Dynamic forecasts - with Bayesian linear models and neural networks

Sigrid Keydana, Trivadis
2017/14/11

About me & my employer

 

Trivadis

  • DACH-based IT consulting and service company, from traditional technologies to big data/machine learning/data science

My background

  • from psychology/statistics via software development and database engineering to data science and ML/DL

My passion

  • machine learning and deep learning
  • data science and (Bayesian) statistics
  • explanation/understanding over prediction accuracy

Where to find me

 

 

Dynamic models for timeseries prediction, - why?

Our task today: forecast men's 400m Olympic winning times

 

It's 2000, just before the Olympics. This is the data we have:

plot of chunk unnamed-chunk-2

This looks like we might even try linear regression...

 

lm(seconds ~ year, male400_1996) %>% coefficients()
(Intercept)        year 
  207.46609    -0.08257 

plot of chunk unnamed-chunk-4

Unfortunately...

 

plot of chunk unnamed-chunk-5

OK then... let's just use ARIMA

 

(yes, the series is not totally regular, but let me make the point ;-))