What is prescriptive analytics?
Prescriptive analytics is a type of data analytics that provides guidance on what should happen next.
Prescriptive analytics is related to descriptive, diagnostic and predictive analytics. Descriptive analytics aims to provide insight into what has happened; diagnostic analytics identifies why it happened; and predictive analytics helps model and forecast what might happen. Given the known parameters, prescriptive analytics helps users determine the best solution or outcome among various possibilities.
At a high level, prescriptive analytics operationalizes insights from the other types of analytics through the use of business tools such as recommendation engines, dynamic pricing models, loan approval engines and machine repair scheduling. It is the cornerstone of business rules management systems.
Indeed, business users may be unaware they are using prescriptive analytics capability since they are likely more concerned about the outcome than the specific analytics technique. The field of prescriptive analytics is growing in popularity as its core techniques have become part of data science and machine learning workflows.
How does prescriptive analytics work?
Prescriptive analytics encompasses a continuum of capabilities. It could be as simple as making a yes/no decision, suggesting inventory levels, adjusting prices or automating stock trades based on market conditions.
In more complex scenarios, prescriptive analytics can also suggest decision options for taking advantage of a future opportunity or mitigating a future risk and illustrate the implications of each decision option. One goal in developing predictive analytics is to continually and automatically process new data to improve the accuracy of predictions and provide better decision options.
The most sophisticated prescriptive analytics models are enabled by stream processing engines, analyzing potential decisions, the interactions between decisions, the influences that bear upon these decisions and the impact of all the above on an outcome to ultimately prescribe an optimal course of action in real time.
Teams develop prescriptive analytics applications using standard data science development processes and tools. This starts with specifying requirements, identifying relevant data sources, organizing the data, developing the model and deploying it into production.
This last part, around deploying the model into production, is one of the most challenging and one in which prescriptive analytics differs from other types of analytics in a couple of ways. First, the prescriptive analytics engine often directly makes a decision rather than presenting an analysis or prediction, so the stakes may be higher if the decision is wrong. Second, prescriptive analytics operationalizes decisions, introducing more stringent latency requirements on data processing.
Prescriptive analytics takes advantage of structured, unstructured or mixed data. Data engineers and data scientists, either at an enterprise or working for vendors to develop prescriptive analytics products, need to consider the best types of data and the most appropriate way of structuring it. For example, a product recommendation engine may ingest all customer transactions into a graph database to identify products purchased by similar customers.
What are the benefits and challenges of prescriptive analytics?
Prescriptive analytics is not failproof. It is subject to the same distortions that can upend descriptive analytics and predictive analytics, including data limitations and unaccounted-for external forces. The effectiveness of prescriptive analytics also depends on how well the decision model captures the impact of the decisions being analyzed. Here are some of the top benefits and challenges of prescriptive analytics.
- It automates decision-making, reducing manual work.
- It speeds complex approval processes, enabling faster time to value.
- It enables faster response to changing market conditions, for example, automating stock trades faster than humans can.
- It improves resilience to fast-changing circumstances, helping enterprises, for example ride out supply chain disruptions.
- It operationalizes predictive analytics insights, increasing the value of existing analytics.
- The impact of bad decisions escalates faster with prescriptive analytics. An example is a stock market crash resulting from automated trading.
- Because of the stringent data engineering requirements of prescriptive analytics, some applications may not be feasible for using this type of analytics. For example, a checkout app using prescriptive analytics might make customers wait too long.
- Bias can be accidentally or intentionally baked into analytics models and is perpetuated by automation.
- Explainability, or the ability to explain how the model works, can be elusive in prescriptive analytics models, putting companies at risk of noncompliance with regulations such as GDPR.
- Prescriptive models require performance infrastructure and processes for ongoing monitoring and adjustment.
- Users must monitor models for unintended consequences. For example, a prescriptive analytics algorithm in a social media algorithm might promote content that is shared more widely. In some cases, viral content is driven by hate, so developers would need to consider ways to mitigate this unintended consequence.
Examples and uses of prescriptive analytics
Prescriptive analytics is the most sophisticated of the various types of data analytics and requires mature date processes to implement.
"It is the next frontier of analytics used in decision support to provide decision options, outcomes of each option and associated risk using techniques from descriptive and predictive modeling," said Terri Sage, chief technology officer at 1010data, a provider of analytical intelligence to the financial, retail and consumer markets.
Prescriptive analytics must be adaptive to allow for changes in economy, environments and dynamics of business. Some of the scientific disciplines associated with prescriptive analytics include machine learning and natural language processing and one or more of the other types of analytics.
For example, customer lifetime value is a calculation that uses descriptive analytics to calculate a value for each customer. The analytics allows companies to find other high-value customers based on their similarity to existing high value customers.
Predictive analytics can then be used by marketers and retailers to predict when customers will purchase again, or time until purchase, through various iterative modeling approaches. Prescriptive analytics then provides the decisions and impact by which to influence a customer's path and time to purchase.
Other examples of prescriptive analytics include the following:
- A ridesharing company introduces surge pricing to incentivize more drivers to action during times of peak demand.
- A maintenance scheduling engine optimizes machine repair in response to a predictive maintenance engine that estimates the cost of downtime, cost of failure, production loads, technician availability and maintenance status of other machines.
- A product recommendation engine suggests other products based on prior purchase history and the purchases of similar customers.
- A loan approval engine automates the approval process based on income, credit score and profession.
- A fraud management application decides whether to approve a transaction based on transaction history, location, amount and type of transaction.
- An AI-powered security management application looks for abnormal activity and blocks suspicious transactions.
Prescriptive analytics tools
Advancements in the speed of computing and the development of complex mathematical algorithms applied to the data sets have made prescriptive analysis possible. Specific techniques used in prescriptive analytics include optimization, simulation, game theory and decision-analysis methods.
Data science and machine learning tools form the foundation of a prescriptive analytics practice. These tools can help automate the process of creating the models required for prescriptive analytics. Enterprises may extend these capabilities to a broader base of users via new tools specifically designed for citizen analytics.
Data management and data processing can also improve prescriptive analytics at scale. Stream processing and stream analytics applications simplify data engineering tasks for ingesting data to support a particular recommendation or decision. Graph databases, data fabrics and data schema tools can simplify the data science task for mapping existing data to new use cases.
Business process tools can help to operationalize prescriptive analytics insights. Business rules management systems can provide the underlying frameworks for connecting multiple prescriptive models together. Intelligent process automation tools can connect the dots between automated decisions and robotic process automation tools.
The future of prescriptive analytics
The future of prescriptive analytics will be driven by the explosion in big data, new modeling approaches, and better artificial intelligence techniques. Hurdles to adoption will include new regulatory and popular pushback regarding privacy; risks associated with bias, discrimination and unintended consequences will also impact prescriptive analytics applications.
The rise in big data opens the potential to mine a far more extensive range and type of data to improve decision-making. For example, some firms are considering approaches to loan approval for people lacking traditional credit scores using data from social media and other third-party data sources. Enterprise data catalogs and new data wrangling techniques make it easier to discover and operationalize new patterns at scale.
Sustainability concerns are driving collaboration on taxonomies and best practices for measuring business decisions' environmental, social and governance impact. Vendors are starting to offer services to measure the supply chain's carbon footprint, relate satellite imagery to business decisions, and detect data leaks and privacy violations.
New modeling approaches are allowing teams to implement more accurate decisions. For example, physics engines can improve predictive and prescriptive maintenance by enhancing AI models. Digital twins can help enhance the context of complex decisions across multiple stakeholders. Improvements in AI natural language processing techniques can connect the dots between news reports and their likely impact on business operations.