About the shape of input feature

In this framework, at each timestep, the agent takes the feature of the whole universe as input (for example Dow 30). If I use 10 features per stock, then the input feature shape will be 300 and it is rational to take it as input. However, when the universe is bigger, such as about 200 crypto futures in Binance, the input dimension will be 2000 which is a quite unusual case.

How do you think about this kind of situations?

I tried myself to let agent takes features for one crpto/stock each and using 'for loop' to iterate over all stocks/cryptos at each timestep. Then concat the output and use it as the mean of diagonal Gaussian for sampling the action (score for each stock which will be normalized as portfolio weight). Though, It doesn't work well now.