Subscribe to our mailing list

* indicates required

Market research, survey analysis and productivity.Subscribe Now

Implementing the 80:20 Rule with Predictive Modeling

80:20 rule, predictive modelingIt has lots of names.  The 80:20 rule.  The pareto principle.  HVC.  Power Laws.  Zipf laws.  Whatever we call it, the idea is that around 20% of our potential customers will account for 80% of our upside.  While the numbers can move around a bit, this basic idea applies in all aspects of customer management.  It applies among your customers: a small proportion of your customers will provide most of your profit.  It also applies among your prospects: a small proportion of the prospects have the potential to deliver much more profit than the rest.

If the 80:20 rule describes your market, how can you leverage it?  Fortunately, it is pretty straightforward!

Get some data and quantify how much you like your customers (or potential customers)

The hardest part of developing a strategy to take advantage of the 80:20 rule is to get some useful data.  You need some data that allows you to quantify how appealing you find your current or prospective customers.  Your customer database can be a good place to start.  Or, conduct a survey.  Or, if you are lucky, maybe you can find some relevant data collected by a government agency.

To be useful, the data needs to contain information that you can use to quantify how appealing you find each of the customers.  If you are using a customer database a great measure is how much profit you have made from each customer in the past 12 months.  If this is hard to compute, you can often get almost us much value by just looking at the revenue, or some combined measure of amount purchased and frequency of purchase. If you are using a survey, you can ask people questions about how much and how often they buy, or you can look at a relevant attitude (e.g., their level of interest in your new products). No matter how you end up getting the data, the end-point is that you need a single score for each of the customers, where you know that the higher the score of each customer, the more you are keen to retain (or acquire) them.

You also need data that can profile the customers, so you can work how to target the customers you find most valuable.  Things like age, gender and where they shop are useful for this purpose.

Build a predictive model

predictive model is fancy-speak for working out the key differences between the profiles of the more appealing customers when compared to the less appealing customers. The idea is easiest to understand with an example.

In this example a business services company created a database of 1,822 of its customers.  The company that undertook the study asked each of these customers to complete a short questionnaire and used this to estimate the total amount of profit each of these companies was providing to the industry (i.e., the profit provided both to the company doing the study and its competitors).  The company also bought lists of prospects which contained the number of employees and industry of each of the prospects (the actual predictive model had some other variables, but I’ve left them out of this post to keep it simple).  These lists were matched with the survey, creating a database containing the estimated profit as well as the industry and number of employees of everybody in the database.  The goal of the predictive model is to work out how the number of employees and industry relate to profit.

A predictive model, created using DataCracker, is shown below.  While called a “tree”, these models are really upside-down trees.  The “top” of the tree has a chart which shows the proportion of firms with different levels of profit.  We can see that most of the firms have profit between $0 and $10,000, but some provide profit of up to $50,000.  The average profit is shown as $5,305.10.

The first splits are by number of employees.  Prospects with 4 or fewer employees have an average profit of $1,834.70, while firms with 30 or more employees have profit of $16,798.80.

The firms with missing data (i.e., the ones where nothing is known about their number of employees), or, with number of employees from 5 to 34, have been split according to industry, with the firms in accommodation and the other firms listed after accommodation having average profit of $6,988.60, whereas the remaining firms have a much lower profit of $2,416.50.


There are lots of different software tools for creating predictive models.  DataCracker has used a variant of a technique known as a regression tree, which is often the best tool because it is easy for novices to use well.  In some situations there can be better tools, such as regression, but it is easy to make mistakes when using these other tools.

Making predictions

Once you have a predictive model the next stage is to use the predictive model to make predictions.

If you are a bricks and mortar retailer and you have found that old men wearing raincoats are your best prospects, the trick is to spot the targets as they enter the store and treat them appropriately (which no doubt they will prefer to the usual greeting with suspicion).

In the example above, the obvious implication is that the business services company should prioritize first the companies with 35 or more employees and those with from 5 to 30 employees in accommodation and the other identified industries.

Making them a good offer (but not too good)

Every Sunday my kids, my wife and I go out to breakfast.  We are regulars.  We were recognized as regulars, and this is a simple predictive model (i.e., “there’s that family we see every week”).  The café rewarded us for or loyalty with a discount card.  We now get 10% off every meal.  Dumb move!  We were always going to go every week, so all their loyalty program has done is erode their profits.

Once you have used a predictive model to work out which customers you want to focus your efforts upon, the trick is to do so in a way that does not erode your profits.  My local café, for example, would be better served to learn my kids’ names and make sure we get preferential seating.  It would cost them less and translate to real loyalty.

My cafe’s mistake is a common one.  When targeting your most high value customers you need to come up with something more clever than price.  The only reason that customers are worth focusing on is because you make more money on them, so giving them a price discount, such as my café does, completely undermines everything that you are trying to achieve. A much better example of this is airlines, who go out of their way to make their most frequent flyers feel super-special.

Have questions? Please post them in the comments below.


Image courtesy of Vichaya Kiatying-Angsulee at


  1. Thank you for the article. I’ve noticed the predictive model in your article classified customers into groups based on number of employees and industry to predict profitability. Just wondering how many categories the model can analyze up to and what is the best practice. Cheers

    • Hi Joanna,

      At one time, the algorithm can compare 10 predictor variables (of course, you can always re-run the model with another 10 next time). The predictive tree can show 5 levels of branches (i.e., picture a tree branch with another branch growing off, another branch growing off the smaller branch, 4 more times).

      For best practice on predictive modelling, please consult this article: