• No results found

.Initial avg Final avg

Test 2: FLS on demand allocation decision

6. Empirical study on real-life problem

6.1. Input data preparation

6.1.1. Demand nodes set

After we position the target service- Next Morning and Next Day EDS 435, we should allocate limited re- sources and efforts to the right customers. In this section, we illustrate how the target market-demand nodes set N- is defined quantitatively, from which most high-end nationwide trans-city EDS demand can generate.  Methods and steps

We consecutively employ a series of multivariate techniques, i.e. correlation analysis, principal component analysis (PCA) and cluster analysis (CA), to identify the target market. The pertinent methods and steps are illustrated in Fig.6-1.

435 Please refer to Tab.1-2.

Figure 6-1: Methods and steps for identifying the target market

(1) Indicator system

Market analysis by the project committee is based on information from expert interviews, official publications and questionnaires handed out to corporate and private clients. Indicator system (see Tab.6-1) was designed accordingly in order to distinguish the high-end market under the consideration of the availability of the sta- tistical data.

General indicators Subordinate indicators

GDP Contribution from secondary industry

Contribution from tertiary industry Export-oriented economy Total export and import volume

Foreign capital utilized

Freight volume

Highway freight volume Railway freight volume

Air freight volume by civil aviation

Industrial economy Industry output

High-tech enterprises output Commercial economy Number of wholesalers and retailers

Sales volume of consumer goods Living standard Disposable income of urban residents

Population size Urban population size

Related business volume Business volume of postal delivery service and EDS Indicator system

Correlation analysis

Principal component analysis

Input

Cluster analysis

Target market

Road traffic Highway mileage Table 6-1: Indicator system for identifying target market

(2) Correlation analysis

First of all correlation analysis is applied to check the co-linearity of variables with a correlation matrix table. The result of the correlation analysis by SAS shows that variables in the indicator system are highly correlat- ed.

(3) PCA

PCA is therefore adopted to eliminate the co-linearity. It replaces the original difficult-to-interpret and corre- lated variables with fewer conceptually meaningful and independent components or factors in order to simplify the evaluation, while retaining most of the information in original data. In our case two independent indicators (or principle components in statistical terminology) abstract more than 80% of the information from the indi- cator system and explain most of the difference among cities (see App.3). The two component scores generated by PCA are taken as clustering variables.

(4) CA

CA is the kern step in the procedure illustrated in Fig.6-1. Many researches and books have pointed out that CA is a useful statistical instrument to identify group with similar characteristics436.

CA, more specifically speaking, non-overlapping hierarchical method437, we use here is a widely adopted meth- od, which typically results in a dendrogram, i.e. a tree structure that represents the hierarchical relations among all objects being clustered. The dendrogram in Fig.6-2 is an example based on Ward’s method (one kind of agglomerative clustering) with the modified project data set438. The X-coordinate represents the joint- ed cluster and the Y-coordinate indicates the loss of homogeneity439. The dendrogram is organized bottom-up that the merger of every possible cluster pair is considered and first minimizes the increase of within-cluster variance when groups are merged and then continues until all cities are clustered in one group.

Actually, clusters themselves are not directly derived by the hierarchical methods. Researchers seeking a solu- tion with a certain number of clusters need to decide how to arrive at those clusters from the tree representa-

436 See Churchill/ Iacobucci (2007), p 351; Myers /Mullet (2003), p.15; Wedel/Kamakura (1999) p.39. Books include such as McDonald/ Dunbar (2004)

and Weinstein (2004).

437 See e.g Everitt et al. (2011), Sec1.1; Romesburg (2004), p.2;

438 Hierarchical structures can be basically derived from two types of algorithms: agglomerative and divisive methods. Agglomerative methods start

with single-subject clusters, and proceed by successively merging those clusters at each stage of the algorithm until one single group is obtained. Divi- sive methods start with all subjects in one single group, and successively separate each group into smaller groups until single-subject groups are ob- tained. The latter category of methods is less popular in applied segmentation research.

439 The semi-partial R-squared (SPR) measures the loss of homogeneity resulting from merging two clusters into a new one at each step. If the value is

small, it suggests that the cluster solution derived at this step results from merging two very homogeneous clusters. On the other hand, large values of SPR suggest that two heterogeneous clusters have been merged to the new cluster.

tion. In our case a break point clearly occurs when three groups are merged into two, indicating the fusion of relatively dissimilar clusters after this point. If component scores of the two principle components from PCA for every city are projected in Euclidean space (see Fig.6-3), all the cities are also intuitionally divided into three groups. Other criteria, such as identifiability, accessibility and stablility440 should also be considered for CA.

Figure 6-2: Process of cluster analysis Figure 6-3: Component scores in Euclidean space

Results

Every city in the database of Company A is regarded as a demand node. The methods and steps described above are adopted to divide all the more than 2000 cities into several groups. Analysis of variance (ANOVA) and hypothesis testing are executed on the result from CA. Finally, the first three groups that have the highest Component Scores are identified as target market (see Tab.6-2). Managers from Company A were quite satis- fied with the results. All the crucial cites listed by experts from the headquarter of Company A are included in the target market. Moreover, the size of the target market (281 cities) is quite ideal for a market development strategy. Therefore, the 281 cities in the first three groups are defined as demand nodes set N. Descriptive Statistics for Different Customer Groups, please refer to App.4.

440 Different criteria are mentioned by former studies. See e.g. Wedel/ Kamakura (1999), p4; Dibb (1999), p108; Tonks (2009), p343; Kotler/ Keller

(2009), p 64. We explain some in the following.

Substantial: The segments are large and profitable enough to serve.

Accessible: The segments can be effectively reached and served. That is, they can be characterized by observably different means. Differentiable: The segments can be distinguished conceptually and respond differently to different marketing-mix elements. Stable: Only segments that are stable over time can provide the necessary grounds for a successful marketing strategy. Familiar: To ensure management acceptance, the segments composition should be comprehensible.

Relevant: Segments should be relevant in respect of the company’s competencies and objectives.

Compactness: Segments exhibit a high degree of within-segment homogeneity and between-segment heterogeneity. Compatible: Segmentation results meet other managerial functions’ requirements.

Group No. Number of cities in group Representative cities

1 10 Beijing, Chengdu, Guangzhou, Shanghai, etc.

2 82 Anshan, Baise, Baotou, etc.

3 189 Bengbu, Shaoxing, etc.

Table 6-2: Target market defined with statistical analysis