Introduction

Marketers are consumed with ‘big data’, struggling to make it useable and under pressure to use data in smart ways to boost results. The sheer volume and complexity of data puts pressure on marketers to discover more effective ways of identifying the right audiences for high volumes of campaigns. This is where predictive modeling plays a critical role in your campaigns. By developing models to identify the best targets for each campaign, marketers can streamline their selection process.

However there have been some challenges that hold marketers back from implementing predictive modeling into campaigns. Traditional modeling tools were complex and required a statistician to use them. And usually these statisticians weren’t either available within an organization, or were prioritized to work on ‘strategic’ models rather than assist with campaign selection. Even when the expertise was available, companies faced other issues such as “siloed” systems, making it difficult to take model results and use them in a campaign.

Thankfully, times are changing and new integrated, marketing friendly platforms combine analysis, modeling, campaign management and reporting so that any marketer can start to exploit the power of predictive modeling.

What Modeling Can You Do?

There are a large number of models that can be invaluable to a marketing or campaign manager. With a willingness to use models, test them in a logical way, and carefully review the results, any marketing team can start to see improvements in campaign results. The trap most marketing teams fall into is they stress about how good the models need to be, with some trying to generate the perfect model. Models are designed to improve what you would have done in business as usual. If you can build a model to select customers for your latest campaign that outperforms your manual selection rules, then it’s successful. There is no such thing as a perfect model; you simply need to create one that is better than you could do manually. It’s generally more effective to generate more models that are “good enough”, rather than focus too much on creating the single, perfect model. After all, as marketing teams are acutely aware, the environment is constantly changing with new channels, competitors, and products coming on-stream every day. A good model available at the right time is better than the perfect model too late.

Some typical uses of modeling in the campaign environment are:

• Who are the best customers to select for my latest campaign?
• How likely is a customer to respond to a particular offer?
• How likely is a customer to buy a particular product?
• How likely are they to use a particular channel?
• When are they likely to buy? Within a week, within a month, within a year?
• What is the next best product that I should offer a customer?
• How likely is it that this customer will lapse or stop buying?

Eight Building Blocks for Sucessful Modeling We believe there are a number of important factors that marketing organizations need to consider when utilizing predictive modeling in your organization. We have outlined our top eight over the following pages.

1. Determine Your Objectives

First off, be clear of the objectives for your modeling activity. Are you looking to build ‘good enough’ tactical models that deliver uplifts to campaigns and are available for most campaigns? Or, are you looking to build fewer more strategic models? Successful marketing organizations have a clear focus about how they plan to use models and have the right team and tools to deliver. Where we see some organizations struggle is when they have traditional modelers, who are not used to working in a high volume business where new models are required for campaign selections multiple times per day and week.

2. Get Your Data in Good Shape

It’s also important to ensure your data is in good shape. If you have a clear data model, which puts customers at the heart of it, then you are instantly in a better place. Companies often struggle when they have ad-hoc data joined together, rather than a clear and robust data model. By utilizing data from multiple sources, such as purchases, interactions across channels, and survey or demographic data, you have a richer dataset on which to create models. By aggregating this data to a customer or individual level, you also have more data to include in any model building. Let’s look at an example. If you look at individual purchase data and create an aggregate, you can better understand each customer. You also can determine how many times they have purchased, the value of their purchase, how many days since their last purchase and how long in between each purchase. These additional ‘fields’ of data allow you to broaden the data on which you can confidently build your models.

3. Introduce Marketing Friendly Modeling

While there are many modeling packages on the market, most need to be driven by Statisticians. However, when your campaign teams are armed with tools designed to be used specifically by them, you enable them to be more productive and effective. Ideally, you want to combine the ability to generate models quickly, but without feeling you need a statistical degree to build the models in the first place. That is why having just the core tools that you will use the most frequently is often the best approach. Too many tools often times presents the dilemma of which tool to use and this can delay your campaigns. Having tools that enable you to predict which customers are the likely to buy, lapse or respond are the most fundamental. As a campaign manager, you want to be able to build a model, see visually that it will work and be able to use it. Having a nice, clear visual view of your model effectiveness certainly helps.

4. Adopt the Template Approach

It helps to have a method where you can use a template of previous models, make some modifications and deploy the results. Being able to use a template helps you get models built faster and also give you more confidence in the results. Having a previous model template that shows how you could build a model to predict a likelihood to buy product x, would be a good starting point to build a model for product y. You can see what fields are used in the model for product x and use some of these or perhaps just change 1 or 2 for your product y model. If you happen to use an analyst to help you build the first model, you can be confident knowing that sensible fields were used in the creation of the model. Using a template approach, which can be managed through your user interface, also puts the power of model building firmly in your hands.

5. Visualize and Test the Outcomes

Once you’ve built your models, and perhaps seen a gains chart giving you a view of your model results, you have some idea that the model will work. However, to truly see who the model will select, you want to quickly be able to visualize those at both the top end and lower end of the model. Being able to create a banded decile field on your model score, showing 10% of customers in each decile, lets you quickly visualize the differences between your ‘best’ and ‘worst’ deciles. For instance, if you build a model to predict purchases of product x, then a quick check to cross tab these people against your model deciles gives you a comfortable view that the model can identify the right audience. By doing some quick visualization of data against your deciles, you will rapidly see if the model seems to make sense, and if you should test it in a campaign, or adjust and run a couple of other variations.

6. Link into Your Campaigns

Most campaign teams are generating high volumes of activity, with many campaigns being planned and executed each day. Therefore, it is essential that you can create a model and use it directly in your campaign selections. You don’t want the hassle of moving datasets from modeling systems onto your campaign system as this will slow you down. This is where an integrated platform that combines analysis, predictive modeling and campaign management makes it easy for you. As you score your customers, being able to drag and drop your high scorers onto your campaign makes you more productive, and means you will develop response lifting models on more of your campaigns. Where it is cumbersome to move data around, we see campaign teams using fewer models.

7. Test Your Segments

Just because a model seems to be able to be predictive doesn’t necessarily mean it will deliver. That is why it is vital to test your models. In the early days of testing your models, you will want to test your standard selections vs those using the model. If you are selling a skiing holiday, your standard selection might be those who’ve previously bought a skiing holiday. While you would probably still select this audience, you might also top up your campaign using the ‘best’ scorers from your model predicting likelihood to taking a skiing holiday. If you are cautious, you might want to initially select a small volume, but then build on it as you gain confidence. You might also want to test a less high scoring decile from your model as well, or depending on volumes, take a sample of records from each model decile. By adding each group into your campaign as a test group, you can track how each performs. As you measure the results of your models and understand their effectiveness on your customer base, then you can start changing the volumes of activity and use the model results to choose more of your campaign volume.

8. Measure Performance

Just because a model seems to be able to be predictive doesn’t necessarily mean it will deliver. That is why it is vital to test your models. In the early days of testing your models, you will want to test your standard selections vs those using the model. If you are selling a skiing holiday, your standard selection might be those who’ve previously bought a skiing holiday. While you would probably still select this audience, you might also top up your campaign using the ‘best’ scorers from your model predicting likelihood to taking a skiing holiday. If you are cautious, you might want to initially select a small volume, but then build on it as you gain confidence. You might also want to test a less high scoring decile from your model as well, or depending on volumes, take a sample of records from each model decile. By adding each group into your campaign as a test group, you can track how each performs.

Be Bold

We see many organizations that are reluctant to try and use predictive modeling in their campaigns. We understand that this unwillingness to try something new can be motivated by a number of factors, but we encourage you to be bold. Many of our clients who take the plunge, develop predictive models and test them in their campaigns really reap the rewards. Don’t look at it as time consuming, difficult, or that it might require specialist skills. With modern, marketing-friendly modeling tools, integrated with campaign management, it is easier than you think.