Information Technology Reference
In-Depth Information
we re-executed the Super Wide Support Vector Machine models for a window size of 40% and one of
60% to evaluate the impact of this choice. The Mean Absolute Errors for the Super Wide Support Vec-
tor Machine with parameter optimization via standard cross-validation for a historical window size of
40%, 50% and 60% are presented in Table 4.
From these results we find that for the chocolate manufacturer's dataset the error decreased when the
window size increased, however, for the toner cartridge dataset the error increased with window size
increase. The Statistics Canada manufacturing dataset showed mixed results with no trend. Therefore,
we did not see a trend that would indicate that a smaller or larger window size would have an impact
on the performance. Additionally, the performance difference between such large window size change
(10%) resulted in very small differences in performance and all of these performances remained better
than the identified next contender which was exponential smoothing. Thus, we found that the Super
Wide Support Vector Machine with parameter optimization via standard cross-validation was relatively
insensitive to the window size and that a window size of 50% seemed to be a good choice.
conclusion and discussion
The purpose of this work has been to investigate the applicability and benefits of machine learning
techniques in forecasting distorted demand signals with a high noise to pattern ratio in the context of
supply chains. Although there are several forecasting algorithms available to practitioners, there are
very few objective and reproducible guidelines regarding which method should be employed. In this
research, we have shown empirically that the best traditional method for a manufacturer is the automatic
exponential smoothing with the first value of the series as the initial value. We have also found that all
of the more advanced machine learning techniques have relatively poor performance as a result of the
noisy nature of data and the limited number of past time periods for any given product. None of the
ML techniques can reliably outperform the best traditional counterpart (exponential smoothing) when
learning and forecasting single time-series. Thus, they are not recommended as forecasting techniques
for noisy demand at the manufacturer's end of the supply chain.
However, one important finding concerns the usefulness of combining the data from multiple
products in what we called a Super Wide model in conjunction with a relatively new technique, the
support vector machine. The domain-specific empirical results show that this approach is superior to
the exponential smoothing. The error reduction found range from 3.11% to 10%, which can result in
large financial savings for a company depending on the cost related to inventory errors. This assumes
that the company is already using the best forecasting method available, or otherwise the performance
gains would be even greater. We feel confident with regards to the generalizability of our findings, since
the work used actual data from a large number of products from two North American manufacturers,
with the additional verification against Statistics Canada manufacturing survey. We also feel that as
the number of products added to the combined time series model (Super Wide approach) increases, the
performance will also probably increase further since there will be more data to learn from. As any
business decision, the use of the technology presented here should be based on a cost-benefit analysis
of the benefits of implementing such technology weighed against its associated costs. If this approach
is viable, in the long run it may be integrated into enterprise resource planning systems for automated
and interaction free forecasting.
Search WWH ::
Custom Search