All Hyndsight posts by date
The latest minor release of the forecast package has now been approved on CRAN and should be available in the next day or so.
Version 8.5 contains the following new features
Updated tsCV() to handle exogenous regressors. Reimplemented naive(), snaive(), rwf() for substantial speed improvements. Added support for passing arguments to auto.arima() unit root tests. Improved auto.arima() stepwise search algorithm (some neighbouring models were missed previously). We haven’t done a major release for two years, and there is unlikely to be another one now.
The International Institute of Forecasters has established interest group sections, devoted to specialized domains of forecasting. One of the first such sections will be for early career researchers.
So if you are a PhD student, post-doc, or otherwise a relatively junior researcher working in forecasting, this is for you! The first events will be during the ISF in Thessaloniki in June 2019, including the following:
ECR welcoming event. A meet and greet event prior to the ISF welcome reception.
This question seems to come up frequently on crossvalidated.com or in my inbox.
I have this time series, however it yields different results when I use the auto.arima and Arima functions.
library(forecast) xd <- ts(c(23786, 25955, 54373, 21561, 14552, 13284, 12714, 11821, 15445, 21307, 17228, 20007, 23065, 32811, 43147, 15127, 13497, 12224, 11412, 11888, 14210,18978, 15782, 17216, 16417, 22861, 42616, 17057, 9741, 10503, 7170, 10686, 9762, 15773, 15280, 13212, 14784, 26104, 29947), frequency = 12, start=c(2014,1), end=c(2017,3)) fit1 <- auto.
This week I’ve been attending the Functional Data and Beyond workshop at the Matrix centre in Creswick.
I spoke yesterday about using ggplot2 for functional data graphics, rather than the custom-built plotting functionality available in the many functional data packages, including my own rainbow package written with Hanlin Shang.
It is a much more powerful and flexible way to work, so I thought it would be useful to share some examples.
Following the highly successful M4 Forecasting Competition, there will be a conference held on 10-11 December at Tribeca Rooftop, New York, to discuss the results. The conference will elaborate on the findings of the M4 Competition, with prominent speakers from leading business firms (Amazon, Uber, Google, Microsoft, SAS, and ProLogistica) and top universities. Nassim Nicholas Taleb will deliver a keynote address about uncertainty in forecasting and elaborate on his claims that “tail risks are much worse now than in 2007” while Spyros Makridakis will discuss how organizations can benefit by improving the accuracy of their predictions and assessing uncertainty realistically.
The annual Melbourne Data Science Initiative (or MeDaScIn, pronounced medicine) is on again next month (24-27 September) with lots of tutorials, and the annual datathon.
This year there will be a “Forecasting with R” workshop (25 September) led my two of my Monash colleagues – George Athanasopoulos and Elena Sanina.
Another great tutorial will be led by Steph Kovalchik (from Tennis Australia) on sports analytics with R (24 September).
For the full list of tutorials, see the MeDaScIn website.
Next week is National Science Week and there are a few mathematics activities happening around Melbourne that are being sponsored by ACEMS.
Elsewhere in Melbourne: Mon 13 Aug 2018, 6:00pm - 7:30pm
Public Talk: Is this your card?
Location: University of Melbourne
Speakers: Anthony Mays & Jen Palisse
Pick a card, any card! The immortal phrase of the magician. In this talk, we’ll look at some great card tricks that have simple maths behind them.
Occasionally R might not be the tool you want to use (hard to believe, but apparently that happens). Then you may need to export some data from R via a csv file. When the data is stored as a ts object, the time index can easily get lost. So I wrote a little function to make this easier, using the tsibble package to do almost all of the work in looking after the time index.
All talks from useR!2018, held in Brisbane last week, are now available on YouTube.
Links to talks from members of my research team are given below.
Many users have tried to do a seasonal decomposition with a short time series, and hit the error “Series has less than two periods”.
The problem is that the usual methods of decomposition (e.g., decompose and stl) estimate seasonality using at least as many degrees of freedom as there are seasonal periods. So you need at least two observations per seasonal period to be able to distinguish seasonality from noise.