Search results

This paper aims to address the effective selection of customers for direct marketing campaigns. It introduces a new method to forecast campaign-related uplifts (also known as incremental response modeling or net scoring). By means of these uplifts, only the most responsive customers are targeted by a campaign. This paper also aims at calculating the financial impact of the new approach compared to the classical (gross) scoring methods.

First, gross and net scoring approaches to customer selection for direct marketing campaigns are compared. After that, it is shown how net scoring can be applied in practice with regard to different strategical objectives. Then, a new statistic for net scoring based on decision trees is developed. Finally, a business case based on real data from the financial sector is calculated to compare gross and net scoring approaches.

Whereas gross scoring focuses on customers with a high probability of purchase, regardless of being targeted by a campaign, net scoring identifies those customers who are most responsive to campaigns. A common scoring procedure – decision trees – can be enhanced by the new statistic to forecast those campaign-related uplifts. The business case shows that the selected scoring method has a relevant impact on economical indicators.

The contribution of net scoring to campaign effectiveness and efficiency is shown by the business case. Furthermore, this paper suggests a framework for customer selection, given strategical objectives, e.g. minimizing costs or maximizing (gross or lift)-added value, and presents a new statistic that can be applied to common scoring procedures.

Despite its lever on the effectiveness of marketing campaigns, only few contributions address net scores up to now. The new χ²-statistic is a straightforward approach to the enhancement of decision trees for net scoring. Furthermore, this paper is the first to the application of net scoring with regard to different strategical objectives.

This chapter introduces the KontoSensor, a digital service offered by Deutsche Bank since September 2018, as an example of data processing using predictive analytics. We present the motivation behind this digital service, the use cases and methods currently implemented, the way they have been created, and measures to increase the usage of the KontoSensor. With KontoSensor, Deutsche Bank offers a digital service to its clients to analyze their transactions on their current accounts using methods from predictive analytics and to inform them when irregularities are found. Twelve months after the start, 90,000 clients are already using this service and experiencing the results of data science firsthand.

This paper aims to focus on different approaches to variable pre-selection for building net score models (also known as uplift modelling or incremental response modelling). The application of these models supports the identification of customers whose response can be traced back to be an effect of the campaign under consideration.

First, a net scoring methodology based on decision trees is presented. Then, derived from research contributions on this subject and analytics performed on real data from the financial sector, different approaches of variable pre-selection are discussed and compared numerically.

Net-χ2 and net information value as well as the rank lift impact correlation for interval variables would be preferred when performing variable pre-selection for net score models. Simulations showed that the results were relatively stable with respect to the number of cross-validation samples.

Variable pre-selection is required since it reduces computational effort that comes along with the complexity of net score models and the availability of a large amount of potential predictors. Some pre-selection methods result in a set of predictors quite close to the application of net scores itself.

Despite its lever on the effectiveness of marketing campaigns, only few contributions address net scores up to now and yet fewer authors deal with variable pre-selection for those models. In this regard, this article is the first to develop and compare different approaches.