Victor S.Y. Lo, PhD A.I. and Data Science Center of Excellence Leader, Workplace Investing, Fidelity Investments
Three Types of Analytics
The ability to analyze data, draw business insights, predict future outcomes, and optimize decisions has become a competitive business advantage. Analytics fields such as Artificial Intelligence (AI), Machine Learning, Statistics, Data Science, Econometrics, and Operations Research are available to meet the challenges of delivering the value of Big Data to drive business successes, as introduced in Lo (2019). Analytics, also known as Data Analytics or Business Analytics, can be classified into the following three types (Figure 1):
1) Descriptive Analytics – reporting what happened, often including business reports, summary statistics, and data visualization (e.g., scatter plots, histograms, box plots, line charts). Multi-dimensional tables and statistical graphics are sometimes employed to provide deeper analysis of data. As an example, reporting the recent weather pattern is a form of descriptive analytics.
2) Predictive Analytics – predicting what will happen such as future customer behavior or probabilities of some events happening in the future. For example, weather forecasting is a form of predictive analytics that can be based on Statistical Modeling or Machine Learning.
3) Prescriptive Analytics – with some knowledge of the future from predictive analytics, you can evaluate potential alternative actions to guide your decision making and determine the best action. For example, if you know there is a high chance of snow impacting your commute tomorrow (from weather forecasting via predictive analytics), you may evaluate the risk and benefit (e.g., safety and productivity) of working from home versus going to the office(using experience or historical data) and then determine the best action by balancing risk and benefit.
Figure 1 Three Types of Analytics
The three types of analytics are closely related and dependent on each other, as described below.
From Descriptive to Predictive
Learning about what happened in the past (descriptive analytics) is often a prerequisite to predicting the future (predictive analytics).
While descriptive analytics such as a summary of past prices, past sales data, revenue and profitability provide essential reports to run your business, learning about the past enables us to identify the trend and pattern, understand what happened, and possibly why something happened
While descriptive analytics such as a summary of past prices, past sales data, revenue and profitability provide essential reports to run your business, learning about the past enables us to identify the trend and pattern, understand what happened, and possibly why something happened. Such knowledge allows data scientists to predict future outcomes (or a range of possible outcomes).As an illustration, suppose you are selling a special kind of juice in a local shop, you may gather historical data on prices and weekly sales and display it in a graph to understand the association between price and sales (descriptive analytics), see Figure 2. To learn about how sales may vary with price (the demand curve), you may fit a regression line as in Figure 2, which is: S =1867-234P, where S = weekly sales, and P = unit price, assuming no other factors are interfering with the relationship. Such a model can predict sales level at any price point (predictive analytics).
From Predictive to Prescriptive
Predictive analytics is typically required as an input to prescriptive analytics, as knowing something about the future (predictive analytics) can help us explore the benefits of alternative options which supports selection of the right action (prescriptive analytics). Evaluating the potential outcomes of options often requires cause-and-effect assessment, known as causal inference in the literature. Continuing with the example above, suppose you want to set the best price for your juice in order to maximize profit, which is a prescriptive analytics (decision-making) problem, you can use predictive analytics to first establish the relationship between sales and price using historical sales data. While the best method to understand the cause-and-effect relationship is through a randomized experiment (A/B testing), we assume all you have in this example is a few data points from some past weeks. The regression line in Figure 2 already provides estimates of sales at various prices which is a form of predictive analytics. The relationship implies that, for every dollar price decrease (increase), you expect to have 234 more (less) units to sell weekly. Since you aim at maximizing profit, you would determine the optimal P such that the profit function is maximized. Assuming the cost of making a cup of juice is $1, your profit function can be expressed as estimated sales multiplied by unit profit, which is f(P)= S(P-1)= (1867-234P)(P-1),as plotted in Figure 3. We can determine the most profitable price by reading off from the chart(or through calculus) which comes out to be about $4.5 per cup, resulting in a weekly profit of $2,849.
Figure 2 Weekly juice sales as a function of unit price
Figure 3 Profit as a function of unit price
This article introduces the three types of analytics (descriptive, predictive, prescriptive), and illustrates them with a simple price-setting example. The link between predictive and prescriptive analytics is established through the cause-and-effect assessment of alternatives. In practice, establishing any cause-and-effect relationship may not be as straightforward (due to potential outliers and confounders). More advanced analytics are often required to understand the causal effect and derive the optimal solution via prescriptive analytics. Further discussion of the data scientist toolbox is available in Lo (2020).