The thief package for R: Temporal HIErarchical Forecasting

I have a new R package available to do temporal hierarchical forecasting, based on my paper with George Athanasopoulos, Nikolaos Kourentzes and Fotios Petropoulos. (Guess the odd guy out there!)

It is called “thief” – an acronym for Temporal HIErarchical Forecasting. The idea is to take a seasonal time series, and compute all possible temporal aggregations that result in an integer number of observations per year. For example, a quarterly time series is aggregated to biannual and annual; while a monthly time series is aggregated to 2-monthly, quarterly, 4-monthly, biannual and annual. Each of the resulting time series are forecast, and then the forecasts are reconciled using the hierarchical reconciliation algorithm described in our paper.

It turns out that this tends to give better forecasts, even though no new information has been added, especially for time series with long seasonal periods. It also allows different types of forecasts for different forecast horizons to be combined in a consistent manner.

Continue reading →

Statistics positions available at Monash University

We are hiring again, and looking for people in statistics, econometrics and related fields (such as actuarial science, machine learning, and business analytics). We have a strong business analytics group (with particular expertise in data visualization, machine learning, statistical computing, R, and forecasting), and it would be great to see it grow. The official advert follows.

Continue reading →

forecast v7 and ggplot2 graphics

Version 7 of the forecast package was released on CRAN about a month ago, but I'm only just getting around to posting about the new features.

The most visible feature was the introduction of ggplot2 graphics. I first wrote the forecast package before ggplot2 existed, and so only base graphics were available. But I figured it was time to modernize and use the nice features available from ggplot2. The following examples illustrate the main new graphical functionality.

For illustration purposes, I'm using the male and female monthly deaths from lung diseases in the UK.

Continue reading →

Melbourne Data Science Initiative 2016

In just over three weeks, the inaugural MeDaScIn event will take place. This is an initiative to grow the talent pool of local data scientists and to promote Melbourne as a world city of excellence in Data Science.

The main event takes place on Friday 6th May, with lots of interesting sounding titles and speakers from business and government. I’m the only academic speaker on the program, giving the closing talk on “Automatic FoRecasting”. Earlier in the day I am running a forecasting workshop where I will discuss forecasting issues and answer questions for about 90 minutes. There are still a few places left for the main event, and for the workshops. Book soon if you want to attend.

All the details are here.

Plotting overlapping prediction intervals

I often see figures with two sets of prediction intervals plotted on the same graph using different line types to distinguish them. The results are almost always unreadable. A better way to do this is to use semi-transparent shaded regions. Here is an example showing two sets of forecasts for the Nile River flow.

library(forecast)
 
f1 = forecast(auto.arima(Nile, lambda=0), h=20, level=95)
f2 = forecast(ets(Nile), h=20, level=95)
 
plot(f1, shadecol=rgb(0,0,1,.4), flwd=1,
     main="Forecasts of Nile River flow",
     xlab="Year", ylab="Billions of cubic metres")
polygon(c(time(f2$mean),rev(time(f2$mean))),
        c(f2$lower,rev(f2$upper)),
        col=rgb(1,0,0,.4),
        border=FALSE)
lines(f2$mean, col=rgb(.7,0,0))
legend("bottomleft", 
       fill=c(rgb(0,0,1,.4),rgb(1,0,0,.4)),
       col=c("blue","red"), lty=1,
       legend=c("ARIMA","ETS"))

NileRiverFlow

The blue region shows 95% prediction intervals for the ARIMA forecasts, while the red region shows 95% prediction intervals for the ETS forecasts. Where they overlap, the colors blend to make purple. In this case, the point forecasts are quite close, but the prediction intervals are relatively different.