Another paper which says it will beat the returns of many active managers: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0212320 I didn't really find anything super novel in this paper except I've tried to do exactly the same thing and failed. Obviously, I am stupid but these guys included the parameters of their layers so I will try and reproduce their results. I don't know why they don't just include Jupyter notebooks. Ah well. May I just say I absolutely love this website. All the important metadata about the paper is there with links. Much nicer than arXiv
Oh shit I figured out what they did different, they used closing price + image. I used OHLC + image. That makes a lot of sense.
This: "Forecasting stock prices is an attractive pursuit for investors and researchers who want to beat the stock market. However, forecasting stock prices is difficult."
I don't want to forecast, just want to pattern match a OHLC time series. I am not sure if I would ever make use of it, but why not try
I too have attempted using a ConvNet-LSTM stack (didnt care for it...LSTM is to ad-hoc for my tastes), but I didn't run images through it... Not sure why you would when you have a more compact representation in the raw-data data that built the chart in the first place.
Also, they say that you experience degradation when your stack is too deep and had a way around it by short-circuiting . I never tried that, and I don't exactly understand it either so that's another rabbit hole.
I agree with you. But I believe there are better solutions for time series modeling than LSTM that are grounded in system-id and Bayesian estimation (at least for my purposes). HMMs are my go to. Also, the stated advantage of LSTM with regard to vanishing gradients is solved in conventional ANNs using a leaky relu
I can definitely agree that LSTM isn't the best solution for our problem. What do you mean by system-id and HMM?