Diagnosing Unexpected Results With Attribution Bridges

Why an Attribution Bridge Is Useful

A monitoring result can show that premium, losses, or another performance measure differs from expectation. It does not, by itself, explain why. An attribution bridge makes the movement traceable: it starts from a defined baseline, applies measured changes in a stated sequence, and compares the resulting endpoint with the observed result.

For ratemaking operations, plausible sources include inaccurate or changed data, rating-rule implementation, policy transactions, portfolio composition, claim experience or maturity, and external conditions such as repair-cost inflation. These are diagnostic categories, not conclusions. A bridge is useful only when each contribution is supported by a reproducible calculation and the categories do not count the same effect twice.

Definition: Let $M_0$ be a baseline value of a clearly specified metric. Let $M_j$ be the value after the $j$th bridge step, with $M_k$ the final bridged value. The contribution of step $j$ is $C_j=M_j-M_{j-1}$. If the bridge reconciles to the observed result, then $M_k$ equals that observed value, and the total change is $M_k-M_0=\sum_{j=1}^{k}C_j$. If the bridge does not reconcile, the difference is an explicit residual, not an attributed cause.

Before calculating contributions, specify the metric and its unit, the observation period, the population included, and the baseline. For example, “total written premium in dollars for policies bound during April” is more useful than “premium was higher.” A change in total dollars can reflect more policies, different policy characteristics, changed rating, or transaction timing; it should not be interpreted as a change in rate adequacy without further analysis.

Constructing a Reproducible Bridge

A practical bridge uses a sequence of controlled replays or reconciled adjustments. Start with the baseline calculation. Then replace or adjust one driver at a time while holding the other definitions and assumptions fixed. The order should be chosen deliberately, recorded, and used consistently when results are compared.

Fix the measurement.
Define the metric, period, unit, included records, and whether the value is observed, estimated, or adjusted for maturity.

Reproduce the baseline.
Recalculate the reference result from its underlying records and document any exclusions or assumptions.

Replay candidate drivers.
Change a single defined driver at each step, such as corrected input data or the production rating logic, and record the new checkpoint.

Reconcile and assess.
Sum the sequential contributions, compare the final checkpoint with the observed result, and investigate any remaining residual.

The arithmetic behind the bridge is a telescoping sum:

\sum_{j=1}^{k} C_j = \sum_{j=1}^{k}(M_j-M_{j-1}) = (M_1-M_0)+(M_2-M_1)+\cdots+(M_k-M_{k-1}) = M_k-M_0.

Every intermediate checkpoint appears once as a positive term and once as a negative term, so the intermediate values cancel. This proves that sequential contributions reconcile to the change from the starting checkpoint $M_0$ to the ending checkpoint $M_k$. When the bridge reconciles to the observed result, $M_k$ is that observed result. The identity is arithmetic; it does not prove that the chosen categories are causally correct.

Bridge contributions can depend on order when drivers interact. For instance, a rating-rule change may have a larger dollar effect on a portfolio with a different mix of risks. A sequential bridge assigns the interaction according to the order of its replays. State that order, and avoid presenting the resulting contributions as order-independent causal effects. If the order materially changes the conclusions, compare alternative orders or report the interaction as unresolved rather than implying false precision.

Worked Premium Bridge

Worked Example: Reconciling Higher Written Premium

A hypothetical insurer is reviewing April written premium. Its comparison baseline is $600,000. A record-level review identifies an exposure-data correction, a difference between the expected and production rating logic, transaction effects, and a change in portfolio mix. A sequential replay gives checkpoints of $612,000 after the data correction, $619,650 after the rating-logic replay, $617,650 after transactions, and $624,000 after the mix replay. Find each contribution and determine whether the bridge reconciles to observed written premium of $624,000.

Method. Use a sequential dollar bridge, with the baseline first and the drivers replayed in the order stated. The contribution at each step is the new checkpoint less the preceding checkpoint.

Data correction: $612{,}000-600{,}000=12{,}000$, a contribution of +$12,000.

Rating logic: $619{,}650-612{,}000=7{,}650$, a contribution of +$7,650.

Transactions: $617{,}650-619{,}650=-2{,}000$, a contribution of −$2,000.

Portfolio mix: $624{,}000-617{,}650=6{,}350$, a contribution of +$6,350.

The contributions sum to $12{,}000+7{,}650-2{,}000+6{,}350=24{,}000$. The observed difference is $624{,}000-600{,}000=24{,}000$, so the bridge reconciles exactly. The result describes this sequence of measured replays; it does not establish that any one driver is an error or that a rate change is warranted.

Common mistake avoided: treating every contribution as positive because the endpoint is above baseline. The transaction contribution is negative: its checkpoint fell by $2,000. Keeping the sign is necessary for the total to reconcile.

Use the Right Metric for Loss Outcomes

Loss bridges require particular care about what is being compared. Reported incurred losses, paid losses, and estimated ultimate losses are not interchangeable. A result affected by claim maturity should be compared with a baseline on a consistent maturity basis, or the maturity adjustment should appear as an explicit bridge step. Likewise, a change in the number or mix of exposures can move total losses even if the loss level within each risk group has not changed.

External changes and claim emergence should be distinguished where evidence allows. For example, a measured repair-cost trend may be an external severity effect; a change in the number of claims reported may be a frequency or claim-emergence effect. The bridge should use the definitions and valuation dates that support the selected metric, rather than assign an unexplained difference to a convenient category.

Worked Loss Bridge

Worked Example: Explaining an Increase in Estimated Ultimate Losses

For a hypothetical book, the expected ultimate loss baseline for a fixed evaluation population is $420,000. After aligning the claim valuation to the current maturity basis, a sequential review attributes $28,000 to claim maturity, $22,000 to external severity change, $18,000 to portfolio mix, and $16,000 to corrected claim data. The observed, maturity-aligned ultimate loss is $504,000. Calculate the checkpoints and test reconciliation.

Method. Build an additive sequential bridge in the order given, starting at $420,000. Each checkpoint equals the prior checkpoint plus the next stated contribution.

Start at $M_0=420{,}000$. After the maturity step, $M_1=420{,}000+28{,}000=448{,}000$. After external severity, $M_2=448{,}000+22{,}000=470{,}000$. After portfolio mix, $M_3=470{,}000+18{,}000=488{,}000$. After the claim-data correction, $M_4=488{,}000+16{,}000=504{,}000$.

The total attributed increase is $28{,}000+22{,}000+18{,}000+16{,}000=84{,}000$. Independently, $504{,}000-420{,}000=84{,}000$. Thus $M_4$ equals the observed $504,000 and the bridge reconciles. The maturity step is an adjustment needed for a like-for-like comparison; it should not be described as an increase in underlying claim cost without further evidence.

Common mistake avoided: comparing immature reported losses directly with a mature baseline and attributing the difference to worsening experience. Here maturity is explicitly adjusted before drawing conclusions about other drivers.

Evidence, Residuals, and Interpretation

A bridge is an accounting of a measured change, not a substitute for diagnosis. A contribution labeled “rating logic” should come from an auditable comparison of the relevant rule or implementation, not from the fact that premium differs from expectation. A “data” contribution should be tied to identified records or fields. A “mix” contribution should reflect a defined change in the distribution of risks, not be used as a catch-all for unexplained movement.

When $M_k$ differs from the observed result $M_{\mathrm{obs}}$, define the residual consistently as:

R=M_{\mathrm{obs}}-M_k, \qquad M_{\mathrm{obs}}-M_0 = \sum_{j=1}^{k}C_j+R.

A nonzero residual can arise from omitted drivers, inconsistent record populations, rounding, timing differences, or errors in the replay. It should be displayed and investigated. Do not spread it among known categories merely to force a zero residual. Conversely, an exactly reconciled bridge can still have poor causal labels if its categories overlap or its replays are not supported.

For any bridge, retain enough detail for another analyst to reproduce the result: baseline definition, data extracts or record counts, calculation sequence, treatment of transactions, claim valuation basis where relevant, and the rounding convention. State dollar contributions to the cent when calculations are in dollars, or use a clearly stated larger unit consistently. If rounded contributions do not sum to the rounded total, disclose the rounding difference.

Key takeaway: A sound attribution bridge defines its metric and baseline, changes one specified driver at a time, reports signed contributions in a stated order, and reconciles the final checkpoint to the observed result or shows a residual. It narrows the investigation; it does not by itself prescribe a corrective action.

Common Exam Traps

Ending the telescoping sum at the wrong checkpoint. The correct identity is $M_k-M_0=\sum_{j=1}^{k}(M_j-M_{j-1})$, not $M_1-M_0=\sum_{j=1}^{k}(M_j-M_{j-1})$ unless $k=1$. The sum ends at $M_k$.
Ignoring negative contributions. A driver can offset an overall increase. Preserve the sign of each checkpoint difference.
Mixing populations or periods. A premium baseline for one exposure population cannot be reconciled meaningfully to an observed result for another without making the population change explicit.
Double-counting a driver. If an input correction is already included in a rating replay, it cannot also be counted as a separate data contribution unless the replays are designed to isolate the effects.
Calling an attribution a causal finding. The bridge records the result of defined replays. Unsupported explanations remain hypotheses, even when the arithmetic reconciles perfectly.

Worked Example: Detecting a Bridge Arithmetic Error

A hypothetical bridge has $M_0=\$250{,}000$, $M_1=\$258{,}000$, and $M_2=\$263{,}500$. An analyst reports contributions of $8,000 and $5,500, then claims that their sum explains $M_1-M_0=\$8,000$. Identify the error and give the correct reconciliation.

Method. Compute each sequential contribution and compare their sum with the change from the first checkpoint to the last.

The first contribution is $C_1=258{,}000-250{,}000=\$8{,}000$. The second is $C_2=263{,}500-258{,}000=\$5{,}500$. Therefore $C_1+C_2=8{,}000+5{,}500=\$13{,}500$, which equals $M_2-M_0=263{,}500-250{,}000=\$13{,}500$.

The analyst incorrectly compared the sum through step 2 with $M_1-M_0$. The correct endpoint is $M_2$, so the bridge explains an increase of $13,500, not $8,000. This example also checks the telescoping identity directly.

Common mistake avoided: stopping at the first checkpoint in the left side of the equation while summing contributions through the second checkpoint.

Check Your Understanding

For each problem, calculate the requested bridge result and check the endpoint or residual.

A hypothetical premium bridge starts at $310,000. Its checkpoints after data and rating-logic replays are $316,400 and $321,000. Find both contributions and the total bridged change. (Answer: +$6,400 and +$4,600; total +$11,000.)
A loss bridge starts at $780,000 and has sequential contributions of −$15,000 for a maturity adjustment, +$42,000 for severity, and +$8,000 for mix. Find the ending checkpoint. (Answer: $815,000.)
A bridged result is $1,046,000, while the observed result is $1,050,500; the baseline is $1,000,000. Find the residual and the total observed change. (Answer: residual +$4,500; observed change +$50,500.)
A two-step bridge has $M_0=\$500{,}000$, $M_1=\$512{,}000$, and $M_2=\$507{,}500$. Find the signed contributions and verify their sum against the endpoint change. (Answer: +$12,000 and −$4,500; sum and endpoint change are both +$7,500.)
A hypothetical bridge reports a baseline of $900,000 and contributions of +$24,000, −$7,500, and +$13,000. The observed result is $932,000. Find the bridged endpoint and residual, using $R=M_{\mathrm{obs}}-M_k$. (Answer: bridged endpoint $929,500; residual +$2,500.)

What You'll Learn