Manuel D. Rossetti, Edward A. Pohl, Hugh R. Medal, Vijith M. Varghese
University of Arkansas
This project presents an object-oriented software framework for intermittent demand forecasting and inventory analysis which structure allows easy implementation and integration of new data sources, forecasting techniques, and forecast metrics. This framework is implemented in Java within a user interface and is named the CID (CELDi Intermittent Demand) Forecaster.
The research plan includes two major components:
1) enhancing/redesigning the software and
implementing the design in Java, and
2) performing applied research to improve the forecasting of intermittent demand.
To redesign the software, we plan to examine the object-oriented model implied by the code and to re-factor the classes to better support user interaction, database usage, and plugging in new algorithms. A graphical user interface will be developed to allow the user
to connect to a database, apply each forecasting technique, invoke the forecast selection classifier, and to view summary statistics concerning the analysis. The will be done using standard object-oriented programming practices and prototyping. Every attempt will be made to use non-proprietary or open source software during the development. The applied research related to this project involves two issues. The first issue is to improve the categorization scheme for picking the best forecasting technique. We will continue to investigate the use of a vector of summary measures as well as the use of a statistical based classifier to pick the best forecasting technique. The basic idea is to train the classifier on a large portion of the intermittent demand space to pick the best forecasting technique based on the summary measures. When a new demand series requires forecasting, the user would preprocess the demand series to compute the summary measures. Then, using the classifier, the user can retrieve the recommended technique. The user can then use the technique to forecast the demand. We have already performed preliminary work on this idea and it is currently at least as good as our previous categorization scheme. The research component would require the testing of which summary measures are better than others for classification and the training/testing of the classifier’s ability to pick the best technique for real data sets. The second research issue is to investigate or test new intermittent forecasting techniques. Since the demand generator has already been proven to reliably reproduce intermittent demand series, we might be able to utilize it to also forecast intermittent demand. The idea here would be to utilize the generator as a “boot strap” forecaster. By pre-processing a demand series, we would estimate the parameters required for the generator. We then generate multiple samples of multiple periods of future demand. These samples would then be averaged to construct a forecast. In addition to this new forecasting technique, a number of intermittent demand techniques have recently been discussed in the literature that involve intermittent series with trend. These types of intermittent series occur because of increasing demand rates due to aging parts or due to slowing demand rates due to end of product life issues. The methodology to evaluate these forecasting techniques is relatively straightforward: apply the technique to simulated or real data and measure its error properties relative to other techniques. Of course, the development of the software will facilitate this analysis.
This software is useful for aiding managers in choosing a forecasting method and stocking policy for a particular intermittent demand item.
Possible extensions to this software include: integration of the open-source R statistical computing software to facilitate further plots and ARIMA models, the automated selection of forecast technique parameters, as well as continued development of the inventory analysis module. Another helpful extension to this software would be the selection of optimal or approximately optimal r and Q values for both the analytical and simulation model.