Common modelling do's and don'ts:

- all parameters are not equally important through time.

- Missing variables are not random events.

- Time series models should be guided by event driven simulations.

- Increased proximity to a forecast should increase in accuracy, and shift relevance to other parameters.

I.E.: If you cannot model a fox chasing a rabbit, you can improve. (A trader chasing an profit ).

I would be pleased to show some neural net , or event driven or modelling examples. Cheers, Stef

Thanks Stephane (and to the author of the article), it was really interesting to see the range of applications that neural networks are being used for.

But as someone who works a lot with modeling financial data, I have to say neural networks are pretty much the worst machine learning models for most financial applications. Some of the stuff that deep learning/NN's have achieved over the past year or two is simply mind-boggling. But what makes them so good at recognizing images or playing video games, is also what makes them poor for finance. They can model things perfectly. This is great for recognizing a leopard in a picture, because, quite literally, a leopard never does change its spots. Overfitting is perfect and viable. Financial data on the other hand regularly goes through time based regime changes (changes its spots), meaning that NN algorithm's are very susceptible to overfitting.

I should have said Jacques not Stephane.

@Graeme I am interested to hear what other models you prefer? I favor artificial neural networks and SVMs.

Author here. Graeme, before getting into it, I also expected deep learning would over-fit on all but the largest datasets. Now, I'm more excited about using it for smaller projects. There are numerous techniques that help prevent over-fitting, for example:

1. Early stopping with the optimization. During training, you monitor model performance on a held-out validation dataset. Stop training if it starts to over-fit.

2. Regularization. Both L1-norm and L2-norm are often used to keep weights small.

3. Dropout. Randomly drop units during training to keep them from co-adapting too much.

4. Autoencoders. First train a network to reconstruct the inputs as closely as possible even though the network is throttled in the middle. Then use the first half of the network for dimensionality reduction and tack on any other classification/regression model that you like to finish the job.

ML models applied to finance/trading need to be adaptive in order to cope with nonstationarity, and this is still a rather new subject. A much related issue is the lag introduced. That being said, since many patterns do recur over and over again in finance, and since momentum does manifest itself and is the basis for successful trading strategies even in the last few years, there is room for having ML/DL within a broader trading strategy.

Greg Harris's article and the references described were certainly thought provoking.

I am currently using deep learning techniques as one tool in managing our fund's algorithmic portfolio.

Overfitting is of course a very real problem, but one that it is also possible to overcome. In particular, I have had success using SAEs, regularization and dropouts.

Likewise non-stationarity is a very real problem, but again one that can be dealt with. In my experience, the tricky part is managing the trade off between lag and training window length.

Seems interesting

What are Neural Networks: they attempt to provide a construct where understanding is lacking. Statistics are sometimes used for the same purpose. In financial markets, many facts are known. Many others have a strong probability. IE if all markets go up by 5% around the globe, the futures premium will be high before the bell on the NYSE. So what is the information you ask of your NN. It must be the unknown missing variable info. Not the white noise. The trick is to feed your NN with a smaller vector set. If it improves on the error, you found a missing mechanism. All your blue estimators should produce a singular information matrix. The resulting NN motor, once out of the learning phase, is wired shut. So if it darkens your white noise, you're on to something. Cheers, Steve

Deep learning for trading is the future. I discussed this topic in a recent Linkedin update article “Artificial Intelligence - Prepare for Change” https://www.linkedin.com/pulse/artificial-intelligence-prepare-change-paul-monsted?trk=hp-feed-article-title-publish

It is no longer a question of if Artificial Intelligence will be used in trading but if you will be a part of it?

I gave a look a this one - Xiong et al. (2015) Deep Learning Stock Volatility with Google Domestic Trends. I am not denying that the paper is presenting some interesting ideas, but if you read it with attention, it reaches some conclusion quite baffling: when showing results, it basically says that simple linear regression model (RIDGE and LASSO) are better at predicting volatility than old but reliable GARCH. Is this reasonable, I ask ?