If the model breaks down today, I can re-estimate it at the end of the day and be ready to trade again tomorrow. There are ways I can adapt to the conditions in real time but the majority of the time the conditions are changing very very slowly so there is a balance.. if the model adapts too quickly it is just curve fitting, if it adapts too slowly then it is not profitible. There are techniques to find structural break-points and such but right now I haven't found that it is necessary. During testing I tried several combinations.. e..g estimate the model with 1 day of data and trade it on 5 days, forward and backward in time, estimate with all 5 days and then trade on all 5 days, etc. You might think 5 days is a very small amount of time but I'm working on a small time scale. I am careful to check the results relative to the return of the entire market that day.
Perhaps you might have had a very consistent performance from (optimised) backtesting, but it looks like you don't have very consistent stataments.
When you feel you need to test and/or paper-trade a much longer period of time, perhaps you should actually do so.
I don't feel I need to or I would have done so. Since the model is recalculated every day (and possibly realtime) then the distant past has no effect on today.
I don't trade trends.. more of a contrarian strategy. The "bigger trend" equivalent in my model is the slow drift in re-esimated parameters each day and I don't want to trade that because it happens on the timescale of weeks or months.
Doesn't a contrarian strategy have anything to to with the bigger picture? It sounds like you have a kewl system though