December 7, 2018
Outbound sales can be a resource-heavy task, but can you perfect the art of cold calling with machine learning?
We used Mind Foundry technology and data hosted on Kaggle to build a machine learning model that can predict the success of car insurance cold calls. In this blog, we explain how.
Why cold calling with machine learning works
“Hello Luke, how are you doing today? I noticed you were up for car insurance renewal, and I was wondering whether you would have a couple of minutes to hear about this amazing offer I can give you to reward you for your loyalty?”
What would you do if you were Luke?
It’s actually very hard to tell, which is why cold calling has an extremely low success rate and requires thousands of calls to bring any value. Because of so many varying factors involved, they are not really pleasant for either party involved.
However, cold calling with machine learning insights might make it possible to understand the factors behind successful calls, better qualify prospects and tailor the content of calls in the future.
We used Mind Foundry to create a machine learning model that could predict the success of car insurance cold calling. The data was provided by the Technical University of Munich and is hosted on Kaggle.
The dataset was collected from a US bank that also offers car insurance and collects information on customers. The goal was therefore to build a model that can predict whether or not a customer will subscribe for car insurance.
A quick overview of this information is provided in the following table:
To understand the features driving conversions, we can try to visualise how the duration of cold calls impacts the outcome. First we need to create the feature through Mind Foundry’s data preparation page by adding a datediff operation.
We can then access the histogram view to eye ball the possible relationships.
Surprisingly, although the call duration does increase the chance of conversion (CarInsurance = 1), there doesn’t seem to be a clear relationship. The hour of the call doesn’t seem to affect the outcome either. Overall there do seem to be patterns in our dataset, so we can be hopeful Mind Foundry will find a machine learning model that performs well!
Our goal is to build a classification pipeline to predict whether a customer will purchase car insurance or not. However, we need to be able to use this model in production, which is why we are going to exclude the duration of the call and the end time stamp of the call, as we cannot anticipate this information before the call.
For the model training, Mind Foundry will perform a 10-fold cross validation and withhold a balanced 10% sample of the original data for final validation purposes. Mind Foundry will also use Optimize, our proprietary Bayesian Optimiser, to efficiently navigate the space of possible data science pipelines.
Once the runs are complete, we can view the performance of the model on the 10% hold out and are reassured to see that the classification accuracy is pretty good (69.9%) and the model health is good!
Interpreting the model
The feature relevance of our machine learning model indicates that the month the customer was last contacted and the outcome of the previous marketing campaign have the strongest impact on them purchasing car insurance. This is followed by the communication channel used and whether they are household insured or not.
A deeper look at the relative influence of the features allow us to understand how their values impact the forecasts.
It's apparent that customers last contacted on their mobiles in March, June or September and responded positively to previous marketing campaigns are most likely to take the cold call offer.
This shows that cold calling with machine learning models, such as the one trained by Mind Foundry, can successfully prioritise customers and thereby improve the success of marketing campaigns.
Unfortunately, this data set doesn’t provide enough insights to tailor the content of these calls, but we could imagine that analysing the transcripts might identify some interesting recommendations!