Walk forward optimization
4/7/99
Today I tested how I would have done with Microsoft using a 3 year lookback window, optimizing for that 3 years, then marching forward one year. The results follow:
Asset: MSFT
Time: 12/31/87 – 4/1/99
System: weekly channel breakout system, long only.
Results:
Optimized
through: Hi_len/Low_len t+1: Results: n (wins):
89 1 / 2 90 31.6% 3 (2)
90 2 / 6 91 47.5% 3 (2)
91 2 / 2 92 -20%
92 1 / 6 93 - 8%
93 1 / 6 94 27% 4 (3)
94 1 / 6 95 26.7%
95 1 / 2 96 43.1% 8 (5)
96 1 / 2 97 25.2%
97 4 / 3 98 84.9% 4 (3)
98 2 / 2 99* (through 4/1) 17.4% (through 4/1)
Avg annual:
Total return: 794% 25.1%
buy and hold: 7,339% 59.1%
The results of this are both encouraging and discouraging. First, the total return was rather poor. Second, the optimal parameters migrated quite a bit, making it unclear if they were truly stable.
However, the return from year to year was positive with 2 exceptions. The past does have some predictive power. Second, if leverage had been used, or if interest from uninvested cash had been taken into account, the model return would have been higher. Of course, one cannot always count on an asset such as MSFT to deliver such a staggering buy and hold result, so the model may add value for those assets that do not perform as well.
No comments:
Post a Comment