The modelling of firms inventory behaviour has been plagued by structural instability and despite a great deal of research effort over recent years little headway has yet been made in producing a structurally stable model of stock levels. Wallis et al. (I987) surveyed the main UK models of inventory behaviour and concluded that 'the tests of predictive failure are particularly powerful when conducted over periods in which data characteristics change, and over half the equations we consider are rejected on the basis of their predictive performance in the early i980s. This is a surprising result, since the poor performance of their predecessors over this period was a prime motivation for the research that has led to the current specifications. The similarity in turning points in the forecast errors, both for the aggregate stockbuilding equations and for the different categories of stocks, suggests the omission of some factor(s) common to all specifications' (Wallis et al., i987, p. 144). Their analysis was conducted using within-sample stability tests, and more recent data would undoubtedly find an even higher rejection rate. Accordingly, this paper seeks to identify important omitted components or misspecification in the existing models of inventory behaviour in response to the suggestion in the last sentence of the Wallis et al. quotation. We take as our point of departure a number of recent papers on company sector behaviour which have focused on intertemporal optimisation under rational expectations. The seminal work of Sargent (1978) on the demand for labour has been followed by a number of studies which seek to treat the firms expectations of future variables in an explicit way. This work includes Nickell (1 984) and Henry and Wren-Lewis (I 984) on the labour market and Hall et al. (I986) on the determination of stock levels. One common finding amongst these, and many other papers is that of a root in the dynamic process under consideration which is close to unity. This means that the long-run solution for the model is poorly determined and this finding is a cause for some concern. Recently, with the growth of the literature on cointegration (see Engle and Granger, 1987; Hall, I986), this finding takes on an even more serious aspect, since a symptom of non-cointegration is the lack of a well-defined long-run solution. So the finding of a near unit root may be indicative that earlier researchers were in fact working with sets of variables which failed to cointegrate, which would inevitably lead to 'spurious regression' problems in the sense of Granger and Newbold (I 974).