I currently test numerous Trend-Following/Momentum and Mean Reversion trading strategies for cryptographic assets with granularities ranging from daily to 3 hourly spaced data points. The general set-up is as follows: I choose an "end date" for the backtest. From here train and test sample dates are determined backwards depending on the specified parameters (I currently use a 75/25 train/test split ratio and a training sample window size of 120 periods). I have the following questions: 1) when I change my "end date" resulting in a shift of my training/test date samples some strategies tend to substantially change dynamics yielding a large deviation in trade statistics. Does this commonly occur in such tests? Are there common ways to mitigate this? 2) I chose the training window of 120 periods more or less arbitrarily. Is there a systematic way to choose for the "optimal" window size?