If an a posteriori probability distribution is a good fit for historical events, it doesn't mean in any way it is going to fit future data points. It may or may not.
Hence, use backtesting with care while trading.
Given the vast amount of parameters that constitute a strategy, naive optimizing of the outcome of the strategy in the past is bound to produce overfitting in the majority of cases...
Simply for the next t-time periods let the model give the user an estimate of the future prices/rates, together with an estimate of how accurate the model expects these predictions to be. This would allow the user to build confidence in the algorithm strategy she/he came up with, before employing it on the open market.
I put that quote on the front page of my masters thesis.