Improving Moving Average Trading Rules with Boosting and Statistical Learning Methods, страница 11

a

  Parameters of the moving average rule [n₁, n₂, b].

b

  Number of moving averages (minor : equal : major).

was 74.58% and the net return corresponding to the fi ltered Boosting model was 69.80%. Also observe that the strategies obtained from fi ltering the statistical learning methods in the rising subperiod become worse than the non-fi ltered strategies, with the exception of the boosting model.

In this rising subperiod the Sharpe ratio of the B&H strategy (0.0631) is only overcome by the best moving average, whose Sharpe ratio is (0.0657). Besides, the best ideal profi t ratio (0.1075) was obtained by the B&H strategy.

These results, which signal the supremacy of the B&H strategy over all the learning methods during a rising period of the market, are not strange.

Nevertheless, as we can see in Table IV, the behaviour of the technical trading rules based on the learning methods was the complete opposite during the falls subperiod. Thus all learning methods overcame the return of the B&H strategy, especially the fi ltered Boosting model. From 2 September 2000 to 31 December 2002, while the return of the B&H strategy was −35.92%, the net return of the fi ltered boosting model was −12.47%, which overcomes the other statistical learning methods (−21.85% for the fi ltered Committee and −14.59% for the fi ltered Bayesian model) as much as the best moving average model (−13.19%). On the other hand, the Sharpe ratio of the fi ltered Boosting model (−0.0224) is higher than that of B&H (−0.0485), and much higher than the Sharpe ratio of the fi ltered Committee (−0.0919) and fi ltered Bayesian model (−0.0618). The same happens with the ideal profi t ratio, which is −0.0249 for the fi ltered Boosting model, while it is −0.0718 for the B&H strategy.

Therefore, although the fi ltered Boosting model is not able to overcome the B&H strategy during rising periods (Table III), the results obtained (Tables I and IV) suggest that the fi ltered Boosting model is able to absorb a considerable part of the falls in the market.