Genetic programs, data-snooping, and technical analysis, страница 10

The returns to such a portfolio are also reported on Table V along with the returns of three popular equity indices over a similar time period. The returns of the futures portfolio are higher than the returns on the Russell 2000 index, but are lower than the S&P 500 and the Dow Jones Industrials. Further, the standard deviation of the futures portfolio is much higher than any of the equity indices.

Table VI reports the correlations among the futures portfolio and the equity indices. As expected, the futures portfolio does exhibit a low correlation with equity returns, between 2% and 0.7%.

These results offer little support for the use of technical analysis in commodity markets. While previous research in commodity markets has indicated the ability to generate profits in excess of transactions costs, when ex ante optimal trading rules are chosen and used, they are not capable of consistently generating profits in excess of transactions costs.

A natural question is why this study finds no support for technical trading rules while others, such as Brock, Lakonishok, and LeBaron (1992) and Lukac et al. (1988b) have. The primary difference is that this study considered a broader universe of technical trading indicators from which to identify optimal rules. This precludes data snooping, but it also precludes direct comparison of results. It does highlight one aspect of research into technical analysis profitability that is frequently overlooked, the necessity of using a trading rule that is optimal (by some measure) during in-sample testing. As Sullivan et al. (1999) points out, “The effects of such data-snooping . . . can only be quantified provided that one considers the performance of the best trading rule in the context of the full universe of trading rules from which this rule conceivably is chosen” (p. 1649). Genetic programming, by design, always selects the optimal trading rule from the in-sample period, previous research has chosen rules based upon their actual use or perceived popularity.

SUMMARY AND CONCLUSION

Technical analysis has a long history in commodity markets, and remains very popular despite a lack of theoretical foundation. Because of this, a rich literature exists on whether technical analysis is actually profitable. Most studies in equities and foreign exchange have failed to find profitability for technical trading strategies, although the evidence for futures markets is more mixed.

As pointed out by Neely et al. (1997) and White (2000), previous studies suffer from data-snooping biases introduced when historically popular trading rules are applied to historical, though ‘out-of-sample’ data. One remedy for data-snooping biases is to produce ex ante optimal technical trading rules from primitive operators, as in Neely et al. (1997) and Allen and Karjalainen (1999). This study uses genetic programming, an evolutionary method for algorithmic design, to evolve technical trading rules for 24 futures markets. When evaluated using data not available to the optimization process, only 2 of the 24 markets could be traded at a statistically significant level of profit.

A portfolio of these commodity returns was constructed that produced returns that were inferior to equity index returns over the time period studied. While the correlation of the portfolio returns to the equity index returns is low, the futures portfolio has a standard deviation significantly higher than the equity indices.

The results of this study are primarily limited by the constraints of genetic programming itself. Evolutionary methods such as genetic programming are not guaranteed to find global, or even local optima. Further, as pointed out by Cooper and Gulen (2001), data-snooping biases can be induced through the process used to choose the data lengths, the operations available, or almost any other parameter of the optimization process.