Quote from hfoptimizer:
Your statements are highly questionable. I've also written and operated software that can generate 10 to 100 great looking strategies per hour per CPU node. Such claims by themselves are meaningless.
What is the mean cross correlation of PnL between these strategies on hourly, daily and monthly time scales? What is the correlation coefficient matrix rank? What is your group autocorrelation of PnL between your in-set and your blind out-of-set periods?
There are four possibilities here:
1) Your cross correlation on a relevant time scale is high, indicating that you are generating close variants of the same strategy. Keep in mind you can have a very low hourly or daily cross correlation, then see that correlation blow up on a monthly time scale. Most likely your strategies are highly related. They will profit together and fail together. In-set/out-set autocorrelation tests are unreliable.
2) Your cross correlations are truly low on an appropriate time scale (there is diversity) but your in-set -> blind out-set autocorrelations are low. That means you've curve fit or over optimized.
3) Size capacity is very low. Or risk or drawdowns are too high and therefore limit size or leverage.
4) Or you have done what is extremely difficult and are well on your way to making 100s of millions. You also won't need investors. If you truly can generate so many viable strategies with low cross correlation you'll have a very high Sharpe ratio and no drawdowns for the portfolio. Just increase your leverage and you're done.
EDIT: The cross correlation test is actually more sophisticated than just mean cross correlation. One must determine the rank and decompose the correlation coefficient matrix. Only then can you measure true diversity of a population of strategies.
