What is predictive analytics?
In business, the data we collect and the information we work with mostly relates to events which have already happened. We have accounts of sales after they close, of expenses after the spending, and employee records concern the people we have already hired. But naturally enough, we want to know the future. How many sales will we have next year? And how many products should we have in stock to meet that demand?
To answer these questions about the future, we can look for patterns in our existing records about past events and project them forward. We call this process predictive analytics.
This forward-looking form of data analytics has many different applications.
- Manufacturers can analyze past failures and predict when to service equipment to prevent future breakdowns.
- Marketers can analyze their best accounts and create campaigns which target similar people in the hope of gaining new, high quality, potential customers.
- E-commerce sites which tell us people who bought this also bought that: We have all been tempted, have we not?
How does predictive analytics work?
At its best, predictive analytics can appear to be deeply insightful about future events. On the other hand, badly implemented systems can quickly lose our confidence. Nevertheless, all systems have some processes in common, and some common keys to success and failure. Here’s the basic process:
- We analyze our existing data to learn statistical patterns.
- From those patterns we create a set of rules — a model — which describes how to apply the patterns to new data.
- We then pass new data through the model and the rules make predictions about what may occur in the future.
For example, we may analyze existing customer data and find that younger people like products with more features, but older buyers are willing to pay a premium for products made with higher-quality materials. From these patterns we can apply rules to new customer responses as they register in our system. If younger, we may successfully offer them more features and if they are older, higher quality products. In this way we hope to optimize our sales.
We often repeat this process regularly, in order to keep the model up to date: It learns new patterns, so you’ll often hear this part of the predictive analytics process called machine learning.
Four important keys to successful predictive analytics
- Most importantly, good predictions rely on good data. If your current records are incomplete or inaccurate, you can’t really expect predictive analytics to make good projections. For example, do you have demographic data about your customers and if so, is it thorough and up to date?
- Good future outcomes rely on choosing the best predictive modeling techniques when looking for patterns. There’s a certain art to this, which forms part of the data scientist’s expertise. But today, predictive modeling uses automated machine learning which can run quite complex statistical modeling experimentally on its own to find the best practical results.
- Ambiguity is inevitable in predictions, and we need to learn to work with imperfect results. We cannot predict the future with certainty — especially when it comes to customer behavior. We need to understand the accuracy of our model and with how much confidence we can use its results. All this may sound challenging, but we do it all the time, for example, with the weather forecast, which is generally accurate enough to be useful, but rarely perfect.
- The predictions made should be actionable insights. That is to say, you should be able to do something useful with the prediction and also be able to test in the future if the prediction turned out to be accurate enough to be helpful.
What is new in predictive analytics?
Predictive analytics perhaps sounds very new. Not really. Some of the statistical techniques — Bayesian analysis and regression — have been around for over 200 years. Nevertheless, contemporary predictive analytics really took off with the development of digital computing from the 1950s when modern algorithms, including neural networks, started to be developed. In recent years, however, there have been very significant improvements, leading to both simpler everyday analytics and advanced artificial intelligence.
The drivers of these new developments are straightforward but powerful.
- We have more data than ever before and storing that data is cost-effective, especially on the cloud.
- We also have more complex data than ever before, with the ability to access not only structured records but images, sound files and documents.
- We have more computational power available to us, again often in the cloud, which means we can handle this scale and complexity.
- And finally, better software design takes advantages of all these developments to make building, testing, deploying and using predictive analytics simpler and more reliable than ever before.
Where is predictive analytics used?
With this new power and these new capabilities, predictive analytics can be found in an ever-growing range of use cases and industries. Here are some examples.
Financial services. Predicting stock prices and other financial indicators is an important practice. However, banks, mortgage lenders and credit card companies also want to identify fraudulent transactions, offer the best rates to their best customers and sell new financial products to new customers. In all these cases, predictive analytics proves its value.
Retail, other consumer-facing industries. Other consumer-facing businesses such as retail and telecoms use similar algorithms when handling customer relationships. They also want to know in advance if customers are potentially unhappy or likely to move to another carrier or another service — what they call churn analysis.
Airlines. Airline carriers predict how many seats they can fill — not always successfully. Remember what we said about managing ambiguity and inaccuracy.
Transport and logistics companies. Predictive analytics is used to optimize supply chains — again we have become familiar with the ambiguities there too.
Overall, predictive analytics have made modern businesses very efficient.
How to get started? Benefits of predictive analytics
With all this in mind, how can we get started with predictive analytics effectively?
I like to recommend three simple scenarios because they can be commonly applied in many different businesses and you likely already have the data you need. Also, the techniques involved are relatively simple. Finally, you can implement and test the results easily.
Time series analysis
A time series is simply any data which records a change in values over time. Think of daily sales, weekly invoices, your monthly costs or your annual budget. But you could also consider the operating temperature of some equipment or the number of visitors to a website.
How this can benefit your business. Now, imagine what you could do if, based on your existing data, you could start to predict. What will our sales be next month? What will our average invoice value be in the next year? Perhaps this equipment runs hotter than usual — dangerously so?
Almost every organization can benefit from these projections somewhere in their business. Modern algorithms to do this can be quite sophisticated, but you could choose even simpler methods. Moving averages are very basic — you can work with them in Excel — and still quite popular in stock market and commodity analysis. Students of statistics, engineering and marketing often learn algorithms such as ARIMA (AutoRegressive Integrated Moving Average) in college and the techniques available in many business intelligence tools. These algorithms can even adjust for seasonal changes.
You can have well informed, effective business conversations about current trends and future possibilities with even simple time-series analysis.
Remember our example of young people buying products with more features and older customers perhaps paying more for better quality? If we could draw a graph of customer age against the amount they spend we might see groups emerging on the page. Older, higher-paying customers, younger customers spending less, perhaps in the middle customers buying the most of all. And scattered around the edges, some outliers who do not fall into any group.
How this can benefit your business. Cluster analysis finds these patterns, of course in much more sophisticated ways than just described. You can then use these groups for targeted marketing or to help you design products with the maximum appeal across groups. You can even look at those outliers — why don’t they quite fit? Perhaps their spending is suspiciously high or disappointingly low. Salespeople want to know.
We all know this use case. At its best, we may not even notice that an analytical engine is nudging us toward some new purchase or behavior.
The most basic recommendation is the familiar customers who bought X also bought Y. Or they watched A and also watched B, so we recommend you try B.
Such systems can be built with very simple data — often just the ID of a customer, the ID of the product or service they chose and perhaps some dates about when these events happened. More sophisticated analysis includes all sorts of data about prices and genres and styles and so on. But you really can start simply.
Challenges to avoid. These systems go wrong if we do not take care to only use patterns and rules which have strong support. If just one customer bought A and B, that’s not much of a recommendation. We also need to be aware of what our customers have already bought in the past — nothing looks more inept than algorithms recommending something the consumer has already bought.
Recommendations need not be about sales or viewing habits. They can also be applied to maintenance of equipment and even to driving routes for navigation applications.
A final prediction
If you have never looked at predictive analytics before, these three quick wins may be just the way to get started. But I think I can be sure of one thing — no matter what your business interest or your business problem, predictive analysis is in your future.