When predictive analytics is not enough

Predictive analytics is in the rise. The typical IoT project is currently shifting from data collection and processing into identifying patterns and providing insights into the future behavior. Examples include predicting the remaining time until the next required hardware maintenance or the expected hardware lifetime, resource demand forecasting, and the emergence of health risks based on past measurements.

Predictive analytics has been the next natural step following descriptive analytics: when you are aware of the current situation in detail, you are likely to ask where current trends are likely to lead to. Currently the typical IoT startup is innovating by adding forecasting features to their platform. And the IoT PaaS offerings are catching up quickly by including machine learning tools to their stack and integrating them with their data collection services. The AWS IoT + Kinesis + Machine Learning tool chain is probably the most well-known case in this respect.

Now consider the following scenario: a hardware monitoring company is collecting data through sensors attached to offshore wind turbines. The data includes everything that affects the lifespan of the turbine such as temperature, humidity, vibrations, etc. Using this data and an existing database of past failures, the service can provide insights into an upcoming failure. The utilities company owning the turbines receives an alert from the monitoring service that a specific installation is likely to fail in the next 24 hours. This form of alert is expected: predictive analytics are probabilistic in nature. No matter how useful this information is, it is incomplete and it is likely to lead the company’s shift engineer to despair. The information that the company does not get is how to delay or prevent the hardware failure from occurring and that’s prescriptive analytics.

Prescriptive analytics deals with the questions that naturally arise when predictive analytics is applied. That is, how to delay, prevent or amplify the predicted fact. Prescriptive analytics generates actions to optimally handle a future outcome by combining predictive models and context-specific rules. It is primarily an optimization and planning technique. Therefore it is of limited use when the available reactions to a prediction are well-known and limited. Prescriptive analytics shines in scenarios where the optimization parameters, the number, combination and sequence of actions is overwhelming.

In order to apply prescriptive analytics, 2 prerequisites need to be met:

  • The predictive model must be actionable, that is, the predicted outcome can be used as a basis for actions and that it can be altered by actions.
  • The effect of actions must be measurable to identify the optimal path of actions and allow the system to learn the relationship of actions and outcomes.

As the demand for actionable insights is picking up, it is likely that cloud and IoT platform providers will start integrating new decision-generating tools in the form of planning and optimization tools into their existing offerings. This will lower the entry point for companies developing reactive IoT services capable of taking smart decisions and adapting to changing conditions.