From rebuilding actuarial packs to reviewing what changed

The problem: assembly work crowding out actuarial judgment

For many actuarial teams, monthly reporting is dominated by assembly work. Model outputs are collected, compared with prior periods, reconciled by hand, and copied into review packs before anyone begins evaluating what actually changed.

By the time the pack is complete, only a small portion of the cycle remains for the work that matters most: challenging assumptions, investigating anomalies, and making informed decisions.

Working the problem

Here is the post-close question worked end to end as one governed session, drawn from real customer conversations. The overnight workflow below runs the same stages on schedule.

The question

The Q2 close numbers are in and the motor book's loss ratio has moved against us again. What's driving the drift versus prior cycles — and is it frequency or severity?

I'll trace it properly. Let me recall the reserving basis and the current assumption memo, rebuild the cohort development from the claims store, split the adverse margin into frequency and severity, and flag anything that needs actuarial judgement.

Based on real customer conversations — all data fictional.

The review starts from validated knowledge—the reserving basis, the triangle template, and the assumption memo—with model changes connected to everything they affect.

What the team already knew

Recalled from memory

Motor loss-ratio basisaccident-year, net of salvage & subrogation · validated

Cohort triangle templatevalidated · recalled in 31 runs

Assumption memo AM-2026-03severity inflation 4.5% p.a. · review threshold at +100 bp deviation

Validated institutional knowledge can include:

Claims and premium pull read-only from the governed stores—the loss side and the denominator on the same accident-year basis.

Governed execution

claimsdwh.sqlPull paid & incurred claims by cohort312,448 rows · 890 ms

policydwh.sqlPull earned premium by cohort5 cohorts · 118 ms

kernel.pythonBuild the development triangle & drift splitcohort × development · freq/sev split

Click a call to see what ran and why · every call audited

sqlLoss ratio by underwriting-year cohort

done

SELECT c.uw_year,
       SUM(p.earned_premium) / 1e6              AS earned_m,
       SUM(c.incurred) / SUM(p.earned_premium)  AS loss_ratio,
       pl.plan_loss_ratio,
       pl.plan_loss_ratio
         - SUM(c.incurred) / SUM(p.earned_premium) AS margin_pp
FROM   claims c
JOIN   policies p USING (uw_year)
JOIN   plan_ratios pl USING (uw_year)
WHERE  c.line = 'motor' AND c.eval_date = '2026-06-30'
GROUP  BY c.uw_year, pl.plan_loss_ratio
ORDER  BY c.uw_year;

Cohort	Earned	Loss ratio	Plan	Margin (pp)
UY 2022	€412m	68.1%	69.0%	+0.9
UY 2023	€438m	70.4%	70.5%	+0.1
UY 2024	€471m	74.2%	73.0%	-1.2
UY 2025	€502m	77.9%	74.5%	-3.4
UY 2026	€260m	79.4%	75.0%	-4.4

Evaluated Jun 30, 2026 · net of salvage & subrogation

Development against prior cohorts and the frequency/severity split—computed in the record, not in a side spreadsheet.

The analysis

The drift is concentrated in the two most recent underwriting years — UY 2025 and UY 2026 are 3.4 and 4.4 points adverse to plan, while mature cohorts sit on pattern. The development curves and the decomposition point the same way: this is severity, not frequency.

Margin vs plan — frequency / severity split

Percentage points of loss ratio by underwriting year

Negative = adverse to plan. Demo data.

Claim severity vs assumption path

Average repair cost, indexed · assumption memo AM-2026-03

Demo data.

The drift lands as a finding with its evidence attached, flagged for actuarial judgement—no automatic assumption change.

The finding — evidence attached

Key finding

Severity inflation is running at 7.2% against the 4.5% assumption

Repair-cost severity on recent cohorts is 7.2% annualised versus the 4.5% in assumption memo AM-2026-03 — past the memo's +100 bp review threshold. Frequency remains within normal variation. UY 2025 and UY 2026 are 3.4 and 4.4 pp adverse to plan; this is flagged for the reserving committee, with no automatic assumption change.

3 audited tool calls · 3 memory items · 3 checks passed · as of Jun 30, 2026 evaluationHow was this derived?

Saved to memory

Motor severity drift7.2% actual vs 4.5% assumed — candidate, pending actuarial validation

The review pack stages itself, and the meeting starts with the exceptions.

The deliverable

I've staged the review pack — cohort table, development, the frequency/severity decomposition, and the exceptions that need actuarial judgement. Every figure links back to its query, run, and the assumption memo.

PresentationMotor book — actuarial review

5 slides

Q2 2026 close · Reserving reviewMotor book — loss-ratio drift review

DevelopmentRecent cohorts are developing above pattern

DecompositionSeverity explains most of the adverse margin

AssumptionSeverity inflation runs at 7.2% vs 4.5% assumed

ActionsItems for actuarial judgement

PresentExport to PowerPoint

What the overnight workflow does

The session above is a one-off investigation. The same stages run unattended as a scheduled workflow - deterministic, with its parameters pinned, so no run brings surprises:

The deployment: one governed view across every model

DataQube runs inside the insurer's own environment—on-premise, in private cloud, or across multiple data centers for resilience.

Model outputs remain read-only. Scheduled workflows rebuild review packs after every reporting cycle, while validated institutional knowledge connects models, assumptions, methodologies, and prior decisions into one governed analytical record.

What changes for the actuarial team

Review, not assembly

The purpose of an actuarial review is not to rebuild reports.

It is to understand what changed, why it changed, and whether it should change the business decision.

DataQube prepares the evidence. Actuaries apply the judgment.