3 Empirical analysis of submarkets
3.3 The specific research design
The brief presentation of the neural network technique above was intended to provide the background necessary to understand the analysis that is reported in this study. It is important to remember that, in the SOM-LVQ classification approach, labels are assigned to each category of observations according to underlying (relative) market characteristics. The label is for recognition pur- poses (e.g., a label might be a symbol for a particular area in which a particu- lar combination of characteristics is typical). Some segments may be based on criteria other than location. For example, a given area might be divided into building stock from two age categories.
It can be argued that the theory of neural network modelling is not actual- ly a theory at all, except in an open sense, with ‘theory’ referring only to the process of moving from empirical findings towards generalisation. Assuming that the proposed neural network classification approach is capable of con- firming segmentation according to a particular criterion, the identification of appropriate criteria for determining segments becomes a relevant issue. The key question concerns whether submarkets are determined primarily with economic or other criteria. Conclusions about possible similarities or differ- ences between the two contexts are based on the answers to this question. First, transaction-price data from metropolitan Helsinki are analysed with respect to each defined criterion (label) for submarket formation. The criteria for determining the submarkets are independent variables that explain prop- erty characteristics, location and other labelling criteria that are chosen in a
10 The results that are generated by non-parametric methods are never exact; the same applies for parametric methods when models have been specified incorrectly.
more flexible manner. After that the same procedure is applied to the data from Amsterdam.
A number of parameters must be considered when comparing two results that have been obtained with different datasets. These adjustments inevita- bly induce certain logical expectations for the robustness of the results. First, compared to the use of aggregated data, individual data is expected to gener- ate a ‘patchier’ feature map, thus providing a more powerful tool for identify- ing submarkets. Second, enlarging the dataset (possibly because of the pre- vious point: replacing mean values with individual observations) and conse- quently defining a larger map size generates a more detailed model, creat- ing a better possibility for identifying segments. Third, using panel data from multiple years instead of a single-year cross-section might generate models in which the time trend alone is too important to enable meaningful assess- ment of submarket classification or other, structural effects of local housing- market dynamics. The validation of all three points (i.e., problems related to the influence of data aggregation, small dataset and map, time trend), howev- er, requires the existence of actual segmentation.
The method for interpreting the outcome of the analysis can now be sum- marised into a few key stages. First, each observation is assigned a label, and each neuron in the feature map is defined as an n-dimensional codebook vector as a basis for the calibration of the feature map (e.g., Kohonen et al., 1996b). Classification accuracy statistics are then calculated for the success- ful ‘hits’ between observations and the feature map, according to this class. Repeating this procedure for each labelling solution can determine which factors contribute to high classification accuracy and are therefore relevant for the observed segmentation. Finally, a loosely formulated theory com- bined with additional knowledge of the local market context is necessary to guide the analyses. It may therefore be worthwhile to perform the analysis with similar types of input in two or more different contexts, with the goal of extracting a new, institutionally sensitive theory. As noted above, however, this would require controlling for the effects of aggregation, enlargement and the temporality of the dataset.
The approach to modelling spatial housing-market structure (submarket classification in particular) that has been elaborated in this chapter contains two general points:
1. All available data sources and previous studies may be used to obtain infor- mation on the diversified (multiple equilibria), fuzzy (based on images and bundles of intangible concepts) and truly nonlinear (e.g., not just log-linear) phenomenon of spatial housing-market dynamics (i.e., how different loca- tions and housing bundles within one urban area differ from each other, both quantitatively and qualitatively).
2. Because all locations are different by definition, location always has a resid- ual influence on the structure and dynamics of the housing market (regard-
less of how well supply and demand factors can be approximated). This idi- osyncratic element has not been fully utilised in the conventional intra- urban location-modelling literature. A bottom-up approach is consequent- ly more efficient than a top-down approach, and induction is a more val- id research strategy than deduction is. It is therefore useful to build up the theoretical framework by generalising from particular cases (instead of specifying an a priori theoretical model) and focusing on average market behaviour, as in the more commonplace approach.