interesting post. I use Nvidia 980Ti cards as they get you the most cuda cores for the money. K40 and K80 cards are way overpriced. I installed cuda support for Theano and it does speed up computations by quite a bit. I use the term data augmentation in a slightly different way but I get what you mean. I did test with inputting time series, for example, or other input data in a reverse order and the output and prediction rate turned out to be almost identical which confirmed that the learning algorithm is capable of building associations and assigning weights regardless of the order of the input data series. Obviously a complete re-ordering of input data is an entirely different story.
Hey, I haven't used these libraries yet (but i will), but since you have experience in this do you run your convolutional network on Theano? would you recommend Theano over Keras or TensorFlow? also since you come from HF industry can you comment on the features that you pass to the algos? are those some price derivatives or you are getting more creative going with image recognition or perhaps weather forecasts etc.? Basically why I'm asking this is because there was a competition on kaggle hosted by Winton Capital that provided over 130 features for price prediction and after much data crunching the general opinion on the forums was that almost none of them had predictive power except for lagged prices, which is a simple ARMA or GARCH model anyway. So I would be very much interested to hear about your experience with feature engineering and what approach do you take in selecting the exogenous variable. Thanks!
Theano should do just fine, Tensorflow is easier to install but the GPU support in the non-commercial version is not yet where I would like to see it. Plus I do not want to lock myself into Tensorflow too early on. Theano is pretty powerful and it works well for what I am doing. I am by far no expert and I think very few worldwide really are. Also, I do not come from a hft background. I started out as otc rates derivatives market maker at an ibank, then transitioned into proprietary trading at another bank and later on hedge funds before I moved into pure quant trading. I am not into market making hence hft is not something I am keen on getting into, aside the fact that it has become prohibitively expensive to join. Do you have a link to your below statement re the Winton Cap challenge? I have heard differently and there are even several deep learning models in public domain for free that definitely beat Garch models in modeling volatility.
By HF I meant hedge fund. You can look at this thread: https://www.kaggle.com/c/the-winton-stock-market-challenge/forums/t/18584/solution-sharing I'm sure there can be solutions for volatility forecasting that are better than GARCH, however for that particular competition many people comment that all of NN stuff failed, but the simplest methods were superior, i.e linear regression with MAE optimization etc. Thus it is interesting to know what approaches do you take in modeling.
thanks for the clarification re HF. I took a look at the link you posted. There was one single guy who mentioned convolutional networks, or so he claimed, but his post was full of simply typos and grammatical errors. Not that English is a hindering point to be a great mathematician but then it makes one wonder how thorough that user was when he did not even bother to spell check his post. Then there were 2 mentions of neural networks and in both cases the users resorted to more "conventional" approaches, both of which did not seem to demonstrate the users had any clue what they was doing. I did not get the impression the submitted work in aggregate was of too high quality. When I mentioned Kaggle competitions earlier I was talking about challenges that attracted the top brass in neutral network research. There was a group that modeled and discovered associations of previously unmapped genes. There was another challenge where the contestants pushed prediction rates of a stated classification problem from a previous industry standard of 95-95% to up to 99.4x percent. Not looking to demean the work done in the challenge you posted but...well...I stated my impression above for what it's worth.
What makes you believe successes in those fields suggest anything about possible success in financial markets? Whole different beast.
My point is that it is much easier to teach an intelligent scientist financial markets than to have the eyes opened of a retail "trader" without the slightest clue of statistics or mathematics nor the inclination to learn such.