Timing Small versus Large Stocks

摘要

U.S. equity managers who might consider an alpha-generating strategy using a small-size bet would earn, on average, a positive expected alpha in the long run. They could also experience long periods of underperformance. The classic small-minus-big strategy, which systematically favors small-caps, might well be too naive, and size timing, even if risky, can present an opportunity to add further value. We show that strategies based on three artificial intelligence approaches—recursive partitioning, neural networks, and genetic algorithms—could successfully time the U.S. size premium over the 1990-2004 period. Of the six individual timing strategies examined—three artificial intelligence approaches conditioned on historical data (1975-1989) and on recent data (1975-month preceding the prediction)—five outperform the SMB premium. None of the six timing strategies systematically outperforms the SMB strategy during the three five-year sub-periods examined. Yet a strategy based on the majority rule (that is, a strategy favored by at least two of the three artificial intelligence approaches) outperformed the SMB strategy in each subperiod. Not only does the consensus strategy benefit from stronger predictive signals, but it also allows the number of bets (transaction costs) to be reduced. Five of the six timing strategies, as well as the consensus strategies, remain profitable even after transaction costs. Although all methods have their merits, recursive partitioning could be favored for its much greater transparency and ease of interpretation. In the case of NN and GA, we deal mostly with black boxes. For investors favoring results over understanding, the black box syndrome is not a serious issue, but when the model does fail (as each method does in one subperiod), the investor will find it quite complex to see what went wrong, given the opaque nature of the model. In this presentation, we consider only extreme bets, 100% long in small-caps and 100% short in large-caps, and vice versa.This follows Fox [1999], who stresses that managers with superior forecasting skills (60% hit rate or higher) should favor more extreme tilts because they improve the entire range of possible returns. Still, we can reasonably conceive of less extreme strategies that would allow for neutral allocation when choices are less clear-cut. Therefore, considering three states of the world— small-cap tilt, large-cap tilt, and no tilt—could be more interesting, as it could yield stronger predictive signals and fewer switches (lower transaction costs).