Examining Long-Term Trends in Company Fundamentals Data

The equities market is generally considered to be efficient, but there are a few indicators that are known to have some predictive power over future price changes. This suggests that the market has some room for identifying inefficiencies. Much work is done on applying machine learning to short-term trading, but there exists little research on using machine learning to identify long-term inefficiencies; almost all mutual funds and hedge funds rely on the judgment of humans to make long-term bets about the market. Therefore, we have reason to believe a priori that there exist long-term market inefficiencies which can be found with machine learning. To test this, I collect a set of fundamentals data across several thousand companies over a fiftyyear period, taken from the CRSP/Compustat Fundamentals Annual database. First I apply linear regression over these fundamentals to see if they predict future returns; I find that they do predict returns better than any common indicators, but do not predict returns well on a risk-adjusted basis. Then I use support vector machines to classify stocks as high or low predicted returns. I find that for well-tuned parameters, an SVM can produce a classification where positive examples have high riskadjusted returns and where this result generalizes well to a large test sample.