Forecasting Supply Chain Demand Using Machine Learning Algorithms - Distributed Artificial Intelligence, Agent Technology, and Collaborative Applications - page 335

providing a unique solution (as opposed to backpropagation neural networks, which may have multiple
local minima and, thus cannot guarantee that the global minimum error will be achieved by training).
A complexity parameter permits the adjustment of the level of error versus the model complexity, and
different “kernels”, such as the radial basis function (RBF) kernel, can be used to permit non-linear
mapping into the higher or lower dimensional space.
r esearch Methodology
Due to the inherent predictive power of universal approximators, it would seem that using ML-based
techniques could provide a simple but powerful solution for forecasting chaotic, noisy and distorted
customer demand at the end of a supply chain. Thus, in this work we set out to investigate whether
ML algorithms, in general, perform better than traditional forecasting techniques. We formulate three
hypotheses to probe our research question:
H1: Machine Learning-based techniques will have better average-performance than traditional tech-
niques for manufacturer's forecasting of distorted customer demand
H2: Machine Learning-based techniques will have better rank-performance than traditional techniques
for manufacturer's forecasting of distorted customer demand
H3: The best Machine Learning-based technique will out-perform the best traditional counter-part for
manufacturer's forecasting of distorted customer demand
To answer our research question, we conducted an experiment to compare the accuracy of ML
forecasting techniques with traditional forecasting techniques in the context of noisy supply chain
demand as seen by the manufacturer. In our study, the traditional techniques were represented by the
moving average, trend, exponential smoothing, and multiple linear regression. Additionally, based on
the M3-Competition (Makridakis & Hibon, 2000), we included the Theta method (Assimakopoulos
& Nikolpoulos, 2000). For completeness we also included the frequently used, ARMA (Auto-Regres-
sive Moving Average) model, sometimes also referred to as the Box-Jenkins model (Box, Jenkins, &
Reinsel, 1994). The ML- based forecasting techniques were limited to the 3 general classes mentioned
previously: artificial neural networks (ANN), recurrent neural networks (RNN) and support vector
machines (SVM).
Sample Size and Statistical Power
The following statistical power analysis is based on recommendations from Russell (2001) for determining
sample size. Determining the required sample size is very difficult, since the variance of a forecasting
technique on an unknown data set is difficult to estimate in advance. One can vaguely guess that the
deviation of a forecast would be the same as the average forecast error. Since an average forecast error
of 26% has been identified (Jain, 2004a), we will use this as the estimate average deviation. However,
since we are concerned only with the manufacturer's end of the supply chain, who by the nature of their
position in the supply chain experiences extremely noisy demand, this estimate is at the most conserva-
tive end of the spectrum.