Five most important pitfalls in predictive analytics
March 9, 2016
Predictive analytics offer plenty of added value but there are also pitfalls on the way to success. Daan van Beek, teacher in the TIAS Master of Information Management program and data scientist Minne van der Sluis describe the five most important pitfalls.
Predictive analytics is a powerful instrument for numerous organizations. It helps them to create a competitive advantage and to make their processes run more efficiently.
For instance, insurance companies and credit card providers use it to track down fraud, the police use it to catch criminals, sometimes even before they have committed a crime, and car dealers use analytics to significantly increase a campaign response.
From our own experience, we are convinced that predictive analytics has plenty of added value. However, predicting is extremely difficult - especially when the future is concerned. Making automated predictions with Predictive Analytics is many times more difficult. There are a lot of pitfalls on the way to success, below are the five most important pitfalls:
1. Simply use all your data
Some people think that predictive analytics is a panacea. They expect the software to find a solution for every business problem, even problems for which it is not known whether they will cause a problem. They collect an enormous amount of data, install data mining tools and wait to see what turns up. In exceptional cases, they will find a rough diamond, but more often the software comes up with obvious, useless or unreal correlations. Before you start with data mining, you need to obtain a clear insight in the application and ask yourself which business problem you want to solve and which data you will probably need to be able to do that.
2. Working with uneducated data scientists
Predictive analytics is a multidisciplinary skill, requiring an in-depth insight into and knowledge of statistics, "data massage" and applying the correct visualizations. There are not many people on this globe who understand this skill and who are able to make a success of it. Our advice: if you have a challenging business case for which predictive analytics could help, hire the best data scientist that you can find. Support and treat these boys and girls well; they are absolutely worth their weight in gold.
3. Forgetting the user
As soon as you have been successful with your predictive algorithm (it works and the results are meaningful), then don't expect everyone to immediately be over-enthusiastic. This certainly applies to business users of predictive algorithms. Give them time to build up confidence and to attach credence to the results. Policemen who used to follow their gut feeling or arbitrary impulses for patrolling in the city, are now being guided by predictive software which tells them where and when they have the greatest chance of catching a thief. And we all know that predictive analytics is not always right every time! That just happens to be the nature of predictive models.
So, don't forget the user's perspective; take the time to improve the analysis process as a joint effort. Build, test and validate the system on a regular basis. That also means: organize feedback from users.
4. Consider predictive analytics as a stand-alone application
In the long term, predictive analyses cannot be successful as independent applications. The solid basis of a data warehouse infrastructure is needed, so that the data can be cleaned up and integrated properly. You really need to anchor predictive analytics in the daily business processes. And at the same time ensure that the predictions can be easily used in other parts of your business intelligence platform (reports, dashboards, analysis and visualization).
5. Using poor quality data
If poor quality data significantly decreases the added value of your regular reports, then poor data quality will completely negate the benefits of predictive applications. Poor data is really a great risk for your company as fuel for your prediction engine. This will result in policemen catching innocent civilians or going poorly armed to the house of a real "bad guy".
This is not open for discussion; there is no such thing as "nearly right". Decisions based on wrong data in reports and dashboards can have a significant impact; decisions based on wrong data in predictive software have an enormous impact on for instance your financial results, customer satisfaction and your organization's reputation. If you are determined to fail with predictive analytics, then you must certainly not bother to increase the quality of your data!
About the authors
Minne van der Sluis and Daan van Beek work respectively as data scientist and managing director at Passioned Group. Daan van Beek teaches in the TIAS Master of Information Management program.
Master of Information Management
Do you have the ambition to architect the future landscape of your company? The Master of Information Management offers you the expertise to align IT processes with your organization’s overall strategy.