Reading the Measure Before Reading the Message
A risk dashboard compresses information so decision-makers can spot changes and decide whether to investigate or act. That compression can also conceal important details. A line may look steep because its axis starts above zero; a count may fall while the underlying rate rises; and a value labelled “current” may be based on an old data cut. The visual message is not reliable until the measure behind it is understood.
Dashboard interpretation is not a substitute for the earlier material on risk indicators, thresholds, and breach responses. It is a check on whether the displayed evidence supports the conclusion being drawn. Before calling a measure improving, deteriorating, or within limit, identify its unit, denominator, time basis, data date, and applicable threshold.
Check Units, Scales, and Denominators
A chart can display multiple measures together, but that does not make them directly comparable. One axis might show dollars in millions while another shows incidents per thousand policies. Separate axis ranges can make two lines appear to move together or by similar amounts even when the quantities and scales differ. Always read the axis labels, units, tick marks, and time period; do not infer magnitude from line height alone.
A second question is whether a count needs an exposure denominator. For a count \(N\) observed over exposure \(E\), a rate is \(N/E\), provided numerator and denominator refer to compatible populations and periods. The rate answers a different question from the count: it describes incidents per unit of exposure, rather than the total number of incidents. A change in exposure can therefore make count and rate move in opposite directions.
For comparisons across periods, use the same definition of event and the same exposure unit. A policy-year denominator cannot be casually compared with a policy-month denominator. If units differ, convert them before calculating or comparing rates.
Worked Example: A Falling Count but a Rising Rate
A hypothetical insurer’s dashboard shows 36 late claim notifications among 18,000 exposed policies in one quarter and 30 among 12,000 in the next quarter. The dashboard headline says “notifications down.” Determine whether the frequency per 1,000 exposed policies also improved, and explain what the headline omits.
Method. Calculate each quarter’s rate using its own exposure, then compare the results. The exposure counts and notification counts refer to the same quarter in each case.
First quarter. The rate is \(36/18{,}000=0.002\) notifications per policy. Per 1,000 policies, this is \(0.002\times1{,}000=2\).
Second quarter. The rate is \(30/12{,}000=0.0025\) notifications per policy. Per 1,000 policies, this is \(0.0025\times1{,}000=2.5\).
The count fell by \(36-30=6\), or \(6/36=0.1667\), which is a 16.67% decrease. But the rate rose from 2 to 2.5 per 1,000, an increase of \(0.5/2=0.25\), or 25%. Thus, “notifications down” is correct as a statement about total count, but it does not establish that notification frequency improved.
Check. Multiplying each calculated rate per 1,000 by its exposure in thousands gives \(2\times18=36\) and \(2.5\times12=30\), matching the reported counts. The common mistake is comparing counts alone when the exposure base changed.
The arithmetic also explains why the denominator matters. Since \(N=E(N/E)\), a total count depends on both exposure and rate. A smaller count can result from a smaller exposure even when the rate per exposure unit has increased. Conversely, a larger count does not necessarily mean deterioration if exposure has grown enough. State both the count and rate when each is relevant.
Make the Comparison Fair
A comparison is interpretable only when the measures line up on key dimensions. Check whether values are actual or projected, gross or net of risk transfer, measured on the same valuation basis, and attached to comparable populations and time windows. A quarter-to-date figure should not be compared with a full-quarter figure as if both cover the same duration.
Different axes are not automatically wrong, but they require clear labels and careful reading. A dual-axis chart can be especially misleading if one scale is compressed or its zero point is hidden. If the task is to compare percentage changes, calculate them from the underlying values instead of judging the apparent slopes. Even then, a percentage change in one measure is not directly equivalent to the same percentage change in a different measure.
Worked Example: Two Axes, Two Different Messages
A hypothetical dashboard plots catastrophe exposure on one axis and written premium on a second axis. Exposure rises from $40 million to $48 million, while premium rises from $200 million to $224 million over the same year. The two lines appear to rise by roughly the same amount on the chart. Compare their percentage changes and the ratio of exposure to premium. Assume the exposure and premium amounts are measured on consistent bases in both years.
Method. Calculate each percentage change from its own starting value. Then calculate exposure divided by premium for each year. The visual distances between lines are not a valid substitute because the axes have different units and scales.
Exposure change. The increase is \($48-\$40=\$8\) million. Relative to the initial exposure, the percentage increase is \($8/\$40=0.20\), or 20%.
Premium change. The increase is \($224-\$200=\$24\) million. Relative to initial premium, the percentage increase is \($24/\$200=0.12\), or 12%.
Exposure-to-premium ratio. Initially, the ratio is \($40/\$200=0.20\), or 20%. At the end, it is \($48/\$224\approx0.2143\), or 21.43%. The ratio rose by about \(21.43\%-20.00\%=1.43\) percentage points.
The chart’s similar-looking line movements conceal different percentage changes: exposure grew faster than premium, and the exposure-to-premium ratio increased. This ratio is only a descriptive comparison; it does not by itself establish whether the exposure is adequately priced or whether risk increased. The common mistake is treating visual distance across two differently scaled axes as a quantitative comparison.
Check Freshness and Threshold Context
A dashboard value needs a clear data date. “Current” may mean the latest available extract, not data through today or through the end of the reporting period. Reporting lag, validation, and data refresh timing can leave a displayed value behind known events. A reader should distinguish the date the dashboard was viewed from the date through which its underlying data are complete.
Thresholds also need context: the applicable limit, its effective date, and whether the displayed measure and threshold use the same definition. A line without a visible threshold is not evidence that a measure is within tolerance. Nor should a historical threshold be assumed to remain applicable if an approved change has taken effect. When those details are unavailable, the responsible conclusion is that status cannot yet be determined from the display alone.
Worked Example: A Value That Is Below Limit Only on an Old Data Cut
A hypothetical dashboard shows $7.6 million of validated incurred losses against an $8.0 million limit, with the value labelled “current.” Its data cut-off is June 30. On July 8, a further $0.7 million of validated incurred losses is confirmed for the same measurement period and on the same basis. Determine the reading using the displayed amount and using the information available on July 8.
Method. First compare the displayed figure with the limit, then update the total with the validated amount omitted by the data cut-off. Because both amounts use the same loss definition and period, addition is appropriate.
Displayed reading. The margin below limit is \($8.0-\$7.6=\$0.4\) million. On the June 30 data cut alone, the figure is $0.4 million below the limit.
Updated reading. The updated loss amount is \($7.6+\$0.7=\$8.3\) million. The excess over the limit is \($8.3-\$8.0=\$0.3\) million. The later information changes the interpretation from apparently below limit to above limit.
Check. The additional $0.7 million exceeds the displayed $0.4 million headroom by $0.3 million, matching the calculated excess. The common mistake is to treat the label “current” as proof that the value includes all information available at the time of review. The updated comparison depends on the new amount being validated and genuinely comparable; if its basis differs, reconcile the definitions before adding.
Association Is Not a Causal Finding
A dashboard may show that two measures move together. This is an association: the measures vary together in the observed data. It does not, by itself, show that a change in one measure caused a change in the other. A third factor may influence both, the direction of influence may be reversed, or the apparent relationship may be an artefact of how the data were grouped or measured.
For example, a positive correlation between training hours and claim errors across business units does not prove that training increases errors. Units with more complex claims might receive more training and also experience more errors. To assess the claim, examine relevant populations, time ordering, other plausible factors, and the design of the analysis. A dashboard association can prompt a question; it cannot, on its own, answer a causal question.
Worked Example: A Reversed Aggregate Relationship
A hypothetical insurer compares claim-error rates for units with lower and higher training hours. Units are also classified by claim complexity. In lower-complexity units, the lower-training group has 45 errors among 900 claims and the higher-training group has 4 among 100 claims. In higher-complexity units, the corresponding figures are 20 among 100 and 162 among 900. Calculate the rates within each complexity group and in aggregate. What conclusion is warranted?
Method. Calculate errors divided by claims for each group. Then pool the counts and denominators separately for each training category. Pooling rates by simple addition or unweighted averaging would be incorrect because the group sizes differ.
Lower-complexity units. For lower training, the rate is \(45/900=0.05\), or 5%. For higher training, it is \(4/100=0.04\), or 4%.
Higher-complexity units. For lower training, the rate is \(20/100=0.20\), or 20%. For higher training, it is \(162/900=0.18\), or 18%. Within each complexity group, the higher-training rate is lower.
Aggregate lower-training rate. The pooled number of errors is \(45+20=65\), and the pooled number of claims is \(900+100=1{,}000\). The aggregate rate is \(65/1{,}000=6.5\%\).
Aggregate higher-training rate. The pooled number of errors is \(4+162=166\), and the pooled number of claims is \(100+900=1{,}000\). The aggregate rate is \(166/1{,}000=16.6\%\).
The aggregate rate is higher in the higher-training group, even though that group has a lower rate within each complexity category. The groups have very different mixes of claim complexity: most higher-training claims are in the higher-complexity category. This example shows why an overall association can mislead when the composition of groups differs. It does not prove that training reduces errors either; the figures are observational and may have other differences. The common mistake is reading the aggregate association as a causal effect without examining the group mix.
A Practical Interpretation Sequence
When a dashboard display seems to signal a change, use a consistent sequence before describing its meaning. This focuses the review on whether the displayed comparison is valid; deciding on a response to a confirmed breach is a separate task.
State exactly what is counted or measured, its unit, and whether it is a count, rate, amount, or ratio.
Confirm that periods, populations, exposure denominators, and measurement definitions are comparable. Recalculate rates when the exposure base changes.
Read all axes and scales. Identify the data cut-off, refresh timing, and any known information not yet reflected in the displayed value.
Confirm the applicable limit, its effective date, and that it uses the same basis as the measure. If any of these are unclear, do not infer breach status from appearance alone.
Describe an observed trend or association as such. Do not describe it as a cause unless the analysis supports a causal conclusion.
Common Interpretation Traps
- Comparing unlike scales. A steep-looking line on a truncated axis may represent a small absolute movement. Read the tick marks and calculate the change from values.
- Reporting only a count. If exposure changes, a count alone may not indicate whether frequency improved. State the denominator and rate where relevant.
- Ignoring the data cut-off. A display date does not necessarily equal the date through which underlying information is complete. Check freshness and reconcile material omitted information.
- Assuming a threshold. A value’s position on a chart does not establish compliance unless the applicable threshold is identified and comparable.
- Turning association into causation. A relationship between two displayed measures is not enough to conclude that one caused the other. Check timing, group composition, and other plausible explanations.
Check Your Understanding
For each case, identify the defensible reading from the stated figures.
- A hypothetical portfolio has 30 complaints among 15,000 policies in one quarter and 36 among 12,000 in the next. Calculate complaints per 1,000 policies for both quarters and state how the count and rate changed. (Answer: 2.0 per 1,000, then 3.0 per 1,000; count rose 20%, while rate rose 50%.)
- A dashboard compares a $12 million exposure increase with a $30 million premium increase. Starting values were $60 million exposure and $300 million premium. Calculate both percentage changes and the exposure-to-premium ratio before and after. (Answer: Exposure increased 20%; premium increased 10%; the ratio rose from 20% to 21.82%.)
- A loss dashboard shows $5.4 million against a $6.0 million limit, but the value is through March 31. A validated $0.8 million for the same period and basis is not included. What is the updated margin relative to the limit? (Answer: Updated amount is $6.2 million, or $0.2 million above the limit.)
- Two teams have these error counts: lower-complexity claims, 12 of 300 for group A and 16 of 400 for group B; higher-complexity claims, 40 of 200 for group A and 45 of 200 for group B. Calculate each within-category rate and state whether an aggregate comparison alone supports a causal claim. (Answer: Lower-complexity rates are 4% and 4%; higher-complexity rates are 20% and 22.5%; an aggregate association alone does not establish causation.)
- A chart’s loss ratio is 74% against a displayed 75% line, but the line is labelled “draft” and has no effective date. What can be concluded about limit status from this display alone? (Answer: Limit status cannot be determined until the applicable approved threshold and its effective date are confirmed.)