- Principle #1: Begin with the Decision in Mind
- Principle #2: Be Transparent and Agile
- Principle #3: Be Predictive, Not Reactive
- Principle #4: Test, Learn, and Continuously Improve
Principle #3: Be Predictive, Not Reactive
In recent years, organizations have spent heavily on technology for managing and using data. Beginning with Database Management Systems and moving through Information Management, Data Quality and Data Integration to Reporting, and Dashboards, these investments are now mostly classified as Business Intelligence and Performance Management. These investments have taken data that was once hidden in transactional systems and made it accessible and usable by people making decisions in the organization.
These investments have been focused on analyzing the past and presenting this analysis to human users. They have relied, reasonably enough, on their human users to make extrapolations about the future. Users of these systems are making decisions based on this data, using what has happened in the past to guide how they will act in the future. Many of these systems can also bring users’ attention to changes in data quickly to prompt decision-making. The value of this investment in terms of improved human decision-making is clear.
These approaches will not work for Decision Management Systems. When a decision is being automated in a Decision Management System, there is no human to do the extrapolation. Passing only historical data into a Decision Management System would be like driving with only the rear view mirror—every decision being made would be based on out-of-date and backward-looking data. It fact it would be worse, as a human driver can make guesses as to what’s in front of her based on what she sees in a rear view mirror. She will be reasonably accurate too, unless the road is changing direction quickly. Systems are not that smart—without people to make extrapolations from data, Decision Management Systems need to be given those extrapolations explicitly. Without some view of the future and the likely impacts of different decision alternatives, a Decision Management System will fail to spot opportunities or threats in time to do anything about them.
Predicting likely future behavior is at the core of using predictions in Decision Management Systems. You need to predict individual customer behavior such as how likely they are to default on a loan or respond to a particular offer. You need to predict if their behavior will be negative or positive in response to each possible action you could take, predicting how much additional revenue a customer might generate for each possible action. You want to know how likely it is that a transaction represents risky or fraudulent behavior. Ultimately you want to be able to predict the best possible action to take based on everything you know by considering the likely future behavior of a whole group of customers.
Decision Management Systems require predictions. They must be given predictions in the context of which they can act instead of simply reacting to the data available at the time a decision is made. They need access to predictions that turn the inherent uncertainty about the future into a usable probability. They cannot be told, for instance, which claims are definitely fraudulent—this is uncertain. They can be given a model that predicts how likely it is that a specific claim is fraudulent.
There are three specific ways in which Decision Management Systems can be given predictions. They can be given models that predict risk or fraud, that predict opportunity, and that predict the impact of decisions. They can use these predictions to direct, guide, or push decision-making in the right direction.
Predict Risk or Fraud
Most repeatable decisions do not have a huge economic impact individually. Despite their limited scope, many do have a significant gap between good and bad decisions. The value of the decision varies significantly with how well they are made. This gap arises when there is a risk of a real loss if a decision is made poorly. For instance, a well-judged loan offer to someone who will pay it back as agreed might net a bank a few tens of dollars in profit. A poorly judged offer will result in the loss of the loan principal—perhaps thousands of dollars. This mismatch between upside and downside is characteristic of risk-based decisions. Similarly, a poorly made decision in detecting fraud can result in large sums being transferred to an imposter or large purchases being made using stolen credit cards.
When Decision Management Systems are being used to manage these decisions, it is essential that the decision-making be informed by an accurate assessment of the risks of the particular transaction or customer concerned. Such models might be focused on fraud, using analysis of patterns revealed in past fraudulent transactions to predict how likely it is that this transaction is also fraudulent. They might be focused on the likelihood of default, using a customer’s past payment history and the history of other customers like him to predict how likely it is that he will fail to make payments in a timely fashion.
Many techniques can be used to build such models from historical data, but all of them require knowledge of which historical transactions were “bad”—fraudulent or in default. These known cases are used to train a model to predict how similar a new transaction is to these “bads.” Once such a prediction exists, a Decision Management System can use it, treating those transactions or customers with particularly high, or particularly low, risk differently.
Many decisions do not involve an assessment of downside risk, but they still have some variability. Not driven entirely by compliance with regulations or policies, these decisions require an assessment of opportunity before an appropriate choice can be made. There is typically no absolute downside if a poor decision is made, simply a missed opportunity. When Decision Management Systems are being used to manage these opportunity-centric decisions, they will need to have some way to manage these tradeoffs.
These decisions are largely, though not exclusively, about how to treat customers. Deciding which offer to make to a customer or which ad to display to a visitor are examples of decisions where the “best” decision is one which makes the most of the opportunity to interact with the customer or visitor. Historical data can be used to predict how appealing a particular offer or product might be to a particular person or to a specific segment of customers. The value to the company of each offer, combined with the likelihood that a particular customer will accept it, can then be used to identify the most effective offer—to make the best decision.
When many such offers are being considered, it may be complex to identify the “best” offer. It may be difficult to manage the tradeoffs between the various decisions. In these circumstances, Decision Management Systems can take advantage of optimization technology that allows the tradeoffs to be explicitly defined, and then the “optimal” or best outcome can be selected mathematically.
Predict Impact of Decisions
Sometimes the effect of an action taken by a Decision Management System cannot be precisely determined. For instance, the value of a subscription for a mobile phone will vary with the use made of the phone. When an action is available for a decision and has this kind of uncertainty about its value, a further prediction is needed.
The likely impact of each action on the profitability, risk or retention of a customer can be predicted by analyzing the behavior of other similar customers who were treated the same way—for whom the same action was taken. The prediction of the likely impact of each action can be combined with predictions of risk and opportunity to improve the quality of decision-making in Decision Management Systems.