streaming services, and many others lean on the algorithms to provide content for a user that match one’s likings. A practical example of such system is Netflix which may recommend movies to a user based on one’s viewing history. “Since you watched X, you might also be interested in Y”. Even though these algorithms are used in multiple services, there are still applications where the power of recommendation systems hasn’t been fully utilized for a public consumer. One of these are publicly traded stocks. Investing into publicly listed stocks is a common way to generate wealth. There are thousands of companies listed in NYSE and NASDAQ stock markets in the USA only. For an investor this is a lot to choose from. Some may prefer growth stocks and others blue-chip stocks with high dividend yield. One can search higher risk-reward returns from stocks that are dropping heavily and other seek steady growth in their preferred stocks. This thesis aims to implement a knowledge-based recommendation system that considers not only stock’s financial data but also historical price development to give meaningful stock recommendations based on an input of a single stock in a casebased manner. The implementation considers two different approaches when combining these distinctly different data types. The experimental development relies on clustering techniques to categorize similar stocks into different recommendation lists and finally sorting the lists using nearest neighbors. The evaluation of the approaches is conducted using machine learning evaluation methods combined with evaluation metrics used in recommender systems. The final best performing implementation is built on top of K-means clustering technique and t-SNE dimensionality reduction method. Trendlines and financial data of the stocks are combined using separately computed distance matrices. Similarity between the trendlines is computed using customized cosine-distance function. Finally the thesis presents a Stock Recommender using Similarity-based Methods (StockRSM).