In general, in the markets, a mechanical trading system that has out-performed in the past, tends to under-perform in the future. If you're really lucky, then previous out-performance is just statistical noise, and your future performance is in line with the market.

In short, to your question "How can I calculate the change that over an X period of time, the strategy will, with a certain percentage of reliability) make money.", the answer is that you can't, and it probably won't: to calculate the probabilities, you need forward information about the markets, which you don't have; so you have to make a set of assumptions just to get the calculation to run: but at least one of those assumptions will be invalid, thus invalidating your calculation. If you want to minimize your chance of going broke, keep your maximum stake size at below 1% of your pot, for any individual trade: trading is a markov chain with an absorbing state at zero, and the trick is to avoid that state.

If you're in the top 1% of financial quants in the world, with access to the fastest sources of relevant information, can inject orders into the system as fast as any of the big players, and can access all their dark pools of liquidity, then your mechanical trading strategy may perform as well in the future as it did in the past, until others hit on how you're doing it, and copy it, diluting your returns. [**edit**:] In the OP's case, as a trader with all that access, then @Zach's suggestion of [the R package Performance Analytics][1] looks very promising. And I'd also suggest sub-sampling fixed-length intervals (e.g. 125-point) at random from the back-tested results, and looking at the statistics of the results: that is to say, pick a lot of dates at random from the range for which you have data: select uniformly on the range from the first date, to 124 days before the end date. For each interval, create a data point calculated as the net change in accumulated Profit & Loss between start and end dates. Then look at the mean, variance, skew, kurtosis of that set of data points. That might be useful: but only because the back-tested results were **not** used for calibrating the trading algorithm [**end edit**]


  [1]: http://cran.r-project.org/web/packages/PerformanceAnalytics/index.html