3
$\begingroup$

What are the main takeaways we, as statistics community, learned from (time series) competitions?

Kaggle, or the M.. competitions, seem such a valuable source, but the (only) main insight I remember is that boosting algorithms such as LightGBM or XGBoost tend to work very well. I am currently teaching a course on practical machine learning and want to tell the students the main insights gained in time series modeling in the wild. I am also wondering whether it still makes sense to talk about the theory behind ARIMA, for example (depending on whether the algorithm is still competitive in practice).

Before raising this question, I had a look at the conclusions of the M5 competition, but some of the key takeaways, such as Exogenous/explanatory variables were important for improving the forecasting accuracy of time series methods. seem rather trivial. What I am looking for is a kind of meta-study aggregating the results of different competitions.

Btw, this is (surprisingly) the only related question I found here.

Improve this question
$\endgroup$
7
  • 3
    $\begingroup$ There is an entire special issue of the International Journal of Forecasting on just the most recent M5 forecasting competition. I would say if an entire journal issue can be devoted to just one aspect of a CV question, the question is too broad. $\endgroup$
    – Stephan Kolassa
    Commented Oct 30, 2022 at 16:04
  • $\begingroup$ Before posting this question, I read the conclusions here, but to be honest, I would have expected more key insights than, for example, Exogenous/explanatory variables were important for improving the forecasting accuracy of time series methods. That is why I am asking. I will edit my question. $\endgroup$
    – Arne Jonas Warnke
    Commented Oct 30, 2022 at 16:07
  • $\begingroup$ I don't think there are truly clear and obvious insights. Yes, the top methods all used boosting. But there was one very well performing white box method (see the paper by de Rezende et al. in that issue), and in any case, there was a lot of instability on the leaderboard, so one wonders how much of the results are down to people submitting high variance forecasts (see the commentary by Ma and Fildes). And in any case, the question remains whether the increase in accuracy is worth the added complexity (see my commentary). And all that is just the M5. The M6 is coming up (and very different). $\endgroup$
    – Stephan Kolassa
    Commented Oct 30, 2022 at 16:15
  • $\begingroup$ Thanks, there was a lot of instability is a crucial insight for me, and I think that is worth an answer! $\endgroup$
    – Arne Jonas Warnke
    Commented Oct 30, 2022 at 16:17
  • 2
    $\begingroup$ "What are the main takeaways" asside from.being broad this type of question is also a bit subjective which can be problematic for the q&a format. $\endgroup$
    – Sextus Empiricus
    Commented Oct 30, 2022 at 16:19

0

Reset to default

Browse other questions tagged or ask your own question.