Of course you can not improve the system by adding a uncorrelated losing strategy, but you can improve the profit by adding a losing strategy with negative correlation. Look at the following equity curves: A: ___/'''''''''''''\____/''''''''''' B: /''''''\_____/''''''''''\____ A is winning, B is slightly losing. Add both curves and I hope you'll see then what I mean.
By the way, there is a huge danger in what you propose. If your strategy A is curve-fit (which is very likely true, given your small training window), and if strategy B is significantly anti-correlated to A, that means it is likely that the two strategies are picking up on the same statistical anomalies. So you are training both strategies to pick up on the same irreproducible pattern. Therefore, if you try to add both of them together, they will both fail as soon as they see new data. In other words, you will get the false sense of security that you are hedged by using B, when in fact, you are experiencing the same risk as when you just ran strategy A.
So you're trading the spread between the strategies, which degenerates to a correlation trade, with all the usual swings and round abouts of a pairs trade (allbeit a more dubious pairs trade than an ordinary pairs trade). It might well be easier to sell one of the strategies to newbies for $10000. And, it probably doesn't matter which of the two you sell
As to why it's good to add uncorrelated strategies to a portfolio, imagine that you trade two uncorrelated strategies after each other. The return is then the sum of both returns, but the drawdown is only the maximum of both drawdowns. I also do not really see why a negative correlation is an indicator of overfitting. An example of a negative correlation is trading stocks from a fire insurance company and stocks from a flood insurance company - that's probably a good thing in a portfolio. But for the sake of theory, just assume that the strategies are not overfit and the equity curves are from real trading.
The probability of inventing two directional mechanical strategies that are profitable is low, and having them also negatively correlated is lower still. If you're able to come up with this stuff, then you should be able to answer your own question. I'd guess that neither strategy is actually working, and you're looking for a magic fix. 2c. Harsh but fair.
Rationalize: Not to be harsh, but why are you posting when you have no idea of the topic? If you can't invent two mechanical strategies that are profitable, that's quite sad, and you can certainly get help - but that does not belong to this thread. My question here was only about experiences with optimal f or other MM systems for uncorrelated mechanical strategies.
The way you have put your problem is unusual. The mechanical strategies I am aware of & have worked on don't have the on/off profitability you're implying. They're more driven by volatility, and operational blunders, as factors determining their day to day small sample efficacy. Your situation, as described, running a couple of strategies that sometimes are profitable but are always negatively correlated sounds pretty unusual. Not the sort of thing banks get excited about.