What types of neural networks are most appropriate for forecasting returns? Can neural networks be the basis for a high-frequency trading strategy?

Types of neural networks include:

  • Radial Basis Function Networks
  • Multilayer Perceptron (standard architecture)
  • Convolutional Neural Networks
  • Recurrent Neural Networks
  • Q-learning Networks or Deep Reinforcement Learning

What about advanced architectures such as:


  • LeNet-5
  • AlexNet
  • Inception networks
  • VGGNet
  • ResNet


  • Bi-directional
  • Attention Models
  • 3
    $\begingroup$ NN and HFT together won't be a good idea. You always have an intensive learning phase with NNs, too slow for HFT - so I propose broadening the scope of your question to trading in general - this is the idea behind my answer below. $\endgroup$
    – vonjd
    Apr 12, 2011 at 11:59
  • 10
    $\begingroup$ I'm not sure it's appropriate to classify support vector machines as a type of neural network. $\endgroup$
    – Zach
    Apr 21, 2011 at 19:28
  • 3
    $\begingroup$ I agree with Zach, SVM are not Neural Networks. $\endgroup$
    – SRKX
    Nov 1, 2012 at 20:17
  • $\begingroup$ @phoenix1886: You can accept one of the answers if you are satisfied by it :-) $\endgroup$
    – vonjd
    Jan 28, 2013 at 15:59
  • 1
    $\begingroup$ @vonjd: But the learning can be carried out once (or perdiodically) and the network put to use in inter-training periods? $\endgroup$
    – rex
    Jan 10, 2014 at 11:10

4 Answers 4


I would say in the context of trading in general (for HFT see my comment above) further developments of recurrent neural networks (RNN), e.g. so called historical consistent neural networks (HCNN) together with forecasting ensembles, are state of the art.

I published an article on that which will be published this month by Springer Verlag (Zimmermann, Grothmann, Tietz, von Jouanne-Diedrich: Market Modeling, Forecasting and Risk Analysis with Historical Consistent Neural Networks)

Just to give you an idea about the new paradigm here is a short excerpt:

In this article, we present a new type of recurrent NN, called historical consistent neural network (HCNN). HCNNs allow the modeling of highly-interacting non-linear dynamical systems across multiple time scales. HCNNs do not draw any distinction between inputs and outputs, but model observables embedded in the dynamics of a large state space.


The RNN is used to model and forecast an open dynamic system using a non-linear regression approach. Many real-world technical and economic applications must however be seen in the context of large systems in which various (non-linear) dynamics interact with each other in time. Projected on a model, this means that we do not differentiate between inputs and outputs but speak about observables. Due to the partial observability of large systems, we need hidden states to be able to explain the dynamics of the observables. Observables and hidden variables should be treated by the model in the same manner. The term observables embraces the input and output variables (i. e. Yτ := (yτ , uτ )). If we are able to implement a model in which the dynamics of all of the observables can be described, we will be in a position to close the open system.

...and from the conclusion:

The joint modeling of hidden and observed variables in large recurrent neural networks provides new prospects for planning and risk management. The ensemble approach based on HCNN offers an alternative approach to forecasting of future probability distributions. HCNNs give a perfect description of the dynamic of the observables in the past. However, the partial observability of the world results in a non-unique reconstruction of the hidden variables and thus, different future scenarios. Since the genuine development of the dynamic is unknown and all paths have the same probability, the average of the ensemble may be regarded as the best forecast, whereas the bandwidth of the distribution describes the market risk. Today, we use HCNN forecasts to predict prices for energy and precious metals to optimize the timing of procurement decisions. Work currently in progress concerns the analysis of the properties of the ensemble and the implementation of these concepts in practical risk management and financial market applications.

Parts of the paper can now be viewed publicly: Here

  • 1
    $\begingroup$ "Today, we use HCNN forecasts to predict prices for energy and precious metals to optimize the timing of procurement decisions." Who's "we"? $\endgroup$ Apr 13, 2011 at 10:01
  • $\begingroup$ @RockScience: Siemens $\endgroup$
    – vonjd
    Apr 14, 2011 at 5:39
  • 2
    $\begingroup$ ok thnaks. would be interested to read more about your testing setup indeed. $\endgroup$ Apr 14, 2011 at 6:11
  • 2
    $\begingroup$ google search found nothing on historical consistent neural network, so very interesting to read something about it $\endgroup$ Apr 14, 2011 at 20:21
  • $\begingroup$ I added a link to the article: See my edits above (unfortunately only parts of the paper are publicly view-able). $\endgroup$
    – vonjd
    Jul 22, 2011 at 15:21

Depends on the data you are trying to model. If your data experience regime change then something with a sigmoid function (arctan, hTan, ...)

If your data is mostly linear but does have some deviation use a radial bias.

These are general guidelines for neural networks. The frequency of the data is not relevant to the above statements.

Remember that any set of basis functions can be made to fit any set of data. The idea is to use functions that reveal some under lying truth about the data.

  • 1
    $\begingroup$ I agree with this answer completely. $\endgroup$
    – Vass
    Apr 26, 2011 at 11:43

It depends on the data, horizon, inputs, etc. Wavelet transforms seems to be good for reducing time, and PCA seems to be good for reducing assets. There's been a lot of work done in this area, so e.g., look at Jurik Research WAV and DDR modules. Their results indicate that you don't know which bars (days for EOD) are the most informative and also which features are the most informative - so collapsing via wavelet offers an advantage. I looked at NN a lot and don't think knowing the past helps predict the future. The patent that Vantagepoint got approved made no sense at all -- but maybe their intermarket analyses helps?


from our experience, working on daily time frame doesn't need such non-linear approach (NN, RNN, SVM,...) Daily open and close are so informative prices/volume that simple VARMA (linear model) explain dynamic stocks prices with great accuracy...but wr have to recognize that puting foots in intraday is a completely different story.

we have to admit that using dynamic non linear approach has a real value compared to classic linear econometric/probabilistics models.

despite that, we continue to prevent a maximum as we can of such non-linear approach, because what is well understood must be ablle to be explained simply, and those kind of models exhibits too much 'hidden' features.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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