What is Prescriptive Analytics?

Prescriptive analytics builds upon the three other types of data analytics which describe the present and make predictions about the future. It then uses heuristics, machine learning and rule-based systems to make specific recommendations based on data and probability-weighted projections.

Let’s look a bit deeper at the different processes and stages of human input for each. You can see below that only prescriptive analysis takes you all the way to a specific recommended next step you should take (“Action”). Certain situations require human intuition and judgment and in these cases, it can provide decision support rather than decision automation.

Ultimately, prescriptive analytics helps you make better decisions about what your next course of action should be. This can involve any aspect of your business, such as increasing revenue, reducing customer churn, preventing fraud, and increasing efficiency. Here are the key benefits in more detail:

• Make data-driven, not instinct-driven decisions. Through advanced algorithms and machine learning, a specific course of action is recommended based on a wide variety of factors including historical and current performance, available resources, and probability-weighted projections and scenarios. This lowers the chance for human bias or error.

• Simplify complex decisions. Prescriptive analysis simulates a variety of scenarios and provides the probability of different outcomes, both immediate to long term. This makes it much easier for you to not only understand the specific recommendation from the tool but also know the probability of a worst-case scenario and incorporate that into your plans.

• Focus on execution rather than making decisions. Your organization is likely flooded with data from a wide variety of sources. And the pace of business today requires you to move fast. The best prescriptive analytics tools first break down data silos to analyze
  an integrated data set and then provide instant, specific recommendations on your best course of action. This allows you to focus your effort on executing the plan.

The high-level prescriptive analytics workflow is similar to the traditional machine learning or AI workflow except that instead of leading to predictive analytics and what-if scenarios, it leads to recommended actions.



The details of your process will vary depending on your specific use case and type of data but below is a high-level overview to get you started.

1. Define the question. As with other data analytics or data science projects, your first step should be to clearly define the problem you’re trying to solve or which question you’d like to answer. This will inform your data requirements and allow your prescriptive model to
   generate an actionable output.

2. Integrate your data. Next you’ll gather the data you need and prepare your dataset. To help your model be the most accurate, you should bring in data representing every factor you can think of. To prepare data for machine learning (ML) projects such as this you’ll need to do the following:
   

   • Ensure your dataset is correctly labeled and formatted.

   • Avoid data leakage and training-serving skew.

   • Clean up any incomplete, missing, or inconsistent data.

   • Review the dataset after importing to ensure accuracy.

   
   You’ll be working with big data, possibly in real time, so you’ll need to find the right tools. As stated above, cloud data warehouses can now cost effectively bring the storage, power, and speed you need.

3. Develop your model. Now you’re ready to build, train, evaluate and deploy your prescriptive model. You can hire a data scientist to code one from scratch or you can leverage an AutoML tool to develop a custom ML model yourself as a citizen data scientist. Either way, this algorithm-based model will need to ingest a mix of structured data, unstructured data, and defined business rules. Analytic techniques used in your model may include simulation, graph analysis, heuristics, optimization, and game theory. You’ll have to tweak your model in
   iterations to get it right and you’ll definitely want to test your model multiple times using new data to see if the recommendations generated meet what you would expect.

4. Deploy your model. Once you’re confident in its performance, you can make your prescriptive model available for use. This may be a one-time project or as part of an on-going production process. For a one-time project, an asynchronous batch recommendation is probably most appropriate. If your model will be integral to a larger process in which other applications depend on fast predictions, a synchronous, real-time deployment is best. Your model should automatically adjust as new data is added over time. This will improve the accuracy of the recommendations.

5. Take action. Now you should review the recommendation, decide if it makes sense to you, and then take appropriate actions. Some situations require human intuition and judgment and in these cases, prescriptive analytics should be viewed as decision support rather than decision automation. Conversely, if your prescriptive model is integrated to a larger process, the downstream actions may happen automatically. Application automation triggering action is one of the top 10 BI and data trends this year.

Both predictive and prescriptive analytics involve the use of statistics and modeling to determine future performance based on current and historical data. But prescriptive analysis goes further by using a mix of machine learning, algorithms, and business rules to simulate a range of approaches to a given business problem. It then recommends an optimal course of action to mitigate risk or achieve a desired outcome. The below chart compares predictive vs prescriptive analytics:

• Expense. Avoid overkill. Prescriptive analysis requires expertise, data storage, and computational power. Not every organization, campaign, or situation requires it. In some cases, the cost and effort of setting it up may not be worth it.

• Certain situations require human intuition and judgment. Sometimes you’ll need to bring additional context such as knowing external market conditions or competitor activities which may not be included in the algorithm. In these cases, prescriptive analytics should be viewed as decision support rather than decision automation.

• Invalid inputs result in invalid outputs. Even the best models will produce poor recommendations if the data and/or assumptions built into your model are not accurate.

• Models take time to refine. Training and evaluating your model is a critical step in getting prescription analytics right. Even with AutoML tools which get you started with a default model, you should adjust parameters to better fit your specific use case and test evaluate your model using new data to see if the recommendations generated meet what you’d expect.