A Claims KPI Change Can Create a Reserving Problem

The claims metric most likely to hurt us is the one everybody likes. A closure rule falls behind target. So we sort pending inventory into the older buckets, tighten what may remain open, or route low-value glass and towing claims below a $2,500 ceiling into straight-through settlement. The dashboard improves. More files close. Pending inventory looks less old. Somebody says the operation is moving. Then the reserve triangle moves.

The triangle does not know whether a movement came from claims experience or from someone changing how a claims team records a file.

It receives cumulative paid or incurred loss by accident year at successive valuation ages: twelve months, twenty-four, thirty-six, onward. It is an instrument for estimating ultimate cost. It is not a witness to what produced the cells. That is the inconvenient part. A claims metric stops being an operations number the moment it enters a reserve. It has crossed from a measure of work into an input to the balance sheet. None of this argues against faster claims handling. A file investigated, evaluated, paid, subrogated, and closed shouldn’t stay open so a development pattern looks familiar. It is an argument against making actuarial infer our intervention from a triangle afterwards. The change notice travels with the data.

The triangle sees a result, not a cause

The portfolio triangle groups claims by accident year or report year and follows aggregate incurred value through valuation ages. The age-to-age factor, or link ratio, divides cumulative loss at one age by the age before. Chain ladder projects the latest diagonal to ultimate through the remaining link ratios and a tail factor. That method rests on one demanding assumption. The expected ratio of cumulative loss at age k+1 to cumulative loss at age k is the same across accident years. We call this “using the data.” We are really assuming the process that turns an accident into a reported, reserved, paid, closed, sometimes reopened file has stayed comparable.

A new litigation-management panel disturbs that.

So does worst-case reserving replacing probable-outcome.

So does a tighter closure rule, a moved first-notice timestamp, or a workflow releasing undisputed amounts earlier.

Each one changes the shape of reported development.

None changes the accidents underneath.

The paid-to-incurred ratio makes the ambiguity visible.

At a given age it shows how fast money has left against the standing estimate.

A rise can mean faster settlement.

It can mean slower settlement.

It can mean a change in reserving basis.

The ratio can’t tell the actuary which.

Nor can a closure count settle it.

Closure ratio is claims closed in a period divided by claims reported in that period.

A ratio below 1.0 means inventory is building.

But the reopen rate belongs beside it: files closed, then reopened within three months. A closure statistic without a reopen rate is a performance claim with its most relevant qualifier removed.

We can clean a backlog by closing files that come back.

The reserve method sees the closure first.

A case reserve is an operating estimate before it becomes IBNER

Within days of assignment, an adjuster sets a case reserve for estimated ultimate cost, indemnity separated from allocated expense.

That number determines authority.

Reserve changes and settlement offers climb a ladder: adjuster, team leader, claims manager, round-table.

The file reserve moves on a monthly or quarterly diary.

It also moves when information changes the estimate: a statement of claim, a plaintiff demand, a medical report, a defence assessment, a coverage opinion.

The discipline is to move it when information arrives.

Not gradually, as pressure becomes harder to ignore.

At portfolio scale that discipline becomes IBNER.

Incurred but not enough reported: development on claims we already know about and have already case-reserved.

It is distinct from IBNR, which estimates claims incurred but never reported to us at all.

Case-reserve strengthening changes IBNER.

It does not generate an unreported claim.

So we cannot log a new reserve practice as a local configuration change.

Worst-case reserving raises incurred amounts on known files even when severity and frequency haven’t moved.

A revision beyond roughly $25,000 or a quarter of the standing reserve typically needs sign-off one level up.

And asking for a reserve update before a settlement offer changes the order in which incurred loss emerges.

The initial-to-ultimate ratio shows this directly.

It sets the early case reserve against incurred at closure.

Clustering near 1.0 suggests honest early evaluation.

Skewed below 1.0 points at systematic under-reserving.

But it is still a forecast score.

Change the initial-reserve rule and its basis moves too.

Nobody should be asked to preserve a bad process so an old triangle keeps looking tidy.

The obligation is less theatrical than that.

Announce the change. Date it. Name the populations.

Then let actuarial decide whether development is still comparable.

The reserve method carries the consequence

Chain ladder is useful where several development periods have emerged and the pattern is stable.

It is least comfortable where we are most tempted to say nothing.

That is the recent accident years, where a small reported amount is multiplied by a large remaining development factor.

Bornhuetter-Ferguson imposes a different discipline.

It weights reported experience by the proportion developed, and an a priori expected loss ratio by the rest.

So it works for the newest accident year and for long-tail lines, where one volatile link ratio swings an answer by multiples.

Neither method absolves us of explaining a change.

Chain ladder needs stable development.

Bornhuetter-Ferguson needs a credible a priori view.

Neither knows we changed the closure queue in June.

The consequences reach pricing too.

Pure premium is expected frequency times expected severity.

Frequency responds to driving exposure and reporting behaviour.

Severity responds to repair, medical, and settlement values.

We trend them separately because an offset between them is not a stable loss trend.

The loss-ratio method gives an indicated change, not a rate: trended developed experience loss ratio over permissible loss ratio.

Permissible loss ratio is one minus expense and profit.

We bring historical premium on-level, as though current rates had applied throughout, before that indication means anything.

If practice moved development, the rate indication needs to know.

Otherwise a process improvement reads as worse severity.

Or we read case-reserve strengthening as worsening frequency.

The arithmetic isn’t the problem.

The attribution is.

The interface should be formal

Here is the boundary I would hold.

Claims operations should not send actuarial a quarterly explanation after a surprising development result.

Both functions need one change-control record beforehand.

That covers any change touching timestamps, reserving, closure, payment, reopening, coding, assignment, or the population a report includes.

The record needs no transformation programme around it.

It needs the old process and the new one.

Effective date. Affected lines and accident years. Systems touched.

Whether a monthly diary becomes a quarterly one.

Whether case-reserve, payment, or closure authority changes.

Expected metric movement, and a named actuarial treatment.

It should say whether the series needs a break, a parallel run, a restatement, a separate triangle segment, or just monitoring.

The first-notice timestamp deserves more respect than it gets.

On broker-placed commercial business, the gap between a broker’s call and our file creation runs two to four days.

Picking one as the start of the clock improves cycle time without improving the claimant’s week.

That is a measurement artifact.

It should be visible before anybody concludes anything.

The same applies to automation.

Straight-through processing suits low-value files: glass, towing, single-item contents.

It should leave an auditable event trail.

What changed, which rule applied, whether payment or reserve moved, whether the file later reopened.

Automation can accelerate a file.

It cannot explain its own effect on a reserve unless we design that in.

Canadian governance is unusually direct here.

A federally regulated insurer has an Appointed Actuary, who opines on policy liabilities under CIA Standards of Practice.

Under IFRS 17 it is discounted and carries a risk adjustment.

So an operational data change booked as liability experience is not a dashboard problem.

It is a basis problem in a signed estimate.

Two functions, one timestamp

The board hears that closure ratio improved, time to first payment fell, and the oldest aged-pending buckets shrank.

We need the question that survives the meeting.

Which process changes altered timestamps, reserving, closure, or payment, and how will we separate their effect from trend?

That question produces a testable record.

It joins the change to the affected population, the rule and system version, the pre- and post-change metrics, the reopen rate, the paid-to-incurred measure, the case-reserve basis, and the actuary’s treatment of the series.

It also makes an important negative finding possible.

No reserve-basis change is required, because the population and the development pattern stayed comparable.

That finding is worth as much as an alarm.

The answer nobody wants isn’t more actuarial claims reporting.

It is to make actuarial reporting less downstream.

Claims owns what its timestamps and statuses mean.

Actuarial owns the interpretation of their aggregate emergence.

Neither does the other’s work from a monthly extract.

The triangle isn’t wrong when it sees movement.

It is waiting for us to say what moved.