I have just been to the EURO2018 conference (29th European Conference on Operational Research) in Valencia (Spain). Many comments come to my mind. Firstly, Valencia is an impressive city to live and to make some sightseeing. People is really, really friendly and lively (and lovely). Secondly, once more I have realized that everybody speaks about big data, big data, big data… Finally, though I did not speak about big data, people kindly listened to me, and I received some support from listeners.
My presentation was entitled (jointly prepared with Marco A Villegas and Diego Villegas) ‘Automatic forecasting support system for business analytics applications based on unobserved components models‘ (slides here) and basically, I tried to do three main things:
- Publicize the use of Unobserved Components (UC) models for forecasting. In the end they are not so bad in comparison to Exponential Smoothing or ARIMA models, and actually I showed that UC are better in some cases and no worse in others.
- Develop a general Automatic Forecasting Support System based on UC models. This is the first time that automatic identification of UC has been proposed. The idea is to fit all possible combinations of trend, seasonal and irregular sensible models and select the one with the smallest BIC. I am sure this procedure sounds familiar to many people.
- Make a thorough comparison among methods and different implementation of each methods. In particular, two implementations of Exponential Smoothing, three of ARIMA, two naive and obviously, the UC. For the database we used the UC was the best.
I do not pretend to say that UC are going to be the best in absolute every situation. The argument is no more than stating that UC models are no worse than other well stablished methods. And they have been around for a while but very little used in forecasting (check any forecasting competition or in general the forecasting literature).
In all this stuff, SSpace plays an essential role. SSpace is a MATLAB toolbox (it will come out soon published in the Journal of Statistical Software) for general and flexible State Space models. Please, feel free to download it from here and try it out.
A second presentation in which I participated (and presented by Juan R. Trapero, check out here) was entitled ‘Implementation of exponential smoothing forecasting method in a GPU for big data problems’. He showed that the use of GPUs is the way to go in forecasting applications when massive databases ought to be forecast. The advantages in computation time are impressive in big data problems. It was interesting to check that many people became interested straight away. We hope this materializes in something else than just interest, but that is another story I will tell you in the near future…