Set the Claim-Cost Prediction Cutoff

Why a Prediction Cutoff Matters

A claim-cost prediction is useful only if it can be made with information that would actually have been available when the prediction is needed. A claim may have occurred before that date but been reported afterward; a reserve may have been revised months later; a policy field may now show a value that was not in force at the time. Treating such records as if they were known earlier gives a model an unfair view of the future.

The solution is to specify an exact prediction cutoff, define the historical period used to create predictors, and state the future period and valuation rule used to define the outcome. These choices turn “information available at prediction time” into a checkable data rule. This discipline is aligned with the SOA Exam PA (Predictive Analytics) syllabus.

Definition: The prediction cutoff, denoted $t_0$, is the exact date and time at which a prediction is assumed to be made. The observation window is the historical interval from which predictor information may be drawn. A field is eligible only if its relevant value was available to the decision-maker by $t_0$, subject to the specified rules for the field and observation window. The outcome window is the future interval over which the target claim cost is defined.

A cutoff should include a time and, where records use different time zones, a time-zone convention. “At renewal” or “at month-end” may be too vague: an event stamped 08:00 on the renewal date can fall on either side of a cutoff at 08:00 depending on the exact inclusion rule. A practical convention is to include records known at or before the stated cutoff, and to define the observation window as ending just before it.

\text{Observation window}=[t_0-L,\ t_0), \qquad \text{Outcome window}=[t_0,\ t_0+H)

Here $L$ is the length of the historical lookback and $H$ is the length of the prediction horizon. These half-open intervals make the boundary explicit: an event at $t_0$ is not historical under this convention. The definitions may instead use calendar periods or another boundary rule, but the rule must be stated and applied consistently.

Occurrence, Reporting, and Availability Are Different Dates

For a claim, at least three dates can matter. The loss occurrence date describes when the insured event happened. The report date records when the claim was reported. The availability date indicates when the relevant information entered a system or otherwise became accessible to the person making the prediction. A report can be delayed, and a system entry can be later than the report itself.

Eligibility is therefore not settled by checking only whether a loss occurred before $t_0$. A historical claim predictor often requires both an occurrence within the chosen observation window and a claim record available by the cutoff. The precise rule depends on what the predictor represents. For example, a reported-claim count requires the claim to have been reported by the cutoff; a count of events known from another verified source could have a different availability rule. Do not infer a field’s availability from its value alone.

Key rule: A predictor must be constructed from information available at the cutoff, not from the latest value currently stored in a database. For dated claims data, check the event date, report date, transaction or entry date, and any later revision date that affects the field.

Worked Example: Reconstructing the Claims Known at a Cutoff

A hypothetical insurer makes predictions at April 1, 2026, 08:00. Its claim-count predictor is the number of claims with loss occurrence dates from January 1, 2026, inclusive, to the cutoff, and with reports available by the cutoff. Apply that rule to the records below.

Claim	Loss occurrence	Reported or available
1	February 10, 2026	February 12, 2026
2	December 28, 2025	January 4, 2026
3	March 20, 2026	March 31, 2026, 17:00
4	March 29, 2026	April 2, 2026
5	March 31, 2026	April 1, 2026, 08:05

Method. Check the occurrence-window condition and the availability condition separately for each row.

Claim 1: February 10 is on or after January 1 and before the April 1 cutoff. Its report date, February 12, is before April 1 at 08:00. It is eligible.

Claim 2: Its report is available by the cutoff, but its loss occurred on December 28, before the observation window begins. It is not eligible.

Claim 3: March 20 is within the observation window, and the claim was reported on March 31 at 17:00, before the cutoff. It is eligible.

Claim 4: Its occurrence is within the observation window, but its report was not available until April 2, after the cutoff. It is not eligible.

Claim 5: Its occurrence is within the observation window, but the record became available at 08:05 on April 1, five minutes after the cutoff. It is not eligible.

Reconstructed predictor: Claims 1 and 3 qualify, so the valid claim count is $1+1=2$. The check is that both eligible records meet both conditions; every excluded record fails at least one.

Common mistake avoided: Claim 1 is not future information: both its occurrence and report dates precede the cutoff. Rejecting it because it happened after the start of the year, or confusing its date with a later claim, would incorrectly reduce the count to 1.

Freeze Policy and Claim Fields as of the Cutoff

Claims are not the only source of leakage. Policy records can be overwritten when coverage changes, a renewal is processed, or an address is corrected. A current extract may show the most recent value rather than the value known at the prediction date. For each policy predictor, identify the field’s effective date and the date its value was recorded or available. Use the version that meets the cutoff rule.

The same care applies to claim financial fields. Payments made after $t_0$, new reserve estimates, and later claim-status changes cannot be used to describe the information available at $t_0$. If a predictor is intended to be “paid to date” or “incurred to date,” calculate it from transactions and reserve values available by the cutoff. Retaining an as-of snapshot, or reconstructing one from dated transactions and versions, is safer than using a present-day total.

Worked Example: Using Claim and Policy Values as They Stood

A hypothetical insurer predicts at July 1, 2026, 12:00. It defines a 90-day historical window beginning April 2, 2026, at 12:00, and counts claims with loss dates in that window that were available by the cutoff. It also wants a historical incurred-cost feature defined as payments to date plus the case reserve recorded by the cutoff.

Claim A occurred May 10 and was reported May 12. Payments of $600 on May 20 and $300 on June 25 were recorded by the cutoff. Its June 30 reserve was $500; a revised July 2 reserve was $700.
Claim B occurred March 30 and was reported April 3.
Claim C occurred June 20 and was reported at July 1, 12:15.
Claim D occurred June 29 and was reported June 30 at 10:00. No payment was made by the cutoff; its June 30 case reserve was $400. A payment was made July 3.

Step 1: Apply the occurrence window. Claims A, C, and D occurred on or after April 2 at 12:00 and before July 1 at 12:00. Claim B occurred before the window began, so it is excluded even though it was reported during the window.

Step 2: Apply the availability cutoff. A and D were reported before July 1 at 12:00. C was reported at 12:15, after the cutoff, so it was not yet available. The eligible historical claim count is $1+1=2$, for Claims A and D.

Step 3: Reconstruct incurred cost from available values. For A, available payments total $600+300=900$, and the available reserve is $500. Its incurred-to-date value is $900+500=\$1{,}400$. For D, available payments total $0 and the reserve is $400, so its value is $0+400=\$400$. The total for eligible claims is $\$1{,}400+\$400=\$1{,}800$.

Step 4: Exclude later revisions and transactions. The July 2 reserve of $700 and July 3 payment were not known at the cutoff. Using them would replace the historical snapshot with future information.

Check: The two included claim records both have in-window occurrence dates and pre-cutoff reports. The cost total uses only pre-cutoff payments and reserves: $900 + $500 + $0 + $400 = $1,800.

Common mistake avoided: Counting Claim B because its report arrived in the window ignores the stated occurrence-date rule. Using the July 2 reserve for A or July 3 payment for D ignores the prediction cutoff.

Define the Future Cost Outcome Separately

A prediction needs a target as well as predictors. A statement such as “predict claim cost” is incomplete until it specifies which future losses count and how their costs are valued. One possible target is the total of paid amounts plus case reserves, measured at a fixed valuation date, for losses occurring during a stated future interval. This is a defined valuation basis, not automatically the ultimate cost of those claims.

It is useful to keep the predictor cutoff and outcome valuation date distinct. The outcome will generally rely on information that was not available at prediction time; that is what makes it an outcome. The leakage rule applies to predictors. For a usable training target, however, the outcome definition must be consistent across observations, and the data must have had enough time to develop under the chosen reporting and valuation rules. A late-reported loss may be absent from an immature outcome extract, so the label’s completeness needs attention.

Y=\sum_{j:\ t_0\leq \text{occurrence}_j<t_0+H} \left(\text{paid}_j(V)+\text{case reserve}_j(V)\right)

In this example definition, $V$ is the fixed valuation date and $Y$ is the total cost assigned to losses occurring in the outcome window. The formula includes only claims meeting the occurrence-date rule and values their payments and reserves as of $V$. Other target definitions are possible, such as paid-only cost, but changing the rule changes the meaning of the prediction.

Worked Example: Assigning Claims to a Future Cost Window

At December 31, 2025, 23:59, a hypothetical insurer predicts cost for losses occurring from January 1 through March 31, 2026, inclusive. It values the outcome at September 30, 2026, as payments plus case reserves recorded by that date. The records are:

Claim A: loss January 15, reported January 16; $4,000 paid and $1,000 case reserve at September 30.
Claim B: loss March 31, reported April 10; $2,500 paid and $500 case reserve at September 30.
Claim C: loss December 28, 2025, reported January 3; $6,000 paid by September 30.
Claim D: loss March 20, reported October 5, after the valuation date.
Claim E: loss April 2, reported April 2; $700 paid by September 30.

Step 1: Apply the outcome occurrence window. A, B, and D occurred from January 1 through March 31, so their occurrence dates satisfy the target rule. C occurred before January 1, and E occurred after March 31; neither belongs in this outcome.

Step 2: Apply the stated valuation rule. A contributes $4{,}000+1{,}000=\$5{,}000$. B contributes $2{,}500+500=\$3{,}000$, even though it was reported after the prediction cutoff: it is an outcome claim, not a predictor at that cutoff. D was not reported by the September 30 valuation date, so it is absent under the stated observed-data rule; this signals a possible incomplete label, not proof that its true cost was zero.

Step 3: Total the observed target. The observed target is $\$5{,}000+\$3{,}000=\$8{,}000$. C and E are excluded because their occurrence dates are outside the future loss window.

Check: The total includes only qualifying occurrences with costs available under the valuation rule. Claim D illustrates why the outcome development and valuation convention must be disclosed: the observed $8,000 may omit a loss that had occurred in the target period but was reported too late.

Common mistake avoided: Claim B’s post-cutoff report does not make it a predictor, but it can contribute to the later outcome. Conversely, Claim C’s January report does not move a December loss into the future outcome window.

Why the Cutoff Rule Works

The basic justification is an information constraint. Let $\mathcal{F}_{t_0}$ denote the information actually available to the decision-maker at cutoff $t_0$. A deployable prediction must be a function of fields that belong to that information set. If a feature depends on a report, payment, reserve revision, or policy change first available after $t_0$, it cannot be reconstructed from $\mathcal{F}_{t_0}$.

State the decision time.
Record the precise cutoff and the inclusion convention for records stamped at that time.

Define each window.
Specify the historical lookback, future occurrence period, and any outcome valuation or reporting runout.

Reconstruct the snapshot.
For each candidate predictor, use the value and claim transactions available by the cutoff, not the latest extract.

Audit boundary cases.
Check losses near the window endpoints, reports near the cutoff, and fields revised after it.

Define and mature the label.
Apply the stated future-loss and valuation rules consistently, and identify outcomes that may be incomplete because of reporting delay.

The justification can be stated directly: a prediction made at $t_0$ cannot use information that was not in $\mathcal{F}_{t_0}$. A feature recorded after $t_0$ is not available to a decision-maker at $t_0$, even if it describes an event that happened earlier. Including that feature lets the analysis learn from knowledge acquired later; any apparent improvement may therefore fail when predictions are made prospectively.

Common Cutoff Traps

Using event date as a proxy for availability. A pre-cutoff loss reported afterward was not yet a reported claim predictor at the cutoff.
Using report date as a substitute for occurrence date. A claim reported during the lookback can still have an occurrence before the lookback began.
Taking current values from a revised database. A reserve, policy status, or address corrected after the cutoff must not silently replace its earlier value.
Letting outcome information enter predictors. Future payments and claim developments can define a target under a stated rule, but they cannot be features available at the earlier prediction time.
Leaving boundaries implicit. State whether the cutoff instant or window endpoints are included and use the same convention in the data construction.

A compact audit table can make the rule operational. For each proposed field, record its meaning, source timestamp, effective date if relevant, and the cutoff-time version used. For claims, separately record occurrence, report, and transaction or availability dates. This makes exclusions explainable and helps distinguish true predictor gaps from records that simply arrived too late.

Key takeaway: A claim-cost prediction table should represent a reproducible historical snapshot. Set $t_0$, specify the lookback and target rules, then use only the policy and claim information that was available by $t_0$. A historical event date alone does not establish that its information was available.

Check Your Understanding

For each problem, apply the stated date and availability rules rather than assuming that the newest database value was known at prediction time.

A prediction is made at 09:00 on August 1. A loss occurred July 20 and was reported August 1 at 09:04. Under a rule requiring reports available by the cutoff, is it a predictor claim? (Answer: No; the report arrived four minutes after the cutoff.)
A policy address was recorded as “North” on June 1 and changed to “South” on September 3. The prediction cutoff is August 31 at 23:59. Which address value may be used as the as-of predictor? (Answer: North, assuming it remained the available value through the cutoff.)
A 60-day observation window runs from May 3 at 00:00 to July 2 at 00:00, exclusive of the endpoint. Claim A occurred May 2 and was reported May 4; Claim B occurred June 30 and was reported July 1; Claim C occurred June 29 and was reported July 2 at 00:01. Reports must be available before the endpoint. How many claims qualify? (Answer: 1; only Claim B meets both the occurrence-window and availability conditions.)
At a December 31 cutoff, the future outcome counts losses occurring January through March and values them at June 30. Claim X occurred March 30, was reported April 5, and has $1,200 paid plus a $300 reserve at June 30. Does it contribute to the outcome, and by how much? (Answer: Yes; it contributes $1,500 under the stated rule.)
A claims extract at a March 31 cutoff shows $2,000 paid by March 31 and a $900 reserve revised April 4. The predictor is incurred cost as known at March 31; the March 31 reserve was $700. What value should the predictor use, and what would be the leakage error? (Answer: Use $2,700 = $2,000 + $700; using the later reserve gives $2,900 and leaks $200 of post-cutoff information.)

What You'll Learn