This article is the epilogue to our case study example about campaign and marketing analytics for an online retail company. You could read the previous parts at
In this part we will synthesise our learning from the previous parts and explore some utilities of response propensity & profit estimation models for business purposes. We will also talk about next best action or next best offer as one of the advanced use cases of customer propensity modelling.
Free Will & Next Best Action
“I am the master of my actions and destiny” – it’s a good feeling to think this way. We all want to be in control of our lives, and want to make independent decisions. However, the new evidence that is emerging from neuroscience contradicts this belief. These are early days but still the evidence is building fast. Benjamin Libet pioneered experiments in the 1970s to question free will. You can find a version of Libet’s experiment in this video.
Recently, far more advanced experiments were designed with the use of an advanced brain imaging technique called fMRI. In one of these experiments, participants were asked to make a binary (yes / no) choice (e.g. press a button or not). FMRI collects the data for pre-decision activities in the brain and the researcher attempts to correlate these activities to decision making. Based on the data from fMRI a multivariate model was developed to predict the choice participants will make. The predictive model, with high statistical significance, was able to correctly predict the choice that the participant made 7 seconds before they were consciously aware of his or her choice. Now this is totally freaky, a mathematical model could predict what you are going to do before you are aware of your own actions.
Let’s assume that these brain activities are the cause of the decision making process. Then we can think of another way to look at the above results. If I could ignite the right areas in your brain before you will make a decision, I could control your decisions. I guess smart marketers are using this trick for a long time through the usage of powerful messaging and advertisements. Luckily only one in a million ads have this kind of power over us.
On some level, propensity models we have discussed in this case study for marketing analytics were also trying to predict human actions. Predictive models may not have access to brain imaging data but they try to model customers’ behaviour and predict customers’ chances to buy a product. However, to make these models useful it is essential that they are deployed properly for business purposes. They need to create the right environment to facilitate decision makers. In order to create useful decision support systems you need to have good understanding of end-user / decision makers and their dilemma.
Decision Maker’s Dilemma and Next Best Action
Imagine you are a telemarketer with a list of 20 products to sell to every customer you call. You are overloaded with information and questions about the products to sell. The following is a small list of questions sales persons have:
- Should I sell product-1 which suits the customer’s need?
- Should I sell product-2 which I am best capable of selling?
- Should I sell product-3 which has the best incentives for me?
- Should I sell product-4 which the customer is most likely to purchase?
- Should I sell product-5 which is the hottest selling product?
Next best action is a decision support framework that helps answer these questions and zero in on one single product to generate the maximum business value. Propensity models are just one of the important links in this framework. To address decision makers’ dilemma the propensity models need to be used in a proper decision support environment. The idea is to create a system that facilitates the right decision making without confusion. This intelligent sales support system needs to be made part of all the sales activity including campaign management and direct to customer marketing that we have discussed in length in this case study. However, designing such a system requires a combination of:
Science, Art, and Creativity – Next Best Action
The job of data scientists is not to make fancy models but to facilitate the decision making process, and help business users. The biggest gratification for your job in data science is through seeing value that your model has generated for the business. You models will only generate value for businesses when they are deployed properly with the complete understanding of the end users. This is by far the most difficult part of a data scientist’s job – developing models in contrast is much simpler. Development of models requires predominantly the understanding of science and business. However, successful deployment of data science requires a combination of both science and art. Most importantly it requires creativity to provide right and succinct information to the decision maker to answer all their questions with simplicity.
Additionally, information technology and system design plays a crucial role in empowering end users either through traditional systems, mobility, or other technologies. Data scientists will play a crucial role in designing these systems because of their understanding of decision engines and the complete business story.
I must say it’s no fun for me to see the collapse of the idea of free will. It will create a deeper identity crisis for us humans in answering the question – ‘who am I?’. I guess in the end taking the focus away from oneself, and enjoying the grand creation is the only road to salvation. If enjoying the rain or a beautiful sunset can create the right ignition in my brain to make decisions while losing free will, I think I won’t mind it as much!
See you soon with a new article on ‘career in data science’.
Again thanks for this article. In my org, we are trying to find out an array of “next best product to offer” to the customer. We are planning to use Latent Markov Model for the same. Do you have similar analysis for this type of question or some material I can benefit from.
Markov model is a good place to start for next best product or offer modeling. Bayesian networks is another technique you may want to investigate to compare results for better performance. Also, if you are investigating several products try association or link analysis to reduce them to a meaningful number. All the best.
You are doing incredible job by sharing your knowledge to all in need to build their career in Data science.
Could you please share base data used in this case study to follow the insights you have provded in the case. Awaiting your response and if possible please share the data at below id :
Hi Roopam. Thanks for the brilliant article. Can you share some more details on Next Best offer specially wrt the Retail case study. I am trying to use this for another case study. Also, how important is to use this while telling impact of a campaign analytics/cross-up sell propensity model. Since, I am new to this area a bit, wanted to understand?
Its an awesome post. it would be great if you could share the data that was used in this case study to email@example.com, currently am looking for a job change , practising this one will definitely helpful, especially people like us