13.5 Intermittent demand challenges
Intermittent demand is complicated and is difficult to work with. As a result, there are several challenges that are related to ADAM model specifically and to the intermittent demand in general that are worth discussing.
First, given the presence of zeroes, the decomposition of intermittent time series does not make sense. The classical time series model assumes that the demand happens on every observation, while the intermittent demand happens irregularly. This makes all the conventional models inapplicable to the problem, although some of them might still work well in some cases (for example, SES in case of slightly intermittent data).
Second, follows directly from the previous point: while in theory it is possible to use any ETS / ARIMA model for both demand occurrence and demand sizes of the ADAM model, some of the specific model types are either impossible or very difficult to estimate. For example, seasonality on intermittent data is not very well pronounced, so estimating the initial values of components of seasonal models (such as ETS(M,N,M)) is not a trivial task. In some cases, if we have several products in a group that exhibit the same seasonal patterns, we can aggregate them to the group level in order to get a better estimate of seasonal indices, and then use them on the lower level.
adam() function allows doing that via
Third, in some cases you might know, when specifically demand will happen (for example, kiwis stop growing in New Zealand from May till September, so the crop will go down around that time). In this case, you do not need a proper intermittent demand model, you just need to deal with demand sizes via ADAM ETS / ARIMA and provide zeroes and ones in the demand occurrence part for the variable \(o_t\). This can be done in
ot contains zeroes and ones for each observation.
Fourth, more specialised models, such as iETS, will produce positively biased estimates of the smoothing parameters, whatever the estimator is used (see explanation in Svetunkov and Boylan, 2019). This is caused by the assumption that the potential demand might change between the observed sales. In this situation the components would evolve over time slowly, while we would only see their values before the set of zeroes and afterwards, which will make the applied model catch up to the data, increasing the values of smoothing parameters. This also implies that such forecasting methods as Croston (Croston, 1972) and TSB (Teunter et al., 2011) would also result in positively biased estimates of parameters, if we assume that demand might change between the non-zero observations. Practically speaking, this means that the smoothing parameters will be higher than needed, implying more rapid changes in components and higher uncertainty in final forecasts.
Finally, summarising this chapter, intermittent demand forecasting is a difficult problem. Differences between various forecasting models and methods in this case might be insignificant, and it would be challenging to select the appropriate one. Furthermore, point forecasts on intermittent demand are difficult to grasp and make actionable (unless you are interested in lead time forecasts, to get an idea about the expected demand over a period of time). All of this means that intermittent demand should be avoided if possible. Yes, you can have fancy models for it, but do you need to? For example, do you need to look at daily demand on products if decisions are made on weekly basis (e.g. how many units of pasta should a supermarket order for the next week)? In many cases thinking about the problem carefully would allow avoiding intermittent demand, making life of analysts easier. But if it is not possible to do, then ADAM iETS and iARIMA models can be considered as possible solutions in some situations.