@dang I think worth to place the year (1993) on the post.<p>...<p>It's quite refreshing to see some results of this competition specially in comparison with the M5 that is the most relevant competition in time series.<p>For practitioners one thing that is important to consider is that this competition had datasets with way longer time intervals (I think M5 has a lot of time series but in a short horizon) and the methods used back in the day was way more sophisticated.
The M-* series of competitions are the current major forecasting competitions in the time series world.<p><a href="https://en.wikipedia.org/wiki/Makridakis_Competitions" rel="nofollow noreferrer">https://en.wikipedia.org/wiki/Makridakis_Competitions</a><p>I believe the top performing models in recent years have all been gradient boosted trees, usually then fed into some secondary ensembling scheme.
> This chapter reports on a competition run through the Santa Fe Institute in which participants from a range of relevant disciplines applied a variety of time series analysis tools to a small group of common data sets in order to help make meaningful comparisons among their approaches. The design and the results of this competiton are described, and the historical and theoretical backgrounds necessary to understand the successful entries are reviewed.<p>The competition has already been run. It's from 1993.
I seriously think the long term future of computing is massive scale time series correlation with analog hardware that doesn't do discrete processing with I/O and on/off transitors<p>I vomited my theory down years ago - though I probably need to refine it.<p><a href="https://kemendo.com/research/streaminference.html" rel="nofollow noreferrer">https://kemendo.com/research/streaminference.html</a><p>Transformers IMO are basically pointing towards time series predictions as being the root of epistemological grounding for "intelligence"<p>We probably need a new type of computer to realize this (nb. A previous good friend did bring up the point that digital sampling at a high enough rate could approximate biological "continuous" sampling - but I think it's flawed theory)