6
$\begingroup$

I know the response rate on this site is lower than most SE sites, but even taking that into account, the response rate to time series and forecasting related posts seems abysmal. Is there are reason for that? Can anything be done to improve on this?

$\endgroup$
  • 6
    $\begingroup$ I don't follow that tag, so I can't really speak to that, but in general it's a question of how many / which users have expertise on a topic & how much of their own time they can dedicate to helping others on the internet for free. $\endgroup$ – gung - Reinstate Monica Jan 24 '18 at 19:49
  • 16
    $\begingroup$ This is a statistical site, so do you have any data for that? Assuming it's true, and I don't disbelieve it, I have one guess only. Forecasting is difficult at best and many posts here that I see are frankly rather naive. They have the flavour "I have data and wish to forecast future values". So far, so good, but it is frustrating, I imagine, even for experts (I'm not one) to spell out what just about every forecasting text not based on snake-oil spells out: many methods, no panaceas, know your system, explore your data, etc., etc. $\endgroup$ – Nick Cox Jan 24 '18 at 19:49
  • 1
    $\begingroup$ The pattern I identify is common across many topics. Again, to forestall a fair comment, I have no data. I am just trading in impressions. $\endgroup$ – Nick Cox Jan 24 '18 at 19:55
  • 3
    $\begingroup$ Related: Why is our answer rate so low? $\endgroup$ – gung - Reinstate Monica Jan 24 '18 at 20:19
  • 7
    $\begingroup$ For some data pertaining to response rates, I searched on the top 8 tags & divided the total # of answers listed by the total # of questions. The resulting rates were: (1) [r] $.964$; (2) [regression] $0.973$; (3) [machine-learning] $0.997$; (4) [time-series] $0.899$; (5) [probability] $1.103$; (6) [hypothesis-testing] $1.048$; (7) [self-study] $0.967$; (8) [distributions] $1.029$. (cc @NickCox) $\endgroup$ – gung - Reinstate Monica Jan 24 '18 at 20:33
  • 1
    $\begingroup$ @gung so my intuition that time series has a response rate lower than the rest was correct. As to your initial comment about users dedicating time - given how involved and detailed many of the answers on this site are, why would time series be the one topic were experts can't be bothered to put too much time into answers? $\endgroup$ – Reinstate Monica Jan 24 '18 at 21:46
  • 8
    $\begingroup$ @Alex, I also had the impression that the answer rate for [t-s] questions was lower. I was surprised at how trivial the difference is. I don't have the impression there is a problem w/ [regression] questions, & I've never heard anyone complain about it, but the difference is less than one tenth of an answer per question. As for why it differs at all (even if only trivially), it again has to do w/ who & how many people there are answering those Qs. Eg, there may be fewer active users answering t-s Qs. $\endgroup$ – gung - Reinstate Monica Jan 25 '18 at 0:50
  • 5
    $\begingroup$ Another statistic to look at would be a proportion of unanswered questions from all questions. It's 5724/15172 = 0.37 for [regression] and 3431/7792 = 0.44 for [time-series]. @gung $\endgroup$ – amoeba says Reinstate Monica Jan 25 '18 at 8:19
10
$\begingroup$

It looks like I'm one of the more active users in the forecasting and time-series tags. Here is my subjective take.

I very much agree with Nick Cox' comment. Many, if not most of the questions especially in the forecasting tag seem to fall into one of two categories:

  1. The OP has quite obviously not read anything on forecasting, such as the excellent free online forecasting textbook Forecasting: Principles and Practice by Hyndman & Athanasopoulos (2nd ed.), which I keep recommending over and over. And over.

    With steadily decreasing enthusiasm, though. By now, anyone who looks through CV before posting, or who performs a simple search (the 1st edition of FPP is the very first hit if you search for "forecasting textbook" at duckduckgo.com - your results on Google will depend on your personal search bubble) should find it.

    If people obviously haven't found FPP, or have found it or a similar forecasting textbook, but have decided to rather post a question than look through the textbook before posting, this seems to be a clear case for downvoting ("shows no research effort"), but considering that these questions often seem to come from new users, slapping them with a downvote on their first question seems to be overly harsh. So I'll often comment that they look at FPP2, instead of writing a full answer.

  2. The question is a slight variation on one of the many, many questions we already have on forecasting. It's a hassle to find the closest match, and even if I do find one, it's often not obvious to the OP why this duplicate contains the answer to their question - if they knew enough about forecasting to understand the connection between the duplicate and their question, they wouldn't be asking in the first place.

    So voting to close as a duplicate will often require some commentary. I'd love to have the time to provide this commentary; unfortunately, my hungry children rely on me to put food on the table, and my employer does not pay me for spending time on CV. So I often write a comment to point to a related question and leave it at that.

    Case in point: I recently looked at a question where the OP essentially asked why the forecast was a flat line. I thought that there should have been tons of answers to this exact question, but I only found similar questions, no duplicates or canonical answers, so this time, I went and wrote a short answer, essentially to have a duplicate to refer to later on, since this kind of question will appear again.

It doesn't help that forecasting seems to be a topic with a high share of drive-by posters. I have a mental image of someone with little to no statistical knowledge being saddled with the task of forecasting sales, because they didn't look busy when the boss came around, or because they are known to be "good with figures". So off they go with little understanding and post a question along the lines of (1) and (2) above here. And even if they do get an answer, they don't accept or upvote. No, I'm not only here because rep tickles my already-inflated ego - but writing an answer and not getting any feedback doesn't encourage me to write more answers the next time a 1-rep user asks an un-upvoted not-very-good forecasting question. Sorry.

Conversely, my mental image (I emphasize that I'm just conjecturing here, I don't have data) of someone who posts in the mixed-model or chi-squared tags is someone who has at least looked into a statistics textbook or taken a stats course at some point in time, maybe while studying psychology or medicine. Answering here is typically more gratifying.

A related problem is that there is no clear point where a forecasting problem is "answered". Here, I am thinking of the time before the question is asked at CV. Someone may have read about Exponential Smoothing or ARIMA and applied it to their data, but they may feel that forecasts could be better. And this is often the point where they come and post questions on more elaborate methods here. In forecasting, it's not clear when the search ends. (Compare many other tags here.) This is exacerbated by my point above, of forecasters typically not having a lot of statistical training, so disabusing them of the notion that adding more data or using more complex models will not automatically improve forecast accuracy is an uphill battle.

So, to answer your question of whether anything can be done to improve the answer rate on forecasting and time-series questions: unless questions improve in quality (unlikely), it seems like we may need to be less stringent in closing as duplicates, or we might want to tolerate (and write) more short answers. Of course, it would help if we got more people to answer in these tags, too, so I'm looking forward to upvoting many answers from you in the future ;-)

$\endgroup$
  • 5
    $\begingroup$ Having been the second or third most active guy in the [forecasting] tag in 2016 and 2017, I largely agree. The issue with full or near duplicates has become quite tiresome for me; I do not want to write yet another nearly identical answer to a nearly identical question but I do not have the time and energy to find the duplicate either. It would be nice if we had some more canonical threads to refer to (5 or 10 could summarize a lot), but creating these threads again requires energy. I agree about the "no research effort" questions, too. As a consequence, I have reduced my activity in the tag. $\endgroup$ – Richard Hardy Feb 3 '18 at 10:03
3
$\begingroup$

Maybe the following analogy/viewpoint may help. Comparing with StackOverflow you could see questions as 'bug-reports'.

On SO those bugs are specific and concrete. If some bug or other type of problem exists in a program or computer language then many other people have encountered the problem before and when they come across somebody else with the same problem then they can easily recognize the problem and also easily explain how to hack/solve it, and this effort (which is not so much) is very rewarding because many people will encounter the same problem and be helped.

  • In statistics problems (bugs) are often much more vague and, unlike mathematics, definitions are not always precise (e.g imagine a question of the type: what is the best optimization/estimate/technique/etc.).

  • There are many very specific questions a computer bug can often be pinpointed to a clear cause. Problems in statistics are much more varied and solution strategies change depending the circumstances. This creates a lot of load, many different questions, that are basically the same. It is difficult to order the questions, and often questions (with little generality) are not so inspiring to answer.

  • There are many badly described/defined problems (with little information or by people that lack the ability to express themselves properly).

  • There are many problems that are very broad. Many questions are not like simple bugs. There are often questions that ask for a full consultation on an entire broad case/problem.

I imagine that this is might be more expressed in questions on 'time series'.

Or at least, from my perspective, I find time series often cumbersome, vague, and also very specific. To solve a problem you can pick out one of many solutions and variations thereof and it seems more like an art than science. Still I like these kind of problems, but often the solution strategy requires a more in-depth look into the problem case, and it is not often inviting to answer them via this medium.

(If I have time I will search for a statistic on the length of the questions. My hypothesis is that 'time-series' questions have much larger length in terms of words/letters, because they are so specific. also anything related to 'forecasting' may do worse. E.g. people dump a lot of data and images, which still doe not provide a good/sufficient picture of the problem, and ask for a solution)

$\endgroup$
  • 1
    $\begingroup$ Good points, but last sentence needs elaborating. Everything until the last sentence could be a better fit for stats.meta.stackexchange.com/questions/1538/…. $\endgroup$ – amoeba says Reinstate Monica Jan 25 '18 at 15:19
  • 3
    $\begingroup$ I will see how to improve it. I have already edited a lot and am working on this answer. My basic idea is that I find time series often cumbersome, vague, and also very specific. Solving a problem you can pick out one of many solutions and variations theroef and it seems more like an art than science. $\endgroup$ – Sextus Empiricus Jan 25 '18 at 15:20
  • 2
    $\begingroup$ The focus of the question is why are time series different (and confirming that they are different is crucial; @gung's evidence suggests that the differences are not enormous and thus should not be over-interpreted). Thus it's a question of how many act in the presence of time series questions rather than how you feel personally. At the same time, I too make guesses as a projection of my impressions! $\endgroup$ – Nick Cox Jan 25 '18 at 20:04
  • $\begingroup$ @NickCox this seems more like a problem that originates in the question being very open. Therefore, subjective input is necessary before some objective answer can be formulated, if such answer is possible at all. If these speculations on effects are over-interpreting the tiny differences then maybe the question should not be asked? (although I believe the effects may even still be present without the differences, since there are many multiple positive and negative effects and verifying/quantifying the OPs propositions by looking at simple differences does not seem so crucial to me) $\endgroup$ – Sextus Empiricus Jan 26 '18 at 12:14
  • $\begingroup$ We agree on essentials. $\endgroup$ – Nick Cox Jan 26 '18 at 12:17

You must log in to answer this question.

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