Looking at the Future through Analytics: Predictive vs. Prognostic

Pitfalls of Predictive Analytics

Predictive analytics is widely practiced as the primary way for business to look into the future. Marketing, retail, healthcare, supply chain management, and many other use cases show the power of prediction as a tool to shape future outcomes. Predictive analytics examines large amounts of historical data to quantify the probability that a thing of interest (consumer, customer, patient, supplier, etc.) will respond to a stimulus (marketing campaign, pricing, treatment, incentive, etc.) in a favorable or unfavorable way. The key here is large amounts of historical data. By examining that data, algorithms identify similar characteristics among the members of a population that responded in a particular way, then use these characteristics to predict future responses. A primary concept of predictive analytics is the use of history to look into the future, making predictions based on what might be thought of as a preponderance of evidence.


Predictive analytics is powerful because predictions are based on evidence
without the burden to understand why things happen.


Herein lies the strength, and potentially the pitfalls, of predictive analytics. It is powerful because predictions are based on evidence without the burden to understand why things happen. The time and complexities of cause-and-effect modeling are avoided. And herein lies the risk of predictive analytics. When predictions are based on evidence of past behaviors there is an underlying assumption that the future will not radically diverge from the past. But many believe that the post-pandemic era will be radically different from the past and the present both economically and socially. If true, then it is reasonable to expect that future behaviors —responses to stimuli—will be different than past behaviors. Evidence from the past will not reliably predict the future. And if that is true, then it is time to ask how reliable, how trustworthy, and how risk-burdened future predictive models will be.


In the post-pandemic era, evidence from the past
will not reliably predict the future.


Potential of Prognostic Analytics

Perhaps it is time to adopt a relatively unheralded analytics discipline known as prognostic analytics. I don’t suggest prognostic as a replacement for predictive, but as a complement—a way to mitigate the risks of unreliable predictive models. Prognostic analytics introduces the discipline of causal analysis to the processes of looking into the future. To understand the difference, look at a simple healthcare example where the problem is to understand which members of a population are at risk for a stroke.


Predictive analytics looks at historical data.
Prognostic analytics looks at cause and effect.

Prognosis is not practical without understanding why things happen.


Predictive analytics looks at historical data to identify patterns of common characteristics among individuals who have suffered strokes and those who have not. It then applies those patterns to the population of interest to classify and cluster by probability of stroke. If asked why a particular individual is at risk the answer is simply that the predictors indicate high probability of stroke. Prognostic analytics looks for causal indicators of strokes, such as extremely high blood pressure. The prediction from a prognostic algorithm is high probability of stroke, just as it is with the predictive model. But the underlying reasoning is distinctly different: the prognosis is that high blood pressure, if left untreated, will lead to a stroke or heart attack. With a prognostic approach we not only know who is at risk of stroke, but why they are at risk. And the prognosis suggests that the future outcome can be changed by taking action to treat the high blood pressure.

Understanding Cause and Effect

The hard part of prognostic analytics is causal analysis. Prognosis is not practical without an understanding, or at least a reasonable hypothesis, about why things happen. To that end, we can turn to the long-standing field of systems thinking for a practical and proven approach to causal modeling. The approach, known as causal loop diagramming, is built on two principles about cause and effect: (1) every outcome is the result of multiple influences, and (2) causal relationships include feedback loops so many causal chains are circular.

Figure 1. Hotel Chain CLD

I’ll quickly introduce a causal loop diagram (CLD) here, then go into greater depth of systems thinking and causal modeling in future articles. The diagram shown in figure 1 illustrates the cause-and-effect model for a hotel chain seeking to manage profitability.

The nature of cause and effect is apparent in this model—the simple reality that things influence other things. In this example, expense influences profit, revenue influences profit, number of room nights influences revenue, etc. Note that every node (text label) in the diagram is expressed as a quantitative thing—something that can be measured. Then look at the lines and arrows. Blue lines are annotated with a plus sign indicating same direction influence: when employee compensation goes up expense goes up, and when employee compensation goes down expense goes down. Red lines are annotated with a minus sign indicating opposite direction influence: when expense goes up profit goes down, and when expense goes down profit goes up.

The entire model is a view of hotel operations as a system—a collection of interacting parts. The collective influences among the parts are the reasons why things happen—the causal dynamics of the system. For data-informed decision making and business management we can examine the causal chains to find the levers that can be used to change outcomes—to increase profit, decrease expense, get more customer referrals, etc. From a performance management perspective, causal modeling is highly informative to see leading/lagging relationships and choose effective metrics and KPIs. In context of this article, I present causal modeling as a fundamental discipline that is needed to adopt prognostic analytics as a complement to predictive analytics.

This is a quick and incomplete view of systems dynamics and causal modeling. I’ve not explained the entire model here, but simply given a sense of how to read it. A well-developed causal model often looks quite simple and sometimes obvious (Doesn’t everyone know that profit is driven by revenue and expense?) but developing them can be challenging. I’ve only introduced the subject of causal modeling here. If it sparks your interest, watch for an upcoming series of articles connecting the practices of systems thinking, causal analysis, and analytics.

Dave Wells

Dave Wells is an advisory consultant, educator, and industry analyst dedicated to building meaningful connections throughout the path from data to business value. He works at the intersection of information...

More About Dave Wells