2
$\begingroup$

Researchers at Stanford University released this paper in 2012:

http://cs229.stanford.edu/proj2012/BernalFokPidaparthi-FinancialMarketTimeSeriesPredictionwithRecurrentNeural.pdf

It goes on to discuss how they used echo state networks to predict things such as Google's stock prices. However to do this once trained, the network's inputs are a day's stock price, and the output is the day's predicted stock price. The way the paper is worded is like this could be used to predict future stock prices for example. However, to predict tomorrows stock price, you need to give the neural network tomorrows stock price...

All this paper seems to show is that the neural network is converging on a solution where it simply modifies its inputs a minimal amount, hence the output of the ESN is just a small alteration of its input.

Here are some Python implementations of the work shown in this paper:

In particular, I was playing with the latter which produces the following graph:

Figure 1

If I take the same trained network, and alter the 7th's day's "real" stock price to say something extreme like $0, this is what comes out:

Figure 2

As you can see, it basically regurgitates its inputs. So, what is the significance of this paper? It has no use in any financial predictions like the network shown in this paper:

$\endgroup$
  • $\begingroup$ Without looking too deeply: The first paper's network is naive about the system it is trying to model, while the second one has more assumptions about the system it is trying to model. $\endgroup$ – k.c. sayz 'k.c sayz' Apr 17 '18 at 22:49
0
$\begingroup$

After a very brief look at the paper, I think that they are predicting the stock price for the next day, and not for the current day, which is quite common and reasonable: see equation (1) where they predict x(t+1). So I don't see any issue with this paper.

But I've only quickly looked at it, so I may have missed something of course...

$\endgroup$
-1
$\begingroup$

Of course.If it is of any use, it won't be published.The fluctuation cannot be predicted by previous time series data. According to the well-known efficient market hypothesis,because lots of traders have access to advanced machine learning techniques such as RNN,lstm or attention based RNN, your model is unlikely to have any predicting power.

$\endgroup$
  • $\begingroup$ Whilst true, I don't think this answers the OP's question, which could apply to any sequence prediction. I think the OP is somehow using the software incorrectly $\endgroup$ – Neil Slater Aug 4 '18 at 17:46
  • $\begingroup$ I think the recent ZM trading fiasco is a strong example to refute the efficient market hypothesis. At least on a daily time scale. $\endgroup$ – Hanzy May 1 at 16:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.