I'm currently working through a time series problem where I'm trying to develop a model that learns the historical relationships between several related financial markets. So, using the closing prices of 8-10 different markets that I think have some sort of connection, predict the price of market #11 based on the other 10.
I figured the best place to learn would be by reading through a lot of research. Most of what I find usually entails the authors doing some kind of of signal noise reduction(wavelet transform or FTT) on the Open, High, Low, Close data. Next, adding some technical indicators to the dataset and applying feature selection (PCA or Stacked auto encoders) and finally feeding a deep learning structure or ensemble structre (ANN or LSTM). However the key part that seems to stay similar is they are usually always using a single market's prior prices(plus indicators based on those prior prices) to to predict future prices.
Ultimately, my belief is that very little success can come from that approach b/c there are too many factors involved in a single market's time series and they change too often. Expecting a model to consistently learn repeatable patterns that hold into the future would be asking to much. Even if it were doable at one point most arbitrage would've been found by now.
With that belief in mind, I'm wondering how different is what I'm trying to do versus to the research I'm reading? Am I tricking myself in thinking that the "learning the historical relationship between markets" is a different task than simply using the prior closing prices of one market plus some technical indicators to predict future closing prices of that market?
Additionally, would following their noise reduction/deep learning structure even make sense for my problem or would it be best to stick to the more traditional more linear type algorithms (Ridge, Lasso, SVR etc)
Any advice would be much appreciated!