algoTraderJo: Many thanks for all your comments. This has helped me clarify several issues I've been working with over the past few years.
I've used Markowitz portfolio optimization in the past. My results may have suffered because 1) I wasn't using enough 'data', and 2) I didn't take into account the need to periodically reoptimize the weights. You bring up another interesting point about data and in particular, having enough data:
When I started out using Markowitz I had the impression that 'data' meant bars. However I'm now of the opinion that it means examples or trades because it is the consistency of a strategy that is important, and when you combine equity curves you are making a bet that the relative relationship between the curves will stay fairly static. However it could also mean time. When you think of 'lots of data', what measure do you use?
I was aware that more frequently trading systems tended to dominate less frequent systems in the combined portfolio but was unaware that equal weighting may have implicitly exacerbated this tendency, or that Markowitz helped to compensate for this. I don't see trade count in the formula, is this compensation done through the covariance matrix or some other means?
Disliked{quote} When weighting systems I usually combine them in order to minimize variance (what is known as a Markowitz optimization)Ignored
QuoteDislikedsince my tests generally use 25 years of data I do not see a substantial deterioration of the performance when I move to live trading (it could happen but I have not seen it yet). Nonetheless I keep rebalancing the portfolio to the weights that reduce variance the most every month using the updated backtests. Since the tests are so long the changes are very gradual if they happen at all.
When I started out using Markowitz I had the impression that 'data' meant bars. However I'm now of the opinion that it means examples or trades because it is the consistency of a strategy that is important, and when you combine equity curves you are making a bet that the relative relationship between the curves will stay fairly static. However it could also mean time. When you think of 'lots of data', what measure do you use?
QuoteDislikedOne thing the Markowitz optimization process does that makes things much better than an equal weights distribution is that it compensate naturally for trade frequency (something that is likely to remain similar). If your weights are equal you are probably under-weighting and over-weighting systems depending on how frequently they trade. If you have a system that trades 10 times per week and another that trades 2 times and you weight them equally the 10 times per week system is actually getting a much bigger cut, simply because it trades more (from...