Value investing with rule-based stock selection and data mining

Paradigm shift is on the way in the financial market and economics theory. Evidence against the previously prevailing assumptions of rationality and market efficiency has become abundant and new models, that are rapidly becoming the main stream, are based on the actual investor behavior that can be empirically observed and provide better fit to the data. The contrast between the old and the new schools of thought serves as a background for this study, and reviewing some contemporary theories and studies of topics such as enhanced value investing and contextual fundamental analysis justifies its results which under the efficient market hypothesis would be anomalous. The first objective of this study is to test in the Nordic markets (Finland, Sweden, Norway and Denmark) Joel Greenblatt's investment formula (GF) that he published in his 2005 book: "The Little Book that Beats the Market". His method aims to buy good stocks when they are cheap and has provided during period 1988-2004 an average 30.8% p.a. return in the U.S. stock market while the S&P 500 index yielded 12.4% annually. Although the average return performance is stellar, the selection rule also accepts some stocks that will be deeply in loss thus making the whole portfolio to suffer periods of underperformance. This undermines the formula's exploitability especially in professional fund management setting. Second objective in this study is to improve GF by developing a model that can filter out the loss producing stocks ex ante using information set that is available at the time of investment. To accomplish the first goal a program that simulates GF investment rule was developed and run with the result of 29.4% p.a net investment return after taxes and trading costs during the research period 2000-2011 while the relevant reference stock index (FTSE Nordic Value Index) returned 7.6% p.a. To develop a model for the second stage, data mining methods were applied in variable and model selection phase. The produced logistic regression model with constant and 8 additional terms (single variables and interactions) was able to predict with 80-90% accuracy the predefined holding period return category for the cases in the research data. Sample data was divided into training, testing and validation partitions to ensure the out-of-sample performance of the models. When this model was applied as a filter in the stock selection phase the annual return for the period increased to 43.8 % p.a