Author: John Hallman

Papers:

Introduction

The Makridakis competitions (aka the M competitions) are a series of open competitions in the space of time series forecasting that were first held in 1982.

The competitions have appeared at highly irregular intervals - the second one was held in 1993, the third one in 2000, the fourth one in 2018, the fifth one in 2020, and the sixth and most recent one is currently ongoing.

Today, we will be exploring the results from the M5 competition!

Table of Contents

Motivation

Although the results from the M4 competition have been analyzed in more detail, there are a couple reasons why the M5 competition is particularly interesting and relevant:

<aside> 💡 Reason 1: data type

The datasets selected for the M4 competition are idealistically diverse with a richness that is rare to find in a typical cloud warehouse, with frequency ranging from yearly to hourly taken from stock, bond, and real estate prices, demographics, education, and more.

On the other hand, the M5 competition focused solely on daily sales data provided by Walmart: the kind of large, streaming datasets we see here at Sisu. (1) All the time series are daily. (2) The data is hierarchical in nature, containing time series for each item, category, store, and more. (3) It is volumetric, just like all our customers' datasets. (4) It displays intermittency, i.e. sporadic zero values.

</aside>

<aside> 💡 Reason 2: focus on uncertainty

For the first time, the M competition focuses not just on forecast accuracy, but also on confidence intervals. This means that models submitted to these competitions are useful not just for forecasting but also for anomaly detection!

</aside>