Chapter 31 Quiz

Twenty questions to check your grasp of data-driven underwriting: the data sources, pre-fill, real-time scoring, the automation frontier, and data quality. Answers are in the collapsed key at the bottom — try the whole set before opening it. All scenarios are constructed teaching examples.

Part 1 — Multiple choice (15)

1. The chapter argues that the data revolution changed which of the following about underwriting? - A. The fundamental decision the underwriter is paid to make - B. The inputs and the speed of the work, but not the underlying judgment - C. The need for a combined ratio - D. The definition of an insurable risk

2. "Pre-fill" (data enrichment) is best defined as: - A. A model that predicts future losses from historical data - B. The automatic population of a submission's fields from third-party data sources - C. A regulatory requirement to disclose all data sources to the applicant - D. The practice of binding a policy before the premium is paid

3. Which is a genuine benefit of pre-fill named in the chapter? - A. It eliminates the need for a combined ratio - B. It can make submissions more honest by filling fields from independent data rather than self-report - C. It guarantees the data is accurate and current - D. It removes the underwriter's accountability for the decision

4. The three pre-fill failure modes identified in the chapter are: - A. Overpricing, underpricing, and mispricing - B. Wrong match, stale data, and false precision - C. Fraud, misrepresentation, and concealment - D. Adverse selection, moral hazard, and morale hazard

5. "Automation bias on a pre-filled field" refers to: - A. A model that is biased against certain classes of risk - B. Trusting a number more because a machine supplied it than if a human had - C. The tendency of automated systems to refer too many risks - D. A regulatory bias in favor of manual underwriting

6. Real-time risk scoring creates a particular psychological hazard because the score: - A. Is always wrong - B. Arrives before the underwriter reads the file, coloring everything read afterward - C. Cannot be documented - D. Is illegal in most states

7. The chapter's correct frame for a real-time risk score is that it is: - A. The decision, which the underwriter should rubber-stamp - B. Irrelevant noise the underwriter should ignore - C. An input — one voice in the room — to be weighed against what it could not see - D. A legal requirement for every submission

8. Which data source is described as "exterior, point-in-time" — powerful but unable to see inside, see intent, or read time? - A. Public assessor records - B. Satellite and aerial imagery - C. IoT/telematics - D. A third-party aggregator feed

9. IoT and telematics data are described as the richest alternative data because they: - A. Are always complete and never raise privacy questions - B. Observe actual behavior and condition over time, rather than a snapshot or proxy - C. Are guaranteed accurate by the vendor - D. Replace the need for a loss run

10. Straight-through processing (STP) is most appropriate for a risk that is: - A. Large-limit and catastrophe-exposed - B. Novel, with no loss history - C. Simple, standard, high-volume, with clean and complete data - D. The subject of conflicting pre-fill fields

11. A good referral rule, in the chapter's words, is essentially: - A. A premium threshold above which everything refers - B. An honest statement of where the data and the model stop being trustworthy - C. A way to avoid automation entirely - D. A list of the carrier's most profitable classes

12. The chapter says automation can be more consistent than a human underwriter. This consistency is: - A. Always a danger - B. An advantage on simple risks but a danger on complex ones (where the risk differs from its class in ways fields miss) - C. Irrelevant to the combined ratio - D. A violation of FCRA

13. The "silent default" is dangerous because it: - A. Refers too many risks to humans - B. Fills a missing price-driving field with a guess (a class average, an optimistic assumption, or a zero), so the submission looks complete and nothing flags - C. Always overprices the risk - D. Is required by state regulators

14. "Garbage in, garbage out" is described as more dangerous in the data age because automation: - A. Creates bad data - B. Removes the human friction that used to catch errors, then acts on the bad data faster and at scale - C. Makes all data current - D. Eliminates the need for verification

15. When a carrier buys a data feed and wires it into automated pricing, the chapter says it has: - A. Transferred the risk of bad data to the vendor - B. Outsourced the data but kept the risk - C. Satisfied its FCRA obligations automatically - D. Eliminated the need for underwriters

Part 2 — Short answer (5)

16. In one or two sentences, explain why the chapter calls a pre-filled risk picture "a first draft written by a machine."

17. Harbor Steel's satellite roof flag agrees with its loss runs, its inspection, and the broker's note. Why is that agreement worth more than the roof read from any single source — and what is the one thing the quarter-old image still cannot settle?

18. Give one risk that straight-through processing should clearly bind and one that judgment must clearly own, and name the single feature that puts each on its side of the line.

19. A carrier automates small-commercial underwriting and reports a lower expense ratio in year one. Explain why you cannot yet call this a win, and name the number against which it must finally be judged.

20. State the chapter's compliance rule for alternative data in your own words, and explain why a data attribute that may proxy for a protected class is not "laundered clean" by being algorithmic.

Answer key (try the whole quiz first) **Multiple choice** 1. **B** — the data revolution changed the inputs and speed, not the underlying selection-and-pricing judgment (§31.1). 2. **B** — pre-fill auto-populates fields from third-party sources (§31.3). 3. **B** — independent data fills fields the applicant might have shaded, narrowing the information gap that adverse selection exploits (§31.3). 4. **B** — wrong match, stale data, false precision (§31.3). 5. **B** — over-trusting a number because a machine, rather than a person, supplied it (§31.3). 6. **B** — the score arriving first colors every fact read afterward (§31.4). 7. **C** — an input, one voice in the room, weighed against what the model could not see (§31.4). 8. **B** — aerial/satellite imagery: exterior, point-in-time (§31.2). 9. **B** — it observes actual behavior/condition over time, not a snapshot or proxy (§31.2). 10. **C** — simple, standard, high-volume, clean data (§31.5, §31.6). 11. **B** — a referral rule states where data and model stop being trustworthy (§31.5). 12. **B** — consistency helps on simple risks, harms on complex ones where context matters (§31.6). 13. **B** — a missing price-driving field filled with a guess, so nothing flags (§31.7). 14. **B** — automation removes the friction that caught errors and acts faster, at scale (§31.7). 15. **B** — outsourced the data but kept the risk (the losses, the mispriced book, the regulator) (§31.7). **Short answer (model responses)** 16. Pre-fill *assembles* the risk picture automatically and fast, but it carries the errors of its sources (wrong matches, stale data, false precision); like a first draft, it is useful and must be *read and checked* — corroborated on the fields that drive the price — not treated as finished (§31.1, §31.3). 17. Three independent sources, gathered for different purposes by parties with no stake in the submission, *agreeing* is stronger evidence than any one of them alone — corroboration raises confidence in the picture. But the aerial image is a quarter old, so it cannot confirm the roof was *not* already replaced; only the inspection settles that (§31.3, §31.7). 18. Bind: a clean personal-auto renewal with no changes — *single feature:* it is simple and standard, fully inside a well-modeled class. Own: a \$20M coastal commercial property with a loss flag — *single feature:* large-limit catastrophe exposure where one wrong bind is existential (and judgment-heavy context the fields miss) (§31.5, §31.6). 19. A lower expense ratio only tells half the story; if the machine bound risks badly, the *loss ratio* will rise and swamp the savings — and those losses surface two or three years later. The verdict belongs to the **combined ratio**, which captures both effects (§31.6; Chapter 3). 20. *If you could not use a fact when a human collected it, you cannot use it because a vendor collected it for you.* The line against unfair discrimination applies regardless of how the data arrived; an attribute that proxies for a protected class is still a proxy whether a human or an algorithm introduced it — being "alternative" or "algorithmic" does not launder it (§31.2; Chapter 35).