Supercharge your data driven approach with trust

Because their inner workings often are hard to grasp, and because outcomes can be manipulated, distrusting data models is a healthy, often even advisable, attitude. Therefore, especially in data-driven marketing, it is like Seth Godin once said, you have to “Earn trust, earn trust, earn trust. Then you can worry about the rest.”

Now building trust takes time and effort. As an online marketing manager you have to prove that using models drives performance. And you have to reassure clients if the model’s outcomes don’t match gut feelings. Putting in your best effort, while constantly being aware of the fact that trust takes years to build, seconds to break, and forever to repair.

So what do you do? 

Our approach is to systematically build trust into all elements that make up your analyses and predictions. Elements like data quality and choice of model. If you build trust into these first, you’ll find trust in the outcome later. 

Using the actual distrust and critical questions from your organisation as essential (!) input to sharpen data and the model, to ultimately improve trust.

Easier said than done? Don’t worry, it does not have to be all that hard. Please find some best practices and examples from our work to help you speed things up a bit.

1. Have you considered all drivers for your model?

One of the reasons that a model can be off may well be that you have not taken into account all the drivers and parameters that influence the prediction. 

Examples from our practice

• Make sure to take into account ALL touchpoints (interactions) with customers for a multi-touch attribution model. Like sales and service phone calls, or advertising leaflets in print. 
• Look at all possible event flows. Next to holiday periods, depending on your business, payday (around 25th every month) may well explain peaks (or drops) in sales.

2. Can you guarantee the quality of your data?

Data quality is a broad topic. The two aspects we have to deal with most are completeness and correctness of the data. Ask yourself: can we be sure that ALL data is collected? What about naming conventions, are they well implemented? Is the data correct? Is the data complete? Is this still the case after preprocessing?

Examples from our practice

• Are all relevant website interactions recorded (tagged) in Google Analytics?
• Are all campaign naming conventions applied consistently?

3. What is the theoretical foundation for matching of model type with its application?

Let’s say you want to predict the number of clicks to be expected given an ad’s average position on a SERP. This relation is likely to be nonlinear but continuous. Therefore it is not recommended to apply machine learning models that have stepped outcomes like tree-based models (since in this example it is not a path but a more straightforward line that you are looking for).

If you’re new to machine learning get ‘All Machine Learning Models Explained in 6 Minutes’ here.

Examples from our practice

• Returns on ad spend like clicks, sales or revenue usually are a continuous function with a certain limit, so returns diminish with cost.

4. Uncertainty quantified

When insights are presented as facts, they may be perceived as being facts. But predictions are not facts, they are always educated estimates. And estimates come with uncertainty. On the bright side uncertainty can be computed, allowing decisions makers to know which insights to rely on more heavily and which to use as indicators merely.

Examples from our practice

• If we produce a sales prediction, we always communicate uncertainties, with confidence intervals. For instance a confidence interval reflects the 95% chance that the actual result will lie within this interval.

That’s it?

That’s it. Consider it a headstart to a continuous endeavor. Keep it going. And with a bit of luck you can start worrying about the rest soon enough!

Need help?

We love answering questions and solving problems surrounding data quality and trust. Also feel free to contact our experienced specialists for sophisticated data collection tooling, within (any, and your) cloud storage solutions. Or let’s discuss how we can guide you through your data-driven transformation. We’re here for you.