8
$\begingroup$

I've noticed that stock trading is a prominent theme among and and questions. I think this is because (1) stock trading is an easy problem to understand (2) the data is readily available and (3) striking it rich on the stock market is a popular fantasy.

Does this warrant the creation of a canonical thread? What should that thread address?

I ask because I'm not aware of another substantive NN/DNN/ML topic which consistently attracts questions. The closest runner-up that I can recall are questions about using neural networks to fit sine waves. (This likely reflects my own biases, since I read most and questions but a much smaller fraction of other questions. Most of these threads are found in this search: https://stats.stackexchange.com/search?q=stock+trade)

Usually, the subject matter itself is of secondary importance on stats.SE and users are asking methodological questions about how to design an experiment or account for an unusual quirk of their data (dependency structure such as spatial or temporal correlation, mixed effects, random effects, etc.).

As a statistical topic, there are some general comments which would appear to apply to all of the shallow stock trading questions, such as how one might frame the discrete choices available to the model (buy, sell, hold, short, etc.) as a reinforcement or more conventional approach. This seems like it's on-topic.

But the broadest versions of these questions are perhaps not suitable because they veer into quantitative finance. The most general question, "How do I make a profitable stock trading algorithm?" must certainly require a deep understanding of QF because feature engineering (how to represent financial data in a manner that is amenable to your algorithm) and computer engineering (how to load and process data to execute trades before your decision is "stale"). I think those kinds of questions are best answered by QF.SE, since it is certainly their core topic.


This question is motivated by this thread, where I opted to write a rather general answer because I had happened to spend a few dreadful weeks learning about QF. How to train a stock trading neural network so that the 'profit' parameter is maximized?

$\endgroup$
  • 2
    $\begingroup$ If you suspect there's value in a canonical thread, you'll very likely be correct. That some questions will veer into quant.SE territory is just part of how any attempt to divide up knowledge is going to work; we'll always have to deal with such issues, so I doubt that's worth more worry than keeping it in mind when we migrate/close threads as having strayed off our topics. $\endgroup$ – Glen_b -Reinstate Monica Nov 28 '18 at 23:41
7
$\begingroup$

I doubt we really need canonical threads on different subject matter for data. It might be helpful to have canonical threads on different data types (binomial vs. normal), but not the meaning behind the numbers (e.g., presence of headache vs. not, presence of gastric distress vs. not, etc.). That seems to me like it negates the need for a canonical stock market thread. I'm not saying it should be off limits, but I mostly don't think it makes sense. In particular, the follow-up question, "What would it address?" seems like a major warning sign.

In the specific case of the stock market (or generalize to include other investment options, such as bonds, commodities, forex, etc.), I do see some value in a short thread that explains the principle of no arbitrage, simply because there is no escaping the occasional ignorant question. However, we arguably already have that: What machine learning algorithm can be used to predict the stock market?

On the other hand, if there were a type of question, say about neural networks, that is contextualized in the stock market and pops up commonly, but is always poorly asked, then a canonical thread on that issue would be helpful. Moreover, it would probably be most helpful to use the stock market as the example situation to motivate the question. I don't follow either investing or these types of questions on CV well enough to provide a great example, but something like modeling / forecasting a mean when the error variance changes unpredictably might be a the sort of thing where it would be helpful.

$\endgroup$
  • 1
    $\begingroup$ +1 These are fair points. I suppose the statistical topics of stock trading are probably covered by our threads about time-series analysis, reinforcement learning, and the usual caveats about out-of-sample generalization and selection bias. No arbitrage/efficient markets and related topics seem to fall in the same bucket as gastric distress -- particular to the subject, not the mode of analysis. $\endgroup$ – Reinstate Monica Nov 29 '18 at 16:35
1
$\begingroup$

I would avoid it. This is my area of research. The two theories that won Nobel prizes were empirically foreclosed as wrong starting in 1963 and have had NO successful validation studies. ANNs as universal function approximators tend to fall into traps that you see in the traditional methodologies and don't work out of sample. It only looks easy.

$\endgroup$
  • 1
    $\begingroup$ I didn't mean to suggest that there is "easy money" to be made in strapping a home-brewed neural net to a web API. If I were to write such an answer, it would emphasize that making a profitable model would be extremely challenging. $\endgroup$ – Reinstate Monica Dec 6 '18 at 4:25
  • $\begingroup$ I have been looking at the verification and validation of non-deterministic systems such as ANNs and I think the field needs a lot more formal mathematical research first. I don't think the field is advanced enough, although there is general work on V&V of ANNs by Marjorah Darrah and a few others. The distributions and processes are a mixture of distributions with first moments and those without. $\endgroup$ – Dave Harris Dec 6 '18 at 4:42

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .