|dc.description.abstract||This thesis was undertaken with the intention of applying forecasting with time series analysis, in a
manufacturing context. This involved two phases: the updating of existing forecasting techniques,
and the application of these techniques to a manufacturing firm.
The existing techniques, developed mainly by Brown in the 1960's, had to be adapted for computer
application, to allow fast and objective computation of forecasts. This required an investigation into
the derivation of each algebraic model, previously computed by hand, and translating those intuitive
steps into routine ones. Furthermore, the revision of each forecast in the light of new data had to be
dealt with mechanically.
As for the application, the data supplied by the client, a large South African manufacturing firm, did
not permit a successful application. This concerned both the manner in which the data were recorded
(inconsistent time intervals), and the volume of data readily accessible. This then led the thesis in an
unanticipated direction to overcome these difficulties. To do this objectively, it became necessary to
generate test data with known characteristics, then to study how many data were required to recover
Generating data required an investigation into random number generation, real data consisting of both
true changes as well as a percentage of random fluctuations. A random data series was, therefore,
added to the series with known characteristics. Such characteristics are unknown for genuine data,
such as those supplied by the client. Empirical experimentation with the generated data, led to the
determination of the number of data required to recover coefficients of various complexity. This
number was found to be contrary to the statements made by Brown on this topic, significantly more
data being required than was previously thought.
Finally, an attempt was made to select an appropriate model for the client's data, based on the
knowledge gained from investigating generated data.||en