The final stage of my optimizing project with training on FX30 and using instruments with long history is done for this time. Due to the long history I could work with a correct conducted Out-of-Sample testing, the cut off was between 1998 and 1999.
Notably with the test periods was that any period close to the cut off had better results and the target calculation preferred the shortest period of time available. Regarding the features it was clear that no one of the top 20 was using any kind of date related information. The features was mostly based on short periods of time and the two most popular calculations come from previous highest highs or the difference between two regression lines from the time series of Closes.
The best setup from each of the 16 three years periods will make up the base for each sub classifier for the IQ19o_FX30 automated trading system.
Hire a freelance quant, contact me at: mikael[at]furesjo.se