In our last article “From Insights to Actions: Why Your Company Must Upgrade From Dashboard Reporting to Prescriptive Analytics”, we defined prescriptive analytics as the data about not just what will happen in your company but how it could happen better if you did a, b, c or d. Prescriptive analytics provides insights as well as recommend actions you should take to optimize a process, campaign or service to the highest degree. Of particular importance to this article is the campaign process optimization using machine learning and big data and specifically using a technique that we briefly described in our last article: Uplift Modelling. As we earlier defined, uplift models seek to predict the incremental value attained in response to a treatment and in our case, in response to an outreach, sales or promotion campaign.
“The old adage in advertising is, ‘I know half of my advertising isn’t working; I just don’t know which half,’” Daniel Porter said. “With uplift modeling, you can identify which half is working and which half isn’t — or more specifically, what customers are most receptive to advertising and what customers aren’t.” Daniel Porter was the Director of Statistical Modelling for the Obama for America 2012 campaign. Actually, the work done by Daniel and his team is attributed for helping Obama win the elections over Mitt Romney and for this reason, I feel obligated to give a brief overview of what exactly happened before we can go ahead.
I know half of my advertising isn’t working; I just don’t know which half ~ most marketers.
In 2010, the Republicans took control of the House of Representatives and many predicted a loss for the then president Obama in the upcoming 2012 elections. Not just the pundits, but even the Obama Campaign Managers themselves predicted that with the exiting base of supporters, Obama would not be able to win against Mitt Romney. With the limited resources they had, how would they increase support for the president through ads, mailings, phone calls and other outreach methods? Daniel Porter and team were tasked with finding an answer to this question and what came up as the best solution was Uplift Modelling. The aim was to identify the persuadable voters ie voters who reacted positively to their outreach campaigns and ensure that every outreach campaign was guided only to these persuadables thereby ensuring that with the limited resources, they could realize the highest level of conversion. In October 2012, the model was put to use by the Obama Campaign and as Porter said, “they guided every door knock and every phone call in the final weeks of the campaign. They used the model to target persuadable voters through direct mail, social media advertising, Facebook messages and TV ads.” “The models enabled the campaign to reach out the voters at an individual level, based on what message they were most likely to be receptive to and what form of contact they were most likely to be persuaded by.” He added. As we all know, the team managed to persuade enough of the swing (persuadable) voters to help Obama win the elections.
Winning elections, Obama proved, is not just an art but also a science. A science that involves statistical modelling and understanding the electorate so as to present to them what they are most likely to positively respond to. Uplift modelling is not just limited to political campaigns, Daniel Porter noted. It can be used in a variety of marketing campaigns like identifying customers who might be persuaded to buy a product and that is where our article goes to next, with our usual focus on the retail industry.
With uplift modeling, you can identify which half is working and which half isn’t — or more specifically, what people are most receptive to outreach campaigns and what people aren’t.
Most of us, at one point, have received promotion messages from a retailer showcasing which products they have put on offers or trying to remind us to go and buy certain products. Not all of us are usually persuaded by such promotions despite the retailers spending hefty amounts to run the campaigns. How then can the retailers optimize their campaigns so as to get the maximum benefits from them? Treatment Analysis, one of the products on the Insense Data Technologies AI4Retail platform was built for this specific purpose and will help us answer this question going forward using a case study with one of the retailers that I will give a hypothetical name “JP Supermarket”.
JP Supermarket has 41,650 customers that it wants to send promotions to. Like everyone else, they have accepted the fact that only 50% of the customers will respond to the promotions and are willing to take the risk. Like fishermen in the ocean, they will cast their nets on a large area and hope that some fish get caught but they run a risk of not catching any fish at all despite the high costs of the promotion. The adage is wrong; in a promotion campaign targeting randomly selected group of people, chances of getting 50% conversion are close to none. As a matter of fact, there are high chances of offending some people thus making them unsubscribe from your service or leave your company for a competitor all together.
A deep analysis of sales and promotion campaigns reveal more losses than gains in the process and as such, retailers (and any other business for that matter) needs to understand how to optimize their campaigns so as to get the most optimal response. To better understand campaigns, it is important to bring on board some form of metrics to gauge the performance of the campaign and eventually form a basis for its optimization. The metrics we use at Insense Data Technologies for retail sector campaigns are Incremental Response Rate (IRR)and Net Incremental Revenue (NIR).
Incremental Response Rate measures how many more customers purchased the product with the promotion as compared to if they didn’t receive the promotion
Net Incremental Revenue depicts how much money is made (or lost) by sending out the promotion.
The aim of every retailer is to maximize the incremental response rate ie to get more customers to buy as well as to maximize the net incremental revenue ie get the highest possible revenue from the promotion. I will not go into the technical details of how we maximize IIR and NIR at our AI4Retail platform but I will mention a few details to put this into perspective. As we earlier mentioned, the target audience for any promotion can be categorized into: sure things, lost causes, sleeping dogs and persuadables. The sure things are people who will purchase whether you incentivize them or not, lost causes are people who will not purchase whether you incentivize them or not, the sleeping dogs are the people who will purchase when you don’t incentivize them but will not purchase when incentivized while the persuadables are the people who will purchase when incentivized but stop purchasing when not incentivized. This means that an amount spent incentivizing the sure things is a waste because they would still have purchased anyway, the amount spent on the lost causes is also lost because the incentives have no effect on them. Incentives spent on sleeping dogs is worse in that you not only lose the money spent but you also lose the revenue you would have earned had you not sent out the promotions. The only sane decision would thus be to target the persuadables so as to maximize your return on investment in the form of high IIR and NIR.
However, this has a drawback in that it discretizes customer behavior which might not always be the case. A better approach is thus to find a probabilistic measure of the response to a treatment by the customers. By this approach, the difference in a customer’s probability to buy when treated and when not treated forms the major basis for optimizing IIR and NIR and the difference is called lift. A negative lift denotes customers who fall in the sleeping dogs category, a lift value of zero, a rare occurrence considering we are dealing with probabilities, denotes the lost causes or the sure things. A positive lift value denotes the persuadables with the value of the lift denoting the “degree of persuadability.” The higher the lift, the more easily persuaded the customer is and this would encompass the sure things and the lost causes at the lower positive probability values with the “highly persuadables” at the high lift values end. This is a lot of technical explanations weighed down to the best of my capabilities; luckily for you, our Treatment Analysis tool on the AI4Retail platform does all the complexities and gives you only what you need: should you send promotions to the customer or not and by sending the promotions, what are the implications in terms of NIR and IIR?
The higher the lift, the more easily persuaded the customer is. A campaign should thus be targeted to the customers who exhibit higher lifts so as to achieve maximal/optimal returns on investement.
No other company realized the benefits better than our hypothetical JP Supermarkets. With a target customer base of 41, 650 people, a promotion to purchase goods averagely worth KES 678.00 would cost them KES 50 per customer. That’s a total cost of KES 2,082,500 to run the entire campaign and with the assumption of “50% will respond”, that’s a total a total revenue of KES 14.11 million. It looks like that is a wise decision to make but the truth is far from the assumption as data shows. From the data, by running this promotion, the retailer actually attains a net incremental revenue of -903,156 meaning it loses KES 903, 156.00 while attaining an IIR of 1%. Why is this so? A profile of the retailer’s customer base shows that when it comes to response to this type of campaign, only 27.5% of the customers are persuadables. 11.7% of the customers are sleeping dogs that should never have been targeted, 14.5% are sure things who would have bought even without the promotion while 46.5% are lost causes on whom money was blatantly lost.
A deeper delve into the data reveals an even more worrying trend. By the company spending a whopping KES 50.00 per customer on the promotions, there is no way they can gain from the promotion and hence the task is to lower their losses from the campaign. This is an accidental value realized in that the model can be used to advise the maximum amount to spend per customer on a promotion so as to realize a gain. We realized that an optimal compromise would be for them to target all customers with a lift above 0.17 in which case only 34.1% of the customers would be targeted among which 72.7% are persuadables, 24.9% are sure things and 2.42% are lost causes.
By targeting this specific audience as compared to targeting random customers or all the customers as initially planned, the company would increase their IIR by 99.49% as well as increase their NIR by 71.23% thus saving the company KES 643,302.00. On the AI4Retail platform, the retailor could even go ahead and try targeting customers with a lift value above 0.3 in which case, even though they would target a much smaller audience of 3.34% of all the customers, their IIR would improve by 111.11%, their NIR by 97.21% and they would save the company KES 877,998.00. The Treatment Analysis tool does not only show you these numbers, it also gives you the list of the specific customers so that you can directly take the recommended actions as that is our primary goal at Insense Data Technologies; to offer tools for prescriptive analytics and make it easy for you to move from insights to actions. In this case, we directly show you what would happen if you targeted customers with a minimum lift of your choice from the generated and discretized existing customer lifts.
Uplift modelling based targetted campaigns would help JP Supermarket to increase their IIR by 99.49% as well as increase their NIR by 71.23% thus saving the company KES 643,302.00.
Uplift modelling does not promise to help you achieve 100% conversion. As a matter of fact, actual conversion might still be quite low but obviously way higher than it would have been had the retailer not used the technique. For instance, only 1.15% of the 41,650 customers purchased with a random selection for treatment, a percentage that increased by 130.43% with a target group selected through uplift modelling and we have already seen the massive improvements in the other parameters like NIR and IIR.
In the old days (last decade and before), targeted outreach campaigns were conducted based on customer demographics and then later based on their interests. You would use Google and Facebook platforms to search for people who love chocolate and then target chocolate campaigns to them. In this decade going forward, we are doing things differently; we are not just looking at the demographics and interests, we dig deeper to understand the customer’s response behaviour. Outreach campaigns are no longer just an art but also a science. The art comes in when we use domain knowledge to understand exactly how to model the customer through a process we call feature engineering. After that, what is left is a scientific process; the system augments the limits of the human brain by analyzing the data to pull to the marketing manager insights and prescriptions of the specific actions to take. The final decision, however, still rests with the marketing manager — the decision maker who again, using his domain knowledge (art) decides on how to use the prescribed actions from the machine. Outreach campaigns are thus both an art and a science and without either part, the company is doomed to lose lots of money in their marketing campaigns.
In the days of old (which includes today for most companies), campaigns were purely based on demographics, interests and what the customer bought before. In the future (which includes today if you use IDT Platform), campaigns are based on the customer’s incremental response rate to a campaign.
Fortunately, you don’t have to worry anymore on how to move from the past to the future; you can contact Insense Data Technologies at firstname.lastname@example.org for your data consultancy work and for a better step-by-step guide on how they can help your business make use of their AI platform to improve not only your campaign outreaches but also lots of other data-related decision making processes in your company.