Since I dabble in No-Limit Texas Hold'em (henceforth NLHE) online, I have invested a bit of time learning the game properly. (IA pretty good source of poker information: Two Plus Two) NLHE is a game of small edges and significant uncertainties, where an average micro/low-stakes winning player makes a mean of 4+ Big Blinds/100 hands with significantly greater standard deviations. 100,000+ long hand histories are the norm where these pro players feel comfortable with the certainty of having an edge. It has opened my eyes around true sample sizing.
Financial Back-Testing, sample size
I've seen plenty of curve fit back-tests going back decades at a time. And there was a time where if the historical test performance displayed great statistics (sortino ratio, high win rate, etc.); with as little as 500 historical trades, I would go "OK great, let's trade it!"
That did not "work".
Anyone can become a master of curve fitting. Well, besides problems off curve fitting, statistical significance simply does not exist for quantitative financial trading. To gain even an illusion of statistical significance, a minimum of 50,000+ trades are needed due to the fact that no significant edges are available today with the presence of growing institutional machines (Impact of institutional trading on stock prices).
Solutions
Now that we know solutions aren't likely to come from more data-mining techniques, where can we look instead? What's worked for me is in learning micro/fundamental structures of traded instruments, and discovering inefficiencies thereof. It's like any other business, where R&D and innovation means everything. Sure it's hard, but not impossible.
Financial Back-Testing, sample size
I've seen plenty of curve fit back-tests going back decades at a time. And there was a time where if the historical test performance displayed great statistics (sortino ratio, high win rate, etc.); with as little as 500 historical trades, I would go "OK great, let's trade it!"
That did not "work".
Anyone can become a master of curve fitting. Well, besides problems off curve fitting, statistical significance simply does not exist for quantitative financial trading. To gain even an illusion of statistical significance, a minimum of 50,000+ trades are needed due to the fact that no significant edges are available today with the presence of growing institutional machines (Impact of institutional trading on stock prices).
Solutions
Now that we know solutions aren't likely to come from more data-mining techniques, where can we look instead? What's worked for me is in learning micro/fundamental structures of traded instruments, and discovering inefficiencies thereof. It's like any other business, where R&D and innovation means everything. Sure it's hard, but not impossible.
0 Reflections:
Post a Comment