1
$\begingroup$

I am working on a multinomial machine learning algorithm that labels stocks with buy/sell signals. My code updates with the most recent quantitative data about the stocks daily, so obviously the data changes daily and each time I run the code. My problem is that each stock is an observation, and for ML purposes the data set I am left with is very small (<10000 obs). I have fiddled with various ways to increase the sample size - but each way has been extremely labor intensive. I have wondered if since the data changes daily, the predictions will also change daily, however slightly (being multinomial gives more levels and maximizes variety). If I save each generated daily data set, can I stack them over time and add them to the latest generated dataset? That is, is it alright to ignore change in time when making a dataset - or is it more complex than that?

$\endgroup$

closed as unclear what you're asking by Michael Chernick, kjetil b halvorsen, mdewey, Yves, COOLSerdash Mar 8 at 11:38

Please clarify your specific problem or add additional details to highlight exactly what you need. As it's currently written, it’s hard to tell exactly what you're asking. See the How to Ask page for help clarifying this question. If this question can be reworded to fit the rules in the help center, please edit the question.

  • $\begingroup$ I think the question is something like "what is the danger in recycling data through a black-box machine learner". I think there are punctuations in the algorithms "others" use, for which I consider Buffet's love of non-transitive dice to be an example. I think he uses "higher quality" random number generation schemes and sequences, and looks for "lower quality" runs to extract value from "hidden" large transactions. Using random walks to disguise the movement before he did that would work, but after, you lose much more in the transfer. $\endgroup$ – EngrStudent Mar 6 at 15:07
0
$\begingroup$

As far goes my past work on predictions, it is hard to infer something from what you've described as the method you've used for predictions wasn't mentioned. Also, the feedback will be different from algorithm to algorithm, data set from data set. So, there is almost no ground truth when it comes to this questionings and for that reason I would suggest is testing both out, in an analytic form.

However, from my point of view I would be surprised if the prediction worsen, as you would get more far to the sample behavior and closer to the to the population one.

$\endgroup$

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