Data Analytics

Study guides: Exam PA, Exam DISC-DA

Data Analytics is the process of examining data to draw conclusions that support decisions. It is commonly divided into four types that answer progressively harder questions: descriptive analytics asks what happened, diagnostic analytics why it happened, predictive analytics what is likely to happen, and prescriptive analytics what should be done. An insurer uses all four, from loss ratio reports to the models that price policies and route claims.

\bar{y}_g = \frac{1}{n_g}\sum_{i \in g} y_i

\hat{y} = \hat{f}(\mathbf{x})

a^* = \arg\max_{a}\, E\big[V(a, Y) \mid \mathbf{x}\big]

• Symbols: The first block is a descriptive summary, the average outcome \bar{y}_g over the n_g records in group g, such as a territory's claim frequency. The second is a predictive model \hat{f} that estimates an outcome Y from characteristics \mathbf{x}, and the third is a prescriptive rule that picks the action a, such as investigate or pay, with the highest expected value V(a, Y) given what the model predicts.
• Each type builds on the last: A diagnostic analysis starts from a descriptive result and breaks it down by segment, period or cause to find what drove it. Predictive analytics fits a model to past data, such as a classification tree or a linear regression, and prescriptive analytics combines a prediction with costs and constraints to recommend an action.
• Data comes first: Every type depends on data quality, because a summary or a model is only as reliable as the data behind it. External data can add characteristics that the insurer's own records do not hold.

Classifying Four Insurer Analyses Classify each analysis as descriptive, diagnostic, predictive or prescriptive.

1. A dashboard shows that personal auto claim counts rose 12% from last year.

2. An analyst finds that the rise comes from one region, where a hailstorm struck in June.

3. A model estimates each open claim's probability of being litigated.

4. A tool assigns open claims to adjusters so as to minimize expected cost, given each adjuster's workload.

Answer

1. Descriptive: It summarizes what happened without explaining it.
2. Diagnostic: It explains why the count rose by breaking the change down by region and event.
3. Predictive: It estimates an outcome that is not yet known for each claim.
4. Prescriptive: It turns predictions, costs and a constraint into a recommended action.

How it appears on the exam

• DISC-DA opens its first topic, Planning an Insurer Data Modeling Project, with harnessing the power of data, and its Basic Data Analysis topic covers analyzing data with classification trees and with linear functions. The exam consists of application-based multiple-choice questions, so a question can describe an insurer's task and ask which type of analytics, or which technique, fits it.

Exam traps

• Diagnostic is not predictive: Explaining why last quarter's loss ratio rose is diagnostic, even when the explanation uses a model. Predictive analytics estimates outcomes that are not yet known, such as next year's claims or whether an open claim will be litigated.
• Prescriptive needs more than a prediction: A fraud score is predictive, and deciding which claims to investigate also needs the cost of an investigation, the expected saving and the investigators' capacity. Calling the score itself prescriptive misses that decision step.