A Comparison of GLOWER and Other Machine Learning Methods for Investment Decision Making

Abstract

Prediction in financial domains is notoriously difficult for a number of reasons. First, theories tend to be weak or non-existent, which makes problem formulation open ended by forcing us to consider a large number of independent variables and thereby increasing the dimensionality of the search space. Second, the weak relationships among variables tend to be nonlinear, and may hold only in limited areas of the search space. Third, in financial practice, where analysts conduct extensive manual analysis of historically well performing indicators, a key is to find the hidden interactions among variables that perform well in combination. Unfortunately, these are exactly the patterns that the greedy search biases incorporated by many standard rule learning algorithms will miss.

One of the basic choices faced by modelers is on the choice of search method to use. Some methods, notably, tree induction provide explicit models that are easy to understand. This is a big advantage of such methods over, say, neural nets or naïve Bayes. My experience in financial domains is that decision makers are more likely to invest capital using models that are easy to understand. More specifically, decision makers want to understand when to pay attention to specific market indicators, and in particular, in what ranges and under what conditions these indicators produce good risk- adjusted returns. Indeed, many professional traders have remarked that they are occasionally inclined to make predictions about market volatility and direction, but cannot specify these conditions precisely or with any degree of confidence. For this reason, rules generated by pattern discovery algorithms are particularly appealing in this respect because they can make explicit to the decision maker the particular interactions among the various market indicators that produce desirable results. They can offer the decision maker a “loose theory” about the problem that is easy to critique.

In this talk, I describe and evaluate several variations of a new genetic learning algorithm (GLOWER) on a variety of data sets. The design of GLOWER has been motivated by financial prediction problems, but incorporates successful ideas from tree induction and rule learning. I examine the performance of several GLOWER variants on a standard financial prediction problem (S&P500 stock returns), using the results to identify one of the better variants for further comparisons. I introduce a new (to KDD) financial prediction problem (predicting positive and negative earnings surprises), and experiment with GLOWER, contrasting it with tree- and rule-induction approaches as well as other approaches such as neural nets and naïve Bayes. The results are encouraging, showing that GLOWER has the ability to uncover effective patterns for difficult problems that have weak structure and significant nonlinearities. Finally, I shall discuss open issues such as the difficulties of dealing with non stationarity in financial markets.