Why Numeric Targets Need Exact Definitions
A numeric prediction can be expressed in dollars, days, counts, or another measurable quantity. But a number without a precise definition is ambiguous. “Predict claim cost” does not say whether cost means amounts paid or amounts incurred, which dates determine inclusion, or whether a policy with no payments has a value of zero. Different reasonable choices can produce different target values for the same record.
To make a target usable, define the measured quantity, its unit, the population or observational unit to which it applies, and the measurement period. The definition should also explain how to represent zero and how to treat incomplete information. These choices establish what the prediction is intended to estimate; they do not determine which predictive model to use. Precise target definitions are aligned with the SOA Exam PA (Predictive Analytics) syllabus.
Specify the Quantity and Its Unit
Begin by naming the quantity rather than using a broad label. For example, “total claim payments posted for a policy during the 30 days after the decision date” is more precise than “claim cost.” The first statement identifies a payment-based quantity and a window. It does not include amounts merely because a claim was reported or incurred during that window.
Units should be explicit and consistent. A total might be measured in dollars per policy over a defined period; a duration might be measured in elapsed calendar days; and an average daily amount might be measured in dollars per day. “In thousands of dollars” is also a valid scale, but it changes how values should be interpreted: a recorded value of 12 then means $12,000, not $12.
For a target defined as a total, the unit belongs to the measured amount. If a rate is intended, state both the numerator and denominator. For example, “total dollars paid per policy-day” is different from “total dollars paid per policy during a 30-day period.” The latter is a period total, not a daily rate.
Zero Is a Value; Unavailable Is Not
For a target such as total payments, zero is appropriate when the period is fully observed and no qualifying payment occurred. It is a meaningful outcome: the amount paid was $0. It is not appropriate to replace an unknown or unobserved amount with zero. If records are incomplete, the value may be unavailable rather than genuinely zero.
Similarly, a duration is generally not zero just because its end event has not yet been observed. If the target is elapsed time from report to settlement, an unsettled claim has no observed settlement duration at the data cutoff. Recording it as zero would assert that settlement occurred at the report time. The target definition and data handling should preserve the distinction between a measured zero, a missing value, and an outcome that has not yet been observed.
Before calculating a target, confirm that the data can establish whether qualifying events occurred throughout the full period. If a policy was only observed for part of the specified window, a zero for the entire window may not be justified. The observation period and the target period must be compatible.
How to Build a Precise Target
Choose a specific amount, duration, count, or other numeric outcome. Distinguish, for example, paid amounts from incurred amounts.
Specify dollars, thousands of dollars, calendar days, or another unit. If the target is a rate, name its denominator too.
Define when the interval starts and ends, whether the endpoints are included, and which event date determines inclusion.
Record zero only when it is a verified outcome for the complete period. Keep incomplete or unobserved outcomes distinct.
Read the definition as if you were a person receiving one prediction. Confirm that the stated value can be interpreted without guessing what was counted.
Worked Examples
Worked Example: Payments Within a Policy Window
A hypothetical insurer wants a numeric target for each policy during the 90 calendar days beginning January 1, 2026. Define the target as the total claim payments posted for that policy during the period. One policy has payments of $820 posted February 10 and $175 posted March 31. Another has no payments posted during the period. State the measurement interval and calculate both targets.
Method. Use the payment-posted date as the inclusion rule and define the period as \([\text{January 1},\text{April 1})\). It includes January 1 through March 31 and excludes April 1. This is 90 days: January has 31 days, February 2026 has 28, and March has 31, so \(31+28+31=90\).
First policy. Both payment dates fall within the interval. The total is \(820+175=995\). Its target is $995 per policy over the 90-day period.
Second policy. The period is fully observed and no qualifying payments occurred, so the total is $0 per policy over the 90-day period. This is a genuine zero, not a missing value.
Check. A payment posted on April 1 would be excluded by the half-open interval, even though it is adjacent to the 90-day period. The common mistake is to include it because it falls on the “next day after March 31”; the stated end boundary resolves the issue. Also, a claim incurred during the period but paid later would not be included under this particular payment-date definition.
Worked Example: Elapsed Calendar Days to Settlement
A hypothetical claims team defines settlement duration as elapsed calendar days from the date a claim is reported to the date it is settled. A claim is reported on January 28, 2026, and settled on February 6, 2026. Calculate its duration under this definition, which uses the difference between the two dates rather than counting both dates as full days.
Method. Calculate the elapsed time between the dates. A contiguous breakdown is January 28 to January 31, followed by January 31 to February 6.
From January 28 to January 31 is \(31-28=3\) days. From January 31 to February 6 is 6 days. Therefore, the total is \(3+6=9\) elapsed calendar days.
As a check using a different split, January 28 to February 1 is 4 days, and February 1 to February 6 is \(6-1=5\) days. The total is again \(4+5=9\) days.
Answer and interpretation. The recorded duration is 9 elapsed calendar days. The common mistake is to count both the report date and settlement date as complete days, producing 10. That would be a different convention, not the elapsed-date difference specified here. If the intended measure were business days, the definition would need a business-day calendar and a rule for holidays.
Worked Example: A Period Total Versus a Daily Rate
A hypothetical adjuster wants to summarize claim payments for one account during the 14-day interval \([\text{June 1},\text{June 15})\). Payments posted on June 1 total $0, on June 7 total $210, on June 14 total $390, and on June 15 total $80. Calculate the period total and the average payment amount per calendar day in the interval.
Method. The interval includes June 1 through June 14. The June 15 payment is at the excluded end boundary, so it is not part of this target.
Period total. Add the qualifying payments: \(0+210+390=600\). The total is $600 during the 14-day period.
Average per calendar day. Divide that total by the number of days in the interval: \(600/14=42.857142\ldots\). Rounded only at the final step to two decimal places, the average is $42.86 per calendar day.
Check. The daily amount has units of dollars divided by days, while the period total has units of dollars. The common mistake is to report $600 as the daily amount or to include the June 15 payment. The zero payment entry contributes zero to the total; it does not remove a day from the 14-day denominator.
Why Interval Boundaries Matter
Clear boundaries prevent events from being counted twice or omitted when periods meet. Consider two contiguous intervals, \([a,b)\) and \([b,c)\). The first includes events at \(a\) but not at \(b\); the second includes events at \(b\) but not at \(c\). Thus, each event time from \(a\) up to, but not including, \(c\) belongs to exactly one of the intervals.
If \(C([a,b))\) denotes the total amount associated with qualifying events in \([a,b)\), then the total over the combined interval is the sum of the two subperiod totals:
This holds because the subintervals do not overlap and together cover \([a,c)\). No event is counted twice, and none is omitted. If instead both subintervals included \(b\), an event at \(b\) could be counted in both totals. The half-open convention is one standard way to avoid this boundary problem; another convention can work if it is stated and applied consistently.
The same reasoning applies to amounts and counts, provided each qualifying event is assigned to one interval and the total is additive. It does not by itself decide which date is relevant: for a payment target, the payment date may determine inclusion, while a different target could be defined using the report date. That choice belongs in the target definition.
Common Mistakes and a Final Checklist
Many target-definition errors are not arithmetic errors. They arise because the number being calculated does not represent the stated business quantity. Before using a numeric outcome, check the following:
- Vague amount. “Claim cost” may mean paid, incurred, or another quantity. Name the intended measure.
- Unspecified unit. Dollars, thousands of dollars, elapsed days, and business days are not interchangeable.
- Unclear date rule. State whether inclusion depends on payment date, report date, or another event date.
- Ambiguous endpoints. Define whether the start and end are included. For adjacent periods, apply one consistent rule.
- Zero used for missing information. A complete period with no qualifying event can have a genuine zero. An incomplete period or an unobserved outcome is not thereby zero.
- Rate confused with total. If dividing by days, policies, or another denominator, state that denominator and the resulting unit.
Check Your Understanding
For each problem, specify the relevant target or perform the calculation using the boundary and unit information given.
- A hypothetical portfolio defines a 30-day target as dollars paid from July 1 inclusive to July 31 exclusive. A policy has payments of $240 on July 1 and $160 on July 30, and $75 on July 31. What is the target? (Answer: $400 per policy over the 30-day period.)
- A service duration is the elapsed calendar-day difference from March 4, 2026, to March 12, 2026. How many days is it? (Answer: 8 elapsed calendar days.)
- A fully observed account has no transactions meeting the definition of “payments posted during the period.” What value should its total-payment target take, and why? (Answer: $0, because the complete period was observed and no qualifying payments occurred.)
- During a 14-day interval, qualifying daily payments total $0 on 10 days, $210 on 2 days, and $390 on 2 days. Find the period total and the average per calendar day, rounded to two decimal places. (Answer: $1,200 total; $85.71 per calendar day.)
- A claim has $300 in payments observed during a target period, but the records do not establish whether additional payments were posted because part of the period is missing. Is $300 a verified total for the full period or a genuine zero? (Answer: Neither; it is an incomplete, unavailable full-period total, not a verified total or a zero.)