A VALUER MAY be instructed to value a commercial property a number of months ahead viz an Estimated Realisation Price (ERP) or a forecast of valuation as defined by the Red Book. This requirement immediately suggests time series analysis whereby the trend pattern in past valuations is projected into the future. There is a theory by Takens from which it may be argued that multivariate influences on a time series are captured in the patters of past valuations and therefore these can be recognised and projected into the future. Our previous work has shown that projections ahead for five months can be achieved using a double time series of gilts and backtracked capital valuations and projecting them forward using neural networks which incorporate lags for both the capital valuations and the gilts. At that time it was not known whether the gilts would be necessary for the forward projection, or whether they would be implicit in the capital time series. Our new work, reported in this paper, shows not only that the single capital series does not seem to produce accurate forecasts of value and that it does not contain all the data necessary for forward projection as the Takens theory would suggest, but also that the addition of local property index data as well as the gilts enhances the prediction accuracy of the valuations. Thus property indices and gilts are necessary along with historic backtracked valuations to make accurate models which can give reasonably accurate forward predictions of value. Our work also obviates the need to predict the gilt series in advance of the capital valuations. Indeed an added argument for these techniques is that it is now practicable to project 12 months ahead instead of the original five with useful accuracy. The methodology used in this work was to refine the forecasting techniques using linear regression, and then when optimal results were obtained, to use artificial neural networks to pick up the underlying non-linearity in the data. Thus regression provided an additional benchmark for the assessment of the ability of neural networks to predict complex trends over time. An additional assessment of results has been introduced into this investigation. Prior assessments of prediction accuracy were made by comparison of the results with actual subsequent portfolio valuations and indeed that method is repeated in this work, but we also introduce here a novel method called forward tracking in which the technique used to calculate past valuations is also projected into the future. These results are then used to compare the network predictions. Such a technique has produced some interesting effects which are described in the conclusion to the work and which beg future investigation.

This item is no longer available.