Dear all, I saw a lot of backtest of last 5 or even more years. And many of them are doing well years before but not so well this year.

In my test, I used 18 month to train (6 month for a WFT loop), and rencent 2 month as out of sample test. The result is better than train 5 years (1 year for a WFT loop), and it passed the reality check.
The problem is, the optimized parameter only work for latest months. And it will fail for data from 2 years ago.

So based on above test result, and since this year's data pattern changed a lot:
1. Is it a good idea to only use 1-2 years data to train and only latest 2 month to test?
2. Is it a good idea to retrain monthly to include latest data patterns?

This depends on the trading system, its characteristics, and the parameters being trained. Some systems hardly need any training at all, some need weekly training!

Thanks, it is really good to know!