Timing Small versus Large Stocks

Abstract

Historical evidence suggests that accurate timing can reap huge potential profits in trading between large and small stocks. Artificial intelligence approaches—recursive partitioning, neural networks, and genetic algorithms—are an aid in development of size-timing strategies to decide when to be long or short with respect to the U.S. Fama-French small-minus-big portfolio. AI strategies are trained on both a static learning sample and on an expandable learning sample. The results here show that most AI strategies could successfully time the U.S. size premium over 1990-2004. When predictions differ, a consensus strategy, or one favored by two of the three AI approaches, outperforms the SMB strategy more systematically, and reduces transaction costs.