casting model using the parts extraction history and
contextual information is introduced. Existing statis-
tical models along with the introduced model are ap-
plied on more than 1300 time series with intermittent
behavior representing GEHC spare parts harvesting
history. A modiﬁcation of the MAPE error is applied
to make it suitable for the intermittent behavior of
parts extractions. The application process of the most
ﬁtted model for each time series is then presented. Fi-
nally, the results of the designed method are discussed
and further research suggestions are provided.
2 LITERATURE REVIEW
Time series forecasting has been widely applied, not
only in academic circles, but also in different indus-
tries and businesses over the last decades. It aims to
collect, to analyze the past observations, and to de-
velop an appropriate model ﬁtted to the structure of
the series. Then, this model can be used to predict
future values (Ivanovski et al., 2018). However, the
application of the forecasting methods has been nar-
rowed to the spare parts clients’ demand mostly in
case of a one-way supply chain.
Some of the applied forecasting models are based
on a trivial logic of calculations like the Average, the
Na
¨
ıve and the Seasonal Na
¨
ıve forecasts. Other mod-
els, a bit more complex yet simple, can be more ad-
equate to the data such as the MA (Moving Average)
which has a property to reduce the noise or the varia-
tion in time series. One of the extensively used mod-
els in demand forecasting is SES (Simple Exponential
Smoothing). It is considered as a statistically simple
model as it cannot deduce trend in data. Nevertheless,
in many occasions, it has outperformed the MA and
robust models (Makridakis and Hibon, 2000).
Another used method is ARIMA which is widely
employed thanks to the Box-Jenkins methodology
that helps identify the optimum parameters (Box
et al., 2015). The limitation of this model is the as-
sumption that there is a linear behavior in the time
series. Thus, non-linear patterns cannot be captured
(Zhang, 2003).
When it comes to sporadic, extremely variable de-
mand, the above models can perform poorly. This cat-
egory of demand is difﬁcult to predict and needs more
sophisticated calculations. J.D. Croston found that
intermittent demand often produces inappropriate in-
ventory levels and that forecasts of constant quantities
at ﬁxed intervals can double the inventory level of the
needed volumes (Croston, 1972). Therefore, a fore-
casting method that helps overcome the problems pro-
duced by intermittent demands was introduced. Nev-
ertheless, the major limitation of Croston’s model is
updating the forecasts only after a positive demand
occurrence which makes the model incompatible with
obsolescence problems (Teunter et al., 2011). Con-
sequently, Teunter, Syntetos and Babai proposed the
TSB-Croston method which updates its periodicity
estimate even if the demand does not occur (Xu
et al., 2012). They used the ME (Mean Error) to com-
pare between TSB-Croston and statistical models like
SES, Croston, and SBA and found that TSB-croston
was the most accurate (Teunter et al., 2011).
Such comparative studies of forecasting methods
have been exhaustively conducted in the purpose of
accurately forecasting intermittent demand. To fore-
cast aircraft spare parts demand, a research study
considered twenty methods and concluded that the
best ones are the moving averages, EWMA (Expo-
nentially Weighted Moving Average), and Croston’s
method (Regattieri et al., 2005). In the same domain,
a study was carried out comparing artiﬁcial intelli-
gence methods like the NN (Neural Network) and the
ABC classiﬁcation method with Croston, TSB Cros-
ton, SBJ Croston, MA, and SES, deduced that NN
with a high number of features outperforms the rest
of the methods (Amirkolaii et al., 2017). In another
paper, the Holt-Winters method and sARIMA perfor-
mances were investigated on a sporadic demand of
spare parts with seasonality and trend components. A
similar performance of the Holt-Winters method and
the best sARIMA was observed on the seasonal de-
mand patterns. However, when a trend component is
also present, sARIMA gave a better accuracy (Gam-
berini et al., 2010).
In the literature, demand forecasting improve-
ments are not only conducted by comparing the fore-
cast methods on the same set of data, but also by ex-
ploiting ﬁeld information and judgmental adjustments
of statistical models. In this regard, a research pa-
per studied the effect of judgmental adjustments made
by forecasters on a commercial statistical forecasting
system used by a pharmaceutical company to forecast
an intermittent demand. The authors concluded that
judgmentally adjusted forecast is more accurate than
the generated forecast by statistical models (Syntetos
et al., 2009). Although the role of judgment in fore-
casting is recognized by researchers and the interest in
this domain is increasing (Lawrence et al., 2006), re-
search works integrating contextual information and
statistical models in the intermittent demand forecast-
ing area are still limited (Pinc¸e et al., 2021).
When applying a demand forecasting model, the
probable occurrence of demand is estimated. In case
of time series forecasting, statistical models that can
differ depending on the demand pattern are exploited.
Forecasting Extractions in a Closed Loop Supply Chain of Spare Parts: An Industrial Case Study
237