A blog by Rob J Hyndman 

Twitter Gplus RSS

Automatic time series forecasting in Granada

Published on 31 January 2014

In two weeks I am pre­sent­ing a work­shop at the Uni­ver­sity of Granada (Spain) on Auto­matic Time Series Fore­cast­ing.

Unlike most of my talks, this is not intended to be pri­mar­ily about my own research. Rather it is to pro­vide a state-​​of-​​the-​​art overview of the topic (at a level suit­able for Mas­ters stu­dents in Com­puter Sci­ence). I thought I’d pro­vide some his­tor­i­cal per­spec­tive on the devel­op­ment of auto­matic time series fore­cast­ing, plus give some com­ments on the cur­rent best practices.


Those attend­ing are asked to do some pre-​​reading. Here are the papers I’ve requested to be read beforehand:

  1. Makri­dakis & Hibon, (JRSSA 1979) was the first seri­ous attempt at a large empir­i­cal eval­u­a­tion of fore­cast meth­ods.  It was fol­lowed by dis­cus­sion which is doc­u­mented in the paper. The dis­cus­sion is heated and enter­tain­ing, and impor­tant in under­stand­ing the dif­fer­ent per­spec­tives on this topic at the time.
  2. Makri­dakis & Hibon (IJF, 2000) can be con­sid­ered a suc­ces­sor to the 1979 paper and describes the largest pub­lished fore­cast­ing com­pe­ti­tion to date. While the algo­rithms used are not dis­cussed in any detail, the result­ing com­par­isons are provided.
  3. Gomez and Mar­vall (2001) must be included, not just because it is by two well-​​known Span­ish time series ana­lysts. It pro­vides the best intro­duc­tion to auto­matic time series model selec­tion and is a use­ful start­ing point on some of the theory.
  4. Hyn­d­man and Khan­dakar (JSS 2008) describes two extremely widely used auto­matic fore­cast­ing algo­rithms. These have been improved in the last five years, but the basic algo­rithms are most clearly described there.

Obvi­ously more has hap­pened in the field in the last 5 or 6 years, and I will also talk a lit­tle about those developments.


Many appli­ca­tions require a large num­ber of time series to be fore­cast com­pletely auto­mat­i­cally. For exam­ple, man­u­fac­tur­ing com­pa­nies often require weekly fore­casts of demand for thou­sands of prod­ucts at dozens of loca­tions in order to plan dis­tri­b­u­tion and main­tain suit­able inven­tory stocks. In these cir­cum­stances, it is not fea­si­ble for time series mod­els to be devel­oped for each series by an expe­ri­enced ana­lyst. Instead, an auto­matic fore­cast­ing algo­rithm is required.

In addi­tion to pro­vid­ing auto­matic fore­casts when required, these algo­rithms also pro­vide high qual­ity bench­marks that can be used when devel­op­ing more spe­cific and spe­cial­ized fore­cast­ing models.

I will describe some algo­rithms for auto­mat­i­cally fore­cast­ing uni­vari­ate time series that have been devel­oped over the last 20 years. The role of fore­cast­ing com­pe­ti­tions in com­par­ing the fore­cast accu­racy of these algo­rithms will also be discussed.

Finally, I will describe meth­ods for eval­u­at­ing fore­cast­ing algo­rithms which use the avail­able data as effi­ciently as possible.


Related Posts:

1 Comment  comments 
  • Adrian S

    Look­ing for­ward to it ! Would there be a video upload ?