Academic journal article Journal of Supply Chain Management

The PMI, the T-Bill and Inventories: A Comparative Analysis of Neural Network and Regression Forecasts

Academic journal article Journal of Supply Chain Management

The PMI, the T-Bill and Inventories: A Comparative Analysis of Neural Network and Regression Forecasts

Article excerpt

INTRODUCTION

The current paper seeks to contribute to forecasting research by applying neural networks to predictions of the PMI, and to predictions of aggregate inventories. Because of its correlation to a number of important output measures, the PMI has a well-established reputation as an indicator of current economic conditions. The PMI and its component series are also used as short-term predictors for the likely direction of several economic variables. In this paper, a quantitative modeling structure is proposed to answer two of three basic operational questions regarding the PMI: first, can the PMI be systematically forecast? If the PMI can be consistently forecast, this effectively extends its analytical reach over time. Because the PMI is closely associated to the business cycle, foreknowledge of the PMI would help anticipate recessions and recoveries. This paper finds that changes in the PMI can be forecast. The results show that changes in the T-bill anticipate changes in the PMI by an average of 10 months. The leading nature of the T-bill is not new. As discussed below, several other researchers have found interest rate levels to lead aggregate output measures such as GDP.

The second operational question is what variables can the PMI in turn systematically fit and/or predict? These include GDP, industrial production, real inventories, real sales and the sales/inventory ratio, among others. Each of these variables should be analyzed and subject to regression and neural network methodology. However, due to space constraints only real inventories and the PMI are analyzed. The results show an 8-month lead by the PMI on changes in real manufacturing and trade inventories. Finally, there is a third operational question not currently answered, and that is whether inventories or other variables can be forecast with PMI forecasts--that is with forecasts generated for the PMI using the T-bill model mentioned above. While this exercise is important, it is beyond the scope of the present analysis. The critical questions posed in the current research are as follows: (1) Can the PMI be predicted? (2) Can the PMI predict? (3) Can predictions of other variables be made with PMI predictions? The first two questions are answered positively, while the third question is not addressed specifically. Most studies of the PMI have used correlation and linear regression as statistical tools. The current study builds on this tradition, but also incorporates neural network methodology to accommodate possible nonlinearity in the relationships. The existence of nonlinearities would point to a more complex association between the PMI and related economic variables than found previously.

BASICS OF NEURAL NETWORK ANALYSIS

While regression is well known, neural network modeling is not as widespread, and merits a brief introduction. Neural network models have varied business applications. These models have been employed in stock price prediction, stock selection strategy, currency exchange rate forecasting, bankruptcy prediction, logistics expenditures, lot-sizing, credit analysis, business cycle recognition and bond ratings (Cichocki and Unbehauen 1993; Altman, Marco and Varetto 1994; Refenes, Zapranis and Francis 1994; Cheng, Wagner and Lin 1996; Lotfi and Gaafar 2000; Bowersox, Calantone and Rodriguez 2003). Neural network modeling is similar to a mathematical portrayal of the functioning of a primitive brain. The system makes a first attempt at modeling a dependent variable by attaching weights to independent variables, very much like regression. Whereas regression stops at this stage, neural systems are used to interpret information from the error term generated in their first approach and attempt to repeatedly minimize the error term. To accomplish this task, the processing units send signals to the system over many weighted links. Values stored in the weights enable the system to learn and to generate data relationships where the structure may be linear or nonlinear. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.