I am training a CNN model for trading using indicator and MA lines to compose a 2D array as input. I want to normalize MA data(ema, sma...) into range between -1 and 1, I have tried several techniques such as min max, z score, percentage change between t and t-1. I have two problems:
The stock prices gradually increases in a very long term, if I use min_max or calculating change between t and t-1, most of variation will concentrate in the later part of time series(because value of change are small due to the price of stock low at beginning).
If I use percentage change, outliers so high/low that my standard deviation is very small and most of my data is close to zero. This problem also happens to my first method. I wonder if it is legit to manually remove outlier or other technique might be helpful.
My question is If I normalize my data as I mentioned, MA series data will be normalized with small standard deviation(0.01-0.1) while oscillators such as RSI are normalized with relatively large standard deviation(0.2-0.4), will the result being greatly affect from the imbalance of data distribution? If so any suggestion on how to solve my problems?