Why Monitor Credibility-Rated Premium Performance?
A credibility-rated loss cost is a forecast, not a guarantee that the next period’s realized losses will equal the forecast. Monitoring compares the predictions made before a period begins with the losses that emerge after that period is sufficiently mature. The purpose is to find persistent or patterned differences that may indicate miscalibration, a changing risk mix, an inadequate complement, or a credibility structure that deserves review.
A single unfavorable result is not, by itself, evidence that the method is wrong. Losses fluctuate, and the amount of fluctuation depends on the coverage, exposure volume, and loss distribution. A sound review therefore looks at both the size and the pattern of deviations: across risks, across periods, and across meaningful groups of risks. The methods here are aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
The comparison is meaningful only when the numerator and denominator have compatible definitions. Match losses to earned exposure, align coverage and valuation bases, and adjust historical amounts to a common cost level when the purpose is to compare periods. If the prediction is a pure loss cost, do not compare it with a premium amount that includes expenses or profit provisions unless those components have been explicitly reconciled.
Portfolio Measures and Their Interpretation
The A/P ratio gives a compact summary, but it does not show where a difference arose. Also calculate dollar residuals \(L_i-e_i p_i\), risk-level ratios \(L_i/(e_i p_i)\) when the denominator is positive, and results for relevant subgroups. Dollar residuals describe the size of the deviation; ratios describe it relative to predicted losses. A small risk with a very high ratio may have little portfolio impact, while a large risk with a modest ratio can matter substantially.
The portfolio ratio must give more weight to risks with larger predicted losses. This follows directly from the definitions. For risks with positive predicted losses, write each risk’s ratio as \(r_i=L_i/(e_i p_i)\). Then:
Thus the portfolio A/P is a weighted average of risk-level A/P ratios, with weights proportional to predicted losses. An unweighted average treats a small predicted amount as if it contributed as much as a large one and generally does not reproduce the portfolio result.
Worked Example: Reconciling Risk-Level and Portfolio Performance
A hypothetical insurer monitors two groups on a consistent coverage and cost-level basis. Group A has 100 earned exposure units, a predicted loss cost of $800 per unit, and realized losses of $84,000. Group B has 50 earned units, a predicted loss cost of $1,200 per unit, and realized losses of $54,000. Calculate each group’s A/P and the portfolio A/P.
Method. Compute predicted losses as exposure multiplied by predicted loss cost, then aggregate dollars before calculating the portfolio ratio.
For Group A, predicted losses are \(100\times \$800=\$80{,}000\). Its ratio is \(\$84{,}000/\$80{,}000=1.05\), or 105%.
For Group B, predicted losses are \(50\times \$1{,}200=\$60{,}000\). Its ratio is \(\$54{,}000/\$60{,}000=0.90\), or 90%.
Portfolio predicted losses are \(\$80{,}000+\$60{,}000=\$140{,}000\), and realized losses are \(\$84{,}000+\$54{,}000=\$138{,}000\). Therefore, \(\mathrm{A/P}=\$138{,}000/\$140{,}000=0.985714\ldots\), or 98.57% to two decimal places.
The weighted-average check is \((\$80{,}000(1.05)+\$60{,}000(0.90))/\$140{,}000=\$138{,}000/\$140{,}000\), the same result. The common mistake is to average 105% and 90% equally, obtaining 97.5%. That is not the portfolio A/P because the groups have different predicted losses.
Comparing Periods and Risk Mix
A period comparison should use forecasts that were available before each period’s experience was observed. For each period, calculate the period A/P from aggregate actual and predicted losses. Comparing period A/Ps is usually more informative than comparing total dollars when exposure changes. If the set of risks or their mix changes, also inspect results within stable segments or other relevant risk groupings. A portfolio ratio near 100% can conceal offsetting overprediction and underprediction among groups.
For a fair time comparison, confirm that loss development, trend, benefit or coverage terms, and exposure definitions are treated consistently. Otherwise, an apparent change in performance may reflect a change in measurement rather than a change in predictive accuracy. Separate periods can be compared on a common basis, but a historical forecast should not be revised after seeing the result merely to make it appear accurate.
Worked Example: Weighting Performance Across Periods
A hypothetical liability portfolio has two monitoring periods. In Period 1, it earns 80 exposure units at a forecast loss cost of $500 per unit and incurs $44,000 of losses. In Period 2, it earns 160 units at a forecast cost of $900 per unit and incurs $144,000. Assume the losses and forecasts have already been placed on a comparable basis. Find each period’s A/P and the combined A/P.
Method. Calculate expected losses separately for each period, then sum expected and actual dollars. Do not average period ratios unless their predicted losses are equal.
Period 1 predicted losses are \(80\times \$500=\$40{,}000\). Its A/P is \(\$44{,}000/\$40{,}000=1.10\), or 110%.
Period 2 predicted losses are \(160\times \$900=\$144{,}000\). Its A/P is \(\$144{,}000/\$144{,}000=1.00\), or 100%.
Combined predicted losses are \(\$40{,}000+\$144{,}000=\$184{,}000\). Combined actual losses are \(\$44{,}000+\$144{,}000=\$188{,}000\). Thus the combined A/P is \(\$188{,}000/\$184{,}000=1.021739\ldots\), or 102.17% to two decimal places.
The unweighted average of the period ratios is \((110\%+100\%)/2=105\%\), which is not the combined result. Period 2 has much greater predicted losses and therefore contributes more to the combined measure. The combined ratio summarizes overall performance; the period results still matter because the excess losses occurred in Period 1.
Interpreting Variation and Credibility Patterns
A/P is descriptive. It does not, by itself, say whether a deviation is statistically unusual or whether a credibility parameter should change. A useful review separates the observed difference from the uncertainty around it. For loss amounts with substantial severity variation, a Poisson model for claim counts is not automatically appropriate. Any standardized residual must be based on a defensible variance model for the quantity being monitored.
For example, suppose predicted claim counts are treated as fixed conditional means and observed counts are independent Poisson random variables. If a group has predicted count \(\lambda\), then its observed count \(N\) has mean and variance \(\lambda\). A standardized residual is \(z=(N-\lambda)/\sqrt{\lambda}\). Under the stated assumptions, a value near zero is consistent with the prediction, while a larger absolute value merits attention. This calculation is not a universal test for loss dollars, nor does one unusually large residual establish a persistent problem.
Worked Example: Avoiding Cancellation Across Credibility Groups
A hypothetical book has two risk groups. Forecasts were fixed before the period. Group H has predicted claim count 40 and observed count 48; Group L has predicted count 160 and observed count 144. Assume, for this illustration only, that counts are independent Poisson variables with those predicted counts as their conditional means. Calculate group residuals and the aggregate result.
Method. Use the Poisson standard deviation \(\sqrt{\lambda}\) within each group, and then calculate the aggregate count ratio and standardized residual.
For Group H, the count residual is \(48-40=8\), and \(z_H=8/\sqrt{40}=1.2649\) to four decimal places. For Group L, the residual is \(144-160=-16\), and \(z_L=-16/\sqrt{160}=-1.2649\) to four decimal places.
The aggregate predicted count is \(40+160=200\), and the observed count is \(48+144=192\). The aggregate A/P is \(192/200=0.96\), or 96%. Under independence, the aggregate count is Poisson with mean 200, so its standardized residual is \((192-200)/\sqrt{200}=-8/14.1421=-0.5657\) to four decimal places.
The aggregate is close to forecast, but the two group deviations point in opposite directions. The common mistake is to conclude from the 96% aggregate ratio that all groups are well calibrated. Persistent group-level patterns could merit investigation even when the total is close to 100%. These results alone do not establish why the differences occurred.
Using Monitoring to Review Credibility and Complements
A credibility-rated forecast commonly combines a risk’s observed experience with an experience complement. Monitoring can help assess whether the resulting predictions perform adequately for the intended population. The key is to examine performance across repeated periods and sensible groups, including groups with different amounts of individual experience or different complements. A recurring pattern among risks receiving low credibility may suggest that the complement does not represent those risks well. A pattern that worsens as experience accumulates may instead prompt a review of the credibility structure, data quality, or the stability of risk characteristics.
These are diagnostic hypotheses, not automatic conclusions. Before changing a credibility method, check whether the observed exposure and losses are correctly matched, whether loss development is mature, whether coverage or operating conditions changed, and whether a few severe losses dominate the result. A single period may be dominated by random fluctuation. Likewise, changing credibility weights solely because one group had an unfavorable result can overreact to noise.
Record the predicted loss cost, exposure basis, coverage, and assumptions before the monitoring period’s outcomes are known.
Match losses to exposure and coverage, apply consistent valuation and cost-level adjustments, and calculate predicted dollars.
Calculate A/P and dollar residuals overall, by period, and across relevant risk or credibility groups.
Look for repeated directional patterns, quantify variation only under a justified model, and investigate data or mix changes before proposing a rating adjustment.
Common Monitoring Traps
- Averaging ratios equally. The aggregate A/P is weighted by predicted losses, not by the number of groups or risks.
- Reading a near-100% total as proof of calibration. Overprediction in one segment can cancel underprediction in another.
- Comparing inconsistent bases. Different exposure definitions, coverage, loss maturity, or cost levels can produce a misleading period comparison.
- Treating one result as a trend. Random variation is expected; a credibility or complement change should be supported by a pattern and a review of alternative explanations.
- Applying a count variance to loss dollars. A Poisson count assumption does not supply the variance of aggregate severity losses without additional assumptions.
Check Your Understanding
For each problem, calculate the requested performance measure on the stated basis.
- A hypothetical segment earns 40 exposure units at a predicted loss cost of $700 per unit and has $29,120 in realized losses. Calculate predicted losses and A/P. (Answer: Predicted losses = $28,000; A/P = 1.0400, or 104.00%.)
- Group A has predicted losses of $30,000 and A/P of 1.10. Group B has predicted losses of $90,000 and A/P of 0.95. Calculate combined A/P. (Answer: \((\$30,000(1.10)+\$90,000(0.95))/\$120,000=0.9875\), or 98.75%.)
- Period 1 has predicted losses of $50,000 and actual losses of $55,000. Period 2 has predicted losses of $150,000 and actual losses of $144,000. Calculate each period’s A/P and the combined A/P. (Answer: Period 1 = 110%; Period 2 = 96%; combined = \(\$199,000/\$200,000=99.50\%\).)
- Under an independent Poisson count model, a group has predicted count 25 and observed count 31. Calculate the standardized residual to four decimal places. (Answer: \((31-25)/\sqrt{25}=1.2000\).)
- Two groups have predicted counts 25 and 75, and observed counts 31 and 69, respectively. Under independent Poisson assumptions, calculate aggregate A/P and aggregate standardized residual. (Answer: A/P = \(100/100=1.0000\); residual \(=(100-100)/\sqrt{100}=0.0000\). The offsetting group deviations remain worth monitoring.)