If only you could easily identify customers that are ripe for the picking to make the channel shift to digital. You could then enhance service levels, improve customer loyalty, save time and costs, and more. Within this context, I’d like to share a recent example of our collaboration with a financial services company to predict the likelihood of customers moving from the traditional, agent-based platform to a new direct, digital channel using machine learning.
The customer’s objective was to provide its field marketing team with an accurate list of customers that they could target and encourage to move to the digital channel – achieve huge savings in time and costs, as a result.
We decided to look at machine learning to solve this challenge. Machine learning is part of the artificial intelligence platform capability that helps solve difficult business problems. The aim was to build a sustainable model which learns on historic data patterns and improves accuracy following every iteration of field testing.
Honing in on Historical Data
We reviewed sample historical data of around 200K customers. The parameters that we reviewed covered attributes such as age, location, value of loan, loan period, customer income, employment status, etc.
We discovered that just around four percent of customers had converted from agent-based, traditional offerings to digital products. We used this data to build the machine learning model based on the following:
- Applying existing business rules to cleanse and transform data for enrichment.
- Identifying the best features influencing channel shift with exploratory analysis and statistical hypothesis testing.
- Splitting the dataset for training, testing and validation. The training dataset was then used to train a supervised machine learning model.
- Supervised learning was applied using different algorithms such as SVM, Logistic Regression, XGBoost and Neural Networks to train and evaluate the machine learning model.
- Evaluation results from the machine learning models aided selection of the best model.
Figure 1: High level overview of Channel Shift Analytics
The developed model was tested on an ‘unseen’ dataset. Around 30 percent of the total customers in this dataset were identified as potential customers who could be targeted for digital products.
This approach maximises the chances of a customer buying the new product and minimises the efforts and cost required to do so. The developed model had an overall accuracy of 70% and a true positive rate (or recall) of 77%.
The improvement over the current system is expected to be at least five times. This means that the customer conversion rate will improve from one customer converting after a 100 calls to five customers converting for every100 calls.
Tools and Technologies
Python programming language was used to clean and explore the data and also for creating and evaluating the machine learning model. NumPy, SciPy, Matplotlib, Seaborn & SciKit-Learn were the Python libraries used to achieve the above tasks.
Artificial intelligence can solve pressing business problems such as predicting the digital shift. We used the machine learning platform to help our customer’s identify a customer set that their marketing team could target to move to the digital channel. Thereby helping them provide immersive customer experiences and a range of financial product offerings, which would help their marketing team achieve customer retention and acquisition objectives.