Saturday, October 5, 2013

lessons of ai: trading underperforms a trending market

its been about 6 months since i published aiSpy (queue cheesy background music,) an artificial intelligence strategy for trading the etf spy. i was a little disappointed that the real-world testing of aiSpy was underperforming the passive investing techniques buy 'n hold and dollar-cost-averaging. i was expecting that when i ran the numbers i would see a reversion to randomness in the real-world numbers. that is to say i thought the numbers would revert to something like a coin-toss with a win/loss ratio very close to 1 and a risk/reward ratio a smidgeon less. instead the numbers worked out to 8 wins and 4 losses for win/loss of 2. the average dollars lost on a losing trade was 66.42 and the average dollars won on a winning trade was 60.33 (both corrected for commissions.) this gives a risk/reward ratio of 1.10 (=66.42/60.33) 

this fits my criteria for a winning strategy whereby:

win/loss > risk/reward

in fact these numbers are rather impressive for automated strategies. for the ones i've evaluated it is rare that an actual edge of this sort holds up and rarer still for the edge to be this large. usually i see something like 1.8 for win/loss and 1.7 for risk/reward.

so why the underperformance? 

i think this has to do with the real-world era under comparison. the last 6 months has been an upward trending period for the market. thus, trading underperforms a trending market. perhaps, this is a bit obvious but the market has just come through a decade long era of sideways action. if you are using aiSpy for trading you should think about how long the market might continue trending. if you believe this trend is not done then buy 'n hold is a better strategy otherwise aiSpy should outperform in the next 6 months. i leave it up to you.

1 comment:

  1. This comment has been removed by a blog administrator.