Improving Moving Average Trading Rules with Boosting and Statistical Learning Methods, страница 9

Table II. Percentiles of the most profi table moving average

Best moving average^a	1993	1994	1995	1996	1997	1998	1999	2000	2001	2002
1993: [1, 140, 0]	100.00	47.22	72.73	26.17	76.57	84.71	49.26	24.31	27.07	96.55
1994: [10, 80, 10]	88.99	100.00	68.18	19.14	79.50	6.37	92.40	92.94	57.17	25.14
1995: [10, 50, 0]	57.27	96.49	100.00	79.69	50.21	99.68	21.81	53.33	100.00	44.53
1996: [1, 200, 6]	100.00	46.19	72.73	100.00	100.00	21.34	79.17	3.92	41.21	94.24
1997: [1, 180, 0]	100.00	37.53	72.73	48.83	100.00	15.29	23.77	11.96	12.53	69.29
1998: [5, 60, 10]	46.70	27.84	34.09	28.91	57.74	100.00	97.55	98.24	80.61	49.33
1999: [5, 80, 0]	42.29	27.42	34.09	60.55	25.94	11.78	100.00	58.24	28.89	80.61
2000: [10, 60, 8]	76.21	86.19	25.00	37.11	69.46	68.79	90.69	100.00	90.91	39.54
2001: [10, 50, 0]	57.27	96.49	100.00	79.69	50.21	99.68	21.81	53.33	100.00	44.53
2002: [10, 120, 7]	100.00	29.69	0.00	76.17	78.66	71.02	60.54	80.39	57.98	100.00

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

the best moving average rule of the present year could be a bad rule next year, being impossible to establish a priori which the good rules would be.

Nevertheless, a wide set of moving averages may have predictive information that could be obtained through a fi ltered Boosting algorithm, able to get predictive information, as much from good rules as from bad rules. In this sense the results shown in Figures 1 and 2 are not strange. This shows how the fi ltered Boosting model is more robust and profi table than any moving average rule for a long period of time.

In conclusion, the contingency and variability of the predictive power shown by the moving average rules, together with the robustness and high Sharpe ratio exhibited by the fi ltered Boosting model, give us suffi cient reason to advise technical analysts to use boosting instead of individual moving average rules.

In order to give our work a deeper economic motivation which could suggest their transcendence and applications, we have repeated our experiment in two special rising and falling subsamples: on the one hand, the subsample from 4 January 1993 to 1 September 2000, a period of high rises in the market; on the other hand, we have also considered the subsample from 2 September 2000 to 31 December 2002, a period of continuous market falls. Both periods are shown in Figure 3.

1 2 3 4 5 6 7 8 9 10 11 12 13 14

Скачать файл