As with all seasonality you have really just curve fitted your strategy to your data mined historical data.
Analyse a smaller section of data and use it to decide which months to be long and which to be short and then use this in sample to test an out of sample and I think you will find the results to be rather disappointing.
I coded something a long while back that used a rolling set of seasonality data to try to avoid this sort of curve fitting but the results of that were disappointing. I think for some commodities considering seasonality might give a slight edge but otherwise I personally find it to be of little use.
