Couple of month and experiments later I get back to co-integration in order to investigate this approach further.
I've been prototyping simple out-of-sample test of synthetic currency pair which consists of two "real" currency pairs. It seems that the most promising pairs are commodity and exotic/developing country pairs like e.g. USDCZK&USDHUF on the pics below. Those pairs have usually higher costs like spread and swap though.
Having said that, I have not managed to do proper/full back test with all the real costs but this particular one has hard coded cost of 10% the trade return which is not really accurate
perf | perf-cost | USDCZK | USDHUF |
Annualized Return | 0.090 | 0.081 | 0.019 | 0.044
Annualized Std Dev | 0.023 | 0.020 | 0.074 | 0.083
Annualized Sharpe (Rf=0%) | 3.989 | 3.989 | 0.263 | 0.531