Abstract
In spite of some methodological problems, we can get some result by applying principal component analysis to time-series data. By calculating the latent root and latent vector of the corelation matrix or of the variance-covariance matrix of variable vector X (Xi : Estimated costs classified by type of dwellings and industries), X can be transfered to a new set of variable U (Ui is a linear combination of X by coefficient β i which is the ith latent vector). We can analyse the transition of building activity pattern by U rather, in some sense, Keenly than by the raw data. The most part of the variation of multivariant X can be explained by a smaller number of variable U, and X can be aproximately reconstructed by a smaller number of variable U. By inner products of vector β in different periods, we can examine the continuity of variations. The direction of the first principal component U_1 is rather stable and closely concerned with the total level of building activities, including time-series trend, building trade cycle and seasonal variations. The second and the other components are comparatively unstable and modifies the pattern explained by the first component. The second expresses a very recent and sudden phenomenon and the third is rather continuous. The forth is the most connected with seasonal variations. Generally, the results are abstract and difficult to be explained in concrete terms of economics, but we can get many implications from them.
