Rough data analysis: aimed at solving classification problems by reducing redundant attributes in order to obtain the minimum subset of attributes ensuring a good approximation of classes and an acceptable quality of classification.
CHAID: the Chi-squared Automatic Interaction Detection algorithm which iteratively segments
a data set into mutually exclusive subgroups.
Logistic regression: a well-known and widely accepted statistical technique to fit a logistic
mathematical model into supplied data.
The Task - Predicting Lapsing in the next month
Models created by these techniques were used to gain insight into factors influencing customer behavior and to predict whether customers (investors) would end their relationship with the company in the next month. The comparison between the techniques was based on accumulating their accuracy values over percentiles of the validation data set. How It Was Done
The company provided a segment of their database containing information about more than 500.000 clients. Time series of registered behavior and static variables for all investors were extracted from the database. Chordiant Decision
Management automatically created predictive a scoring model to rank cases in terms of their likelihood of lapsing. The model was used to rank higher rank were those more likely to lapse, (i.e. the percentage of stoppers is highest in the first
percentile). The 20th cumulative value (horizontal axis in Figure 1) equals the percentage of stoppers in the percentiles 1 to 20. The 100th value equals 38.6 % being the percentage of stoppers in the entire validation data set.
The Outcome
Figure 1 shows that the model created by Chordiant Decision Management outperformed all the other models, especially in first percentiles, where accurate predictions matter the most. In these percentiles the customers are most at risk of lapsing. The Conclusions
Models created by these techniques were used to gain insight into factors influencing customer behavior and to predict whether customers (investors) would end their relationship with the company in the next month. The comparison between the techniques was based on accumulating their accuracy values over percentiles of the validation data set. How It Was Done
The company provided a segment of their database containing information about more than 500.000 clients. Time series of registered behavior and static variables for all investors were extracted from the database. Chordiant Decision
Management automatically created predictive a scoring model to rank cases in terms of their likelihood of lapsing. The model was used to rank higher rank were those more likely to lapse, (i.e. the percentage of stoppers is highest in the first
percentile). The 20th cumulative value (horizontal axis in Figure 1) equals the percentage of stoppers in the percentiles 1 to 20. The 100th value equals 38.6 % being the percentage of stoppers in the entire validation data set.
The OutcomeFigure 1 shows that the model created by Chordiant Decision Management outperformed all the other models, especially in first percentiles, where accurate predictions matter the most. In these percentiles the customers are most at risk of lapsing. The Conclusions
- Proprietary data was able to be used effectively to improve customer retention.Chordiant Decision Management outperformed other advanced analysis techniques in identifying those most at risk.Chordiant Decision Management enabled the investment bank to adapt its targeting policy and perform directed actions to retain its clients.
