(Please note I have read through "How to handle changing input vector length with neural networks" but somehow this is different)

As most of us know, financial trade data exhibit different volume and velocity depending on predictable and unpredictable factors, from national holidays to wars. The fine detail of these financial signals is mostly considered "noise" and a lot of conclusions can be drawn by so-called timeframes, i.e. averages and aggregates over 1 minute, 1 day, or even 1 year. I would like to hunt for the "signal in the noise" with some deeplearning package, but variable length vectors do not seem to be an option. I can think of workarounds, and simply adding the length of the vector (the number of ticks) to a standardized timeframe is not bad at all, but I am not happy with my workarounds. Certainly a spectral parametric description would go a long way towards capturing the waveform in a fixed vector length, but I would be happy to hear of a "natural way"-

  • $\begingroup$ Having trade time series stretched out across single vectors doesn't make a lot of sense since so much information is lost. Other data structures are possible. For instance, treating each unit of time -- whether high frequency or aggregated -- as a unique observation within a trading cross-section that is described by a set of fixed features is a highly flexible approach. This would create a panel or pool of information which would be much more informative than a single vector. $\endgroup$ – Mike Hunter Aug 1 '16 at 18:04
  • $\begingroup$ I do not properly grok your comment, perhaps because you are a statistician and I lack the vocabulary - I will give it a go with the help of some dictionary later. But regarding the first sentence, I am of course prepared to lose "so much information" as an improvement over losing/leaving out all that information. But true, I am looking for a way to lose much less, in more ways than one :) $\endgroup$ – fidetrainerNET Aug 1 '16 at 18:39
  • $\begingroup$ The variable length comes from the number of ticks, right? Can each tick me an input? $\endgroup$ – roundsquare Aug 1 '16 at 20:53
  • $\begingroup$ guessing that "me" is a "be". Each tick should be an input, we are debating what to do if you got 0 ticks per ms versus 100 ticks per ms. $\endgroup$ – fidetrainerNET Aug 2 '16 at 10:36

Your Answer

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

Browse other questions tagged or ask your own question.