\( \newcommand{\mathbbm}[1]{\boldsymbol{\mathbf{#1}}} \)

Chapter 8 Introduction to ARIMA

Another important dynamic element in ADAM is the ARIMA model (developed originally by Box and Jenkins, 1976). ARIMA stands for “AutoRegressive Integrated Moving Average”, although the name does not tell much on its own and needs additional explanation, which will be provided in this chapter.

The main idea of the model is that the data might have dynamic relations over time, where the new values depend on the values on the previous observations. This becomes more obvious in the case of engineering systems and modelling physical processes. For example, Box and Jenkins (1976) use an example of a series of CO\(_2\) output of a furnace when the input gas rate changes. In this case, the elements of the ARIMA process are natural, as the CO\(_2\) cannot just drop to zero when the gas is switched off – it will leave the furnace in reducing quantity over time (i.e. leaving \(\phi_1\times100\%\) of CO\(_2\) in the next minute, where \(\phi_1\) is a parameter in the model).

Another example where AR processes are natural is modelling the temperature in the room, measured with five minute intervals. In this case, the temperature at 5:30 p.m. will depend on the one at 5:25 p.m.: if the temperature outside the room is lower, then the one in the room will go down slightly due to the loss of heat. Every five minutes it will go down on average by some quantity \(\phi_1\).

Both these examples describe the AR(1) process, and in both of them the ARIMA model can be considered a “true model” (see discussion in Section 1.4). Unfortunately, when it comes to time series in the social or business domains, it becomes very difficult to motivate ARIMA usage from the modelling point of view. For example, the demand for a product does not reproduce itself and in real life does not depend on the demand on previous observations. So, if we construct an ARIMA for such a process, we turn a blind eye to the fact that the observed time series relations in the data are most probably spurious. At best, in this case, ARIMA can be considered a very crude approximation of a complex process (demand is typically influenced by price changes, consumer behaviour, and promotional activities, etc.). Thus, whenever we work with ARIMA models in social or business domains, we should keep in mind that they are wrong even from the philosophical point of view. Nevertheless, they still can be useful (as was pointed out by one of the original authors, George Box), which is why we discuss them in this chapter. We focus our discussion on forecasting with ARIMA. A reader interested in time series analysis is directed to Box and Jenkins (1976) or to more modern editions of that textbook. ARIMA is also well explained in Chapter 6 of Ord et al. (2017).

This chapter will discuss the main theoretical properties of ARIMA processes (i.e. what would happen if the data indeed followed the specified model), moving to more practical aspects in the next chapter. We start the discussion with the non-seasonal ARIMA models, explaining what the forecasts from those models would look like, then move to the seasonal and multi-seasonal ARIMA, then discuss the classical Box-Jenkins approach for ARIMA order selection and its limitations. Finally, we explain the connection between ARIMA and ETS models.


• Box, G., Jenkins, G., 1976. Time Series Analysis: Forecasting and Control. Holden-day, Oakland, California.
• Ord, J.Keith., Fildes, R., Kourentzes, N., 2017. Principles of Business Forecasting, 2nd ed. Wessex Press, Inc, New York, New York, USA.