Image Classification with two very similar and random-ish classes

Question

I am trying to build a neural network which takes a picture of a FOREX chart (currency exchange) and outputs an "up" or "down" prediction. I'm classifying up or down based on where the stock is after one hour, and currently I have 400 examples in each class. However, my training attempts are stuck at around 49-51% accuracy. I've come up with a few potential reasons:

1) A bad model. I don't think this is the case - I've tried transfer learning on Inception V3 and two of my own models. However, is there a different type of model that could somehow account for the randomness of markets?

2) Not enough data: I think this is the most likely explanation. From what I've read, you need thousands of images, not a few hundred, for good image classification.

3) The problem is impossible. I chose FOREX over stocks because FOREX isn't as heavily affected by headlines during the span of a few hours. Most FOREX trading in that time frame is done by technical analysis, which is doing things with charts - trend lines, moving averages etc.

I know that two is the most likely explanation, but before I go out and spend days collecting more data, I would like a bit of feedback to ensure that this is actually possible.

Thanks!

Answer 1

This is likely a combination of 1, 2, and 3. Without any data to back this claim up, 800 data points seems woefully small for the scale of problem you are trying to solve.

Assuming the problem is solvable, why not try to build a model directly onto the time series of the underlying data? What extra information are you getting from the image that you can't get from the time series? I would recommend reading up on some time series models.

Keep in mind that people using way more sophisticated methods are likely trying to build similar models, and likely are not succeeding.