ARIMA models can be quite adept when it comes to modelling the overall trend of a series along with seasonal patterns.

In a previous article titled SARIMA: Forecasting Seasonal Data with Python and R, the use of an ARIMA model for forecasting maximum air temperature values for Dublin, Ireland was used.

The results showed significant accuracy, with 70% of the predictions ranging within 10% of the actual temperature values.

Forecasting More Extreme Weather Conditions

That said, the data that was being used for the previous example took temperature values that did not particularly show extreme values. For instance, the minimum temperature value was 4.8°C while the maximum temperature value was 28.7°C. Neither of these values lie outside the norm for typical yearly Irish weather.

However, let’s consider a more extreme example.

Image for post

Braemar is a village located in the Scottish highlands in Aberdeenshire, and is known as one of the coldest places in the United Kingdom in winter. In January 1982, a low of -27.2°C was recorded at this location according to the UK Met Office — which deviates strongly from the average minimum temperature of -1.5°C that was recorded between 1981–2010.

How would an ARIMA model perform when forecasting an abnormally cold winter for Braemar?

An ARIMA model is built using monthly Met Office data from January 1959 — July 2020 (contains public sector information licensed under the Open Government Licence v1.0).

#time-series-analysis #outliers #time-series-forecasting #data-science #statistics #data analysis

Limitations of ARIMA: Dealing with Outliers
10.70 GEEK