Predictive Models of Data in Digital Advertising
Digital Ritesh
Using analytics in the organization, and/or expecting to invest heavily in the staff and IT is required to deliver high-quality predictive analytics, then it makes sense to have at least some appreciation of what a company is investing in and why it will bring benefit.
It's a myth to have a background in mathematics or statistics to be able to understand how a predictive model works, or how it can be used.
It's true that using predictive analytics to build a predictive model is a technical task, done by nerdy types who enjoy it, but understanding what predictive models are and how to use them does not require specialist training.
A typical credit scoring model, used by banks the world over, works by simply adding up the relevant points to get a score. The higher the score the more creditworthy someone is.
If someone starts talking about the predictive models that they could build for a company, before they have asked it what want to achieve, there is something strange.
When discussing predictive models, the starting point should always be some objective within the organization.
Predicative analytics then maybe the right tool to help deliver what is the objective. Models are used to predict all sorts of different things, but whether or not a predictive model is going to help to meet objectives boils down to just three things:
1. Will the model improve efficiency?
2. Will the model result in better decision-making?
3. Will the model enable to do something new that the company has not been able to do before?
Talking about efficiency is replacing a manually based decision-making process with an automated one. Sometimes this results in people being redeployed productively elsewhere, but more often than not efficiency means job losses or a devaluation of people's skills.
This is important because it means implementing predictive models for the first time, or they are deploying in a new area where they have not been used before, or a company will meet resistance and will need a strategy to deal with it.
With regard to the second point, the evidence from many different studies is that models created using predictive analytics make better predictions than their human counterparts, and in many situations, better predictions mean making more money.
However, having a model that can predict something with a high degree of accuracy it’s not enough. Perhaps the biggest mistake people make when developing predictive models is to deliver a model that is not then used for anything.
The predictions generated by the model are to do something to influence or control people's behavior, which in turn generates some benefit for them or for you. Identifying people who are likely to purchase something is fine, but then it needs to act on this information to increase sales.
This could be by encouraging existing customers to spend more (e.g. discount off their next purchase), or to use them as a conduit to attract new customers who would otherwise have spent their money elsewhere (e.g. two for one deal for the customer and a friend).
Let’s think about a model that predicts the likelihood of default on a loan. Knowing how likely someone is to default does not make to the company any money. The decision is whether or not to offer loans. Therefore use the model score as the basis of our decision about whether or not to give someone a loan.
Making the right decision on the basis of the score is far more important than the score itself. When thinking about predictive analytics, always have these three things in mind:
How will a predictive model improve operational efficiency?
How will it improve the decision-making process? And
what value does the model provide?
If it is impossible to answer these questions then maybe it's better to reconsider whether it's worth proceeding. Otherwise, the risk is wasting a lot of time and money creating something that might be very predictive but doesn't do anything to help to achieve goals.
The example of a score distribution shows a lender who has to decide whether or not to grant loans to customers who had applied for one. The score distribution is the key tool that underpins the use of all predictive models and is the basis for assessing how well a model performs.
To translate a word into a concrete example, let’s assume that a company had to set up a national online store, it sells selected quality wine also in-store, and it sees online selling as a low cost and low-risk strategy for reaching a much larger customer base than it currently has access to.
The setup and running costs will be low because the company uses its existing warehousing and IT systems, so all that remains to be done is to find the right customers to target and persuade them to buy. The biggest challenge is to identify which customers to target, and then develop an appropriate marketing campaign to attract them to buy online.
So let's start by thinking about what the business wants to achieve. Let’s assume that the board has laid out the following objectives:-
Objective 1. Recruit a good number of customers in the first year of operation. i.e. at least 25,000 buyers.
Objective 2. Make a profit during the first year. Therefore, the revenues generated from the product sales only need to cover the cost of the wine plus the marketing cost of acquiring new customers.
The board has given the marketing team a budget of INR1,2m to meet the objectives. Analysis of in-store transactions shows that the average profit on a single bottle case of wine is INR75. A direct marketing campaign costs INR2 for each person targeted.
A typical campaign includes texts, e-mails, mailshots, and voice messages, delivered over several weeks. The marketing department has access to a contact list, supplied by a database marketing company, containing details of 5 million people.
The list contains names and contact details, plus geo-demographic information such as income, occupation, age, car ownership, and so on, but the list does not contain information about the company's product buying behavior.
Using the budget provided by the board, the company could run a direct marketing campaign targeted at 600,000 people, selected at random from the contact list.
Another option would be to forget the list and go for a mass communication strategy, spending the recruitment budget on TV ads, sponsorship of websites, articles in magazines, and so on.
However, the Marketing Director knows that neither of these strategies will be successful because the products the company is selling are the domain of a very specific consumer segment.
In order to meet the business objective, the Marketing Director knows that they need a way to target just those people who like that product, and he believes predictive analytics may be the tool to help them do this.
So in this case, predictive analytics is going to be used as an efficient tool to reduce the costs associated with targeting and recruiting profitable new customers. So what does the marketing department do? If they want to use predictive analytics then they need some consumer data to analyze.
To construct a model using predictive analytics, two types of data are required:
• Predictor data. This is the information that is going to be used to make the prediction; i.e. the things that could feature in the model. For this problem, the predictor data is the geo-demographic information such as income, occupation, age, and so on, that has been supplied with the contact list.
• Behavioral data. This is information about the behavior we want to predict. For this example, the behavior is whether the customer buys wine or not. In technical terms, this is what a data scientist calls the "Dependent variable," "Target variable" or the "Modeling objective."
Predictive analytics is all about understanding the relationships between predictor and behavioral data. You can't do predictive analytics if one of these types of data is missing.
For behavioral data, you also need representative samples of each type of behavior so that the differences between behaviors can be analyzed.
For our case study, this means that information about people who did not buy the product is just as important as information about those that did.
The predictive analytics process then analyzes the data to identify how the predictor data can be used to differentiate between each behavior.
The marketing department has lots of predictor data that was supplied with the contact list, but it has no information about outcomes. Therefore the marketing team needs to obtain some before it can construct a predictive model.
To obtain data about wine-buying behavior, the marketing team undertakes a test campaign targeting 100,000 people selected at random from the contact list, for a total campaign cost of 200,000. As the test campaign progresses, the first sales come within just a few hours, Sales peak after about a week, followed by a gradual decline over several weeks.
At the end of the sixth week, the campaign winds up as new sales dropped to zero. Summarizing the key findings from the test campaign. Out of 100,000 people who were contacted, 1,600 responded by buying wine.
This is a response rate of 1.6%. Each contact cost INR2. Therefore the average spend required to secure one sale is INR125 (INR2*100,000/1,600). The information in supports the Marketing Director's belief that a random targeting strategy would be unsuccessful.
This is because, in order to generate the required number of sales specified for objective 1, it would be necessary to target 1,562,500 people at a cost of more than INR3m. However, with the budget, they can only afford to target 600,000. In fact, with a response rate of 1.6%, a random contact strategy can be expected to generate just 9,600 sales from 600,000 contacts.
We can also establish that a random contact strategy would be loss-making. This is because the average profit of INR75 per case won't cover the average costs to recruit each customer, so the marketing department would fail to meet the second objective as well.
However, the main purpose of the test campaign is not to generate sales, but to gather data about wine buying behavior. The marketing department now has some behavioral data to work with and sets about the task of 37 building a predictive model using the data it has acquired from the test campaign.
To build the model, the marketing team's data scientist takes all of the available information about the 100,000 people contacted as part of the trial and loads it into a statistical software package for analysis.
There are many specialist software packages that one can use to build predictive models, but the most popular ones are SAS, SPSS, and R. These days it's pretty rare to actually do any math yourself when building a predictive model, the software takes care of all the required calculation to generate a model using the appropriate mathematical and statistical technique.
The role of the data scientist is to decide what technique or parameters to use and explore the range of possible models, using their experience and expertise to derive the best model that they can. Obviously, you want the model to be as predictive as possible, but often there are business requirements and constraints that need to be taken into account.
It is very common to sacrifice a small amount of predictive accuracy to ensure that these business requirements are met. The model developers may be required to force certain variables to feature in the model or ensure that certain ones are excluded.
After a week of developing models, tweaking them, and exploring different options, the data scientist comes up with a type of model that is called the "Decision Tree."
With the decision tree, you can divide 100,000 targeted consumers, classifying them for gender, income level, family composition, marital status, age, and so on, so for example all the male people with less than 35 years addressed by the test campaign will be under the same class.
For each class, the data scientist will be able to calculate the response rate, the cost for response, and the profit for a response. The marketing team thinks will apply the decision tree to the remaining five million people on the contact list who have not yet been targeted.
To decide how to spend their remaining $1m budget, the marketing team calculate what kind of classes they can afford to mail and at how much cost. So they choose what is the cut-off class that the marketing team is going to use for this model.
Those scoring better than the cut-off class are targeted, those scoring below the cut-off are not. So the decision tree will help the marketing team to meet both of the objectives set by the board.
 |
| Decision tree |
Decisions were made using only the scores from the models, no humans are involved in determining what the outcomes should be. Each and every decision about a customer is made automatically on the basis of the score alone.
However, in many practical applications, it is common to use business rules to override score-based decisions, in order to meet strategic objectives beyond the scope of the model or to ensure that certain actions are or are not taken for some individuals, regardless of the score that they receive. One reason for override rules is legislation. There are laws that require you to treat certain people in different ways.
Also Read,
Digital marketing vs traditional marketing
Digital advertising, a continuous disruption
0 Comments