I am training a stacked LSTM that takes as input a sequence [1,...,n] and outputs a sequence [1,...,m], for m<n, to predict stock prices.
Upon training the model with multiple architecture nuances, I am getting the same result: the yhat sequences all have the same pattern.
What might be causing this?
*All input columns are z-scored
This is probably because there is no benefit is trying more nuanced architectures anyway. Stock prices are usually incredibly noisy data, neural networks excel when you have data with a complex structure yet not noisy (e.g. images, text, speech). Stock prices are the opposite, simple structure a lot of noise.
I would advise trying to model your data with something more simple, such as ARIMA or exponential smoothing.
