What I have done is write a small program which can take a set of allocations for portfolio components and then calculate an expected Sharpe ratio for the whole portfolio by using a Monte Carlo simulation run over hundreds of years using historical returns data. It then uses simulated annealing to do a constrained optimization over the solution space. This is only feasible because there are only about 20 available portfolio components. There are also no capital-gains taxes, transaction costs, shorting or borrowing, all of which keeps things reasonably simple.
Of course historical returns are a poor predictor of future performance. But historical correlation coefficients are a reasonably good predictor of future correlations, which is mainly what I am after with trying to find the best diversification.
Actually at the moment I have the majority in a money-market fund just because I'm trying to time the market and play a hunch. But that's usually a poor strategy and I wouldn't recommend it to anyone else.