Predictive Analytics: 3 Things You Should Know

Many organizations make decisions based on historical data. In fact, Trend Analytics form the bedrock of all Business Intelligence tools. Businesses spend enormous resources in building Data Warehouses that house historical data. Executives routinely make decisions based on these reports. So, is it all good, sufficient and accurate?  Perhaps not.

There is always a caution about extrapolating past performance to predict future results. For example, Security and Exchange Commission’s Rule 2a-7 explicitly requires companies soliciting funds to include a caveat “Past performance does not guarantee future results.” So what exactly is wrong in using historical data to base our decisions?

In short, the circumstances and context matter.

Focusing exclusively “What” aspect of past performance gives you information. But rarely does a sales report tell you “why” the sales performance was such. A depressed sales performance may be because of lack of demand, but can also be because of unavailable on-hand inventory. Unless Inventory positions are correlated with Sales, it is very easy to draw obvious, but incorrect conclusions.  Focusing on “Why” will tell you the factors that led to those results. When the context is not factored in, it is easy to make incorrect decisions about the future. Consider the circumstances under which certain business performance was achieved:

• An individual fund manager is the cause of superlative market performance. Once this manager exited, the performance tanked, even as other funds performed better.
• A new product launched has cannibalized Sales of another product.
• A competitor has exited the business and your Product sales boosted.

In each of these cases, clearly, there is a causal relation. But if we ignore the cause and focus only on the results, it can lead to very incorrect assumptions and expensive mistakes.

How do you get around this?

Enter Machine Learning and Predictive Analytics. Imagine you can list down all the known causal factors that impact the results. A few examples of causal factors that would impact Sales are:

• Investment in Marketing
• New Product Launch
• Seasonality
• Known events such as a super bowl, Holiday spending,
• Geography
• Opening new stores

If you can capture all the data related to the causal factors as they happened, it would help in two different ways:

1. You will be able to draw more meaningful conclusions from your past history
2. You will have built a foundation for Predictive Analytics.

At a Customer- individual level, the causal factors take a different form. These are related to customer buying journey. A typical activity pattern on your website might look like below:

• Visited website
• Checked pricing pages
• Searched using a specific keyword
• Downloaded an App
• Put an item in the cart
• Registered for promotions
• Clicked a link

If you start tracking the causal factors, then you start establishing a pattern. In time, you will be able to detect the “Winning Behaviors.”

Why does Predictive Analytics need Machine Learning?

Let’s take an example, to understand why you need Machine to learn and predict. Consider the table below that lists a number of different activities that your prospective customers has engaged over past 2 weeks and the final results as to whether a particular visitor has eventually become a customer!

What do we see here?

• 4 people did Email registration, out of which 3 turned out to be Customers.
• 5 people did more than 10 web visits in past 2 weeks. 3 of these turned out to be Customers.
• 4 people did 5 more Facebook likes. 2 of these turned out to be Customers.

What about combining these factors?

• It appears, people who visited more than 10 times AND had more than 5 likes they became Customers.
• People who sent in at least 1 Email AND engaged in at least 1 customer service chat became Customers.

Now imagine, thousands of Customer records and dozens of such factors.  Identifying what factor has caused the prospective Customer to buy from you, is not humanly possible, without sophisticated pattern detection mechanism. In general, it is a combination of factors that matter. In the above example, what types of activities mattered for visitors to become Customers? Does the Website Visit matter more? Did Facebook page make a difference? We don’t know until we analyze a large enough sample of Customer data, along with all the predictors identified and the data captured in a similar fashion.

Machine Learning and Predictive Analytics help connect the dots, establish patterns and tell you the likelihood of a visitor becoming a customer.

That brings us to the 3 takeaways that you should keep in mind while embarking on Predictive Analytics.

3 Key Takeaways:

1. Historical performance may not be indicative of future performance, because the factors that led to historical performance may not be true anymore.
2. You need to identify and capture causal factors, to get started with Predictive Analytics.
3. Because you cannot capture ALL causal factors, not every prediction will turn out true.

Predicting likelihood of Customer conversion is only one of the many uses of Machine Learning and Predictive Analytics. Same principles apply to a variety of situations such as predicting the likelihood of equipment failure, predicting delinquency, customer churn so on and so forth.

In summary, look beyond and around the historical facts and figures. Historical data gives good sense of what has happened, but extrapolating is fraught with risks. Check out VectorScient’s Predictive Lead Segmentation to see how Machine Learning helps.

Thank you for reading the post. If you liked it, please share it with your network or leave a comment for us.