© 2000 John Petroff
|
|
© 2000 John Petroff |
Time-series analysis is a special but important type of nonlinear regression analysis. What is special is that the exogenous variable is the endogenous variable itself. In other words, what was previously called autocorrelation and looked upon as a disturbance that must be removed, is analyzed here as the source of an informative signal.
The method is derived from quality control. As long as variations in output are random and within acceptable limits, a machine is operating adequately. Time patterns that are totally random are called "white noise". Such white noise cannot be explain and cannot be controlled. As soon as some discernible pattern emerges, the time series is no longer white noise. For a machine, it indicates that future output is likely to deteriorate further to unacceptable levels, and that a corrective action is called for. Extension of time series analysis from physics to social sciences is primarily attributable to G.E.P. Box and G.M. Jenkins.
The technique is particularly applicable to finance because many - if not all - financial decisions are dependent on next period forecast, i.e. short run forecast. Think of purchases of inventory to meet next month sales, or advertising to boost next year sales. Predicting next day stock price is also an area of application, but that will be studied separately in the last section of this chapter. In commercial and social behavior, existing patterns do not change significantly in the short run. Thus, to identify a pattern in a time series is for most company data quite sufficient to make a good short run prediction.
This is also true in several major aggregate series maintained by the US Department of Labor and US Department of Commerce, as well as many national and international bodies handling economic statistics. See, for example, unemployment statistics that are seasonally adjusted:. The reported unemployment rate is not the actual proportion of individuals unemployed in a particular month, but an annual rate representative of current conditions and calculated by multiplying the actual unemployment rate by an index of seasonal variation for that particular month. For instance, it is well known that December has a higher employment level than January. Thus the seasonal index for December is higher than the one for January. If the actual January employment level in a given year drops less than the index, the January annual unemployment rate for that year will be reported to have increased. Several other series are also seasonally adjusted, such as new housing construction.
| Previous: Regressions |
|
Next: Box-Jenkins |