I found that adding a new strategy tends to increase the overall Sharpe ratio even though it's own Sharpe ratio alone is not so great. So far I have added about two strategies per year but I am worried about trying to find edges repeatly in the same instrument space might increase the risk of data mining/over-fitting.
At the same time, if their drawdowns overlap it can be catastrophic. A higher Sharpe does not exclude that possibility. But indeed, ideally one would look at a lot of different instruments and markets. I have an issue with trying to find edges in the same market with a similar trading style and with same risk parameters that would minimize drawdowns and work with the exposure levels, trouble is that it's extremely hard to meet all the conditions.
Hi folks, this is my point of view about this interesting topic: Even if I trade only forex spot contracts, about the golden rule of assets diversification I'm in strictly according with you. For this reason my several strategies, each one edging on different price movements (trend following, mean reversion and volatility breakouts) are applied to many decorrelated cross pairs. I don't need to generate periodically new strategies, but only to refine the weights of existing in according to changing market behaviour and microstructure in prices formation. I usually optimize systems weights of my real market portfolio with a proprietary algo that measure each system pearson correlation vs others and his last drawdown level. This algo follows a simple rule based on maximizing the lotsize of systems that report a low correlation levels and low drawdowns, and minimizing those systems that report a high correlation levels and high drawdowns. Obviusly keeping always the same max portfolio leverage in strictly according to my risk appetite. So each week I change systems order sizes, calculated by the algo on a fixed rolling window (usually on last two months of trading) as in sample period. I'm not a fan of periodical optimization methods as WFO, so I prefear build robust strategies backtested in long historical runs that performs well in any market. It means that my strategies are not very profitable in their ideal market conditions but keep always tiny drawdons in their adverse market conditions. The good performance of entire portfolio should be derived from different and decorrelated strategies interaction. Indeed changing systems weights in according to their last performances and their behaviour inside the entire portfolio I think is itself another kind of optimization. Excuse for my bad english, but I'm still learning..
That's pretty interesting. I am curious as to whether there are certain patterns in the change-in-correlation that could lead that system to chase itself somewhat?
Yes, I learned the lesson about overlapping drawdowns sometime ago. Now I calculate the exposure assuming the perfect correlations among different strategies/instruments and then limit that exposure.
Number of times. Forex is such a market where one particular strategy would not work for very long period of time. I like to observe market closely and then i take my decision. I do like to take calculated risk and that is why i have been in this industry for long period of times.
Searching recurring patterns that anticipate the future behavior of a system, by studiing the changes in its levels of correlation inside the entire portfolio and its drawdwon levels compared to the performances of other systems, is a very complicated task that could possibly be resolved through sophisticated machine learning techniques. By my small experience I can say that, if these patterns exist, they are very difficult to find. In fact is not always the case that when a trend following system outperforms on a certain symbol, then a mean reversion system must outperform on a symbol uncorrelated to the first, or even that a trend following system applied on a highly correlated symbol to the first must outperforms too. I think this difficult exists because in a portfolio composed by several strategies applied to different symbols, the changes measured on correlation level of their returns and the changes obtained in their net performances usually not have the same speed. In practice the levels of correlation of strategies returns traded on different symbols do not explain the changes occurred in the price bias of the symbols traded, and so on cannot explain their future performances. What I do is to adapt the systems market exposition measuring both their last performance (its drawdown level) and their behaviour inside the entire portfolio (its mean correlation level vs all others used systems), minimizing exposure or stopping the unsafe systems and maximizing overperforming and uncorrelated systems. Maybe this method suffer a certain delay, by since now it has done its dirty work with good returns and lower drawdowns. My 2 cents..