340 likes | 474 Views
Correlation Estimation for Property and Casualty Underwriting Losses. Fred Klinker Insurance Services Office, Inc. Mathematical vs. Physical Models for Correlation.
E N D
Correlation Estimation for Property and Casualty Underwriting Losses Fred Klinker Insurance Services Office, Inc.
Mathematical vs. Physical Models for Correlation • Mathematical models/ treatments: convenient and parsimonious ways of encoding what we know about correlation: simulation, Fast Fourier Transforms, copulas, etc. • Physical models for the drivers of correlation that therefore capture the structure: parameter uncertainty, natural and man-made catastrophes, mass torts.
Estimation of Correlation • For a number of lines of business, companies, and years, estimate expected losses or loss ratios • Measure deviations of actual ultimates from these expectations • Estimate correlations among these deviations as the correlations relevant to required capital
Issues • Deviations about long-term means not the most relevant, because they probably include a predictable component driven by known rate and price indices, trends, knowledge of current industry competitiveness, losses emerged to date, etc. • What is relevant are unpredictable deviations from expectations varying predictably over time.
Thought Experiment 1 • Rose Colored Glasses Insurance Company—will probably estimate larger correlations than a company that estimates its expected losses more accurately. • A cautionary conclusion—the correlations we estimate to some extent depend on how we estimate the expectations.
Thought Experiment 2: How We Might Like to Estimate Correlations • Mimic P&C industry real-time forecasting: rolling one-year-ahead forecasts based on what industry would have known compared to estimated actual ultimates • What we need: Multiple decade time series of loss ratios and predictors, one decade to calibrate the time series, plus more to check for time varying correlations • We lack the requisite data
An Alternative Calculation • One decade of data, no predictors • By LOB, a generalized additive model with main effects for company and year • Year effect captured by a non-parametric smoother • Fitted values respond to both earlier and later years, as opposed to one-year-ahead forecasts
A Question • Could the year smoother “forecast” even better than the best true one-year-ahead forecast, thereby understating deviations and covariances? • Perhaps, but probably not vastly better.
A Correlation Model Based on Parameter Uncertainty • From recent papers by Glenn Meyers, assuming frequency parameter uncertainty only:
where: • Lijk is annual aggregate ultimate loss for LOB i, company j, and year k. • δii´ is 1 if and only if i = i´ and 0 otherwise. Likewise for δjj´. • δGiGi´is 1 if and only if first and second LOBs are in the same covariance group, otherwise 0. • μi and σi are the mean and standard deviation of the severity distribution associated with LOB i. • Eijk = E[Lijk] • gi is the covariance generator associated with LOB i.
Recall the definition of covariance: Define the normalized deviation: Divide the original equation by EijkEi’j’k to find: .
Model for Expected Losses • Model loss ratios, then multiply by denominators. • By LOB, a generalized additive model with main effects for company and year • Year smoothing parameters chosen so that model responds to long term trends without responding much to individual year effects. • Loss ratio volatility declines significantly with increasing company size; a weighted model strongly recommended.
Appearance of roughly parallel lines supports main effects model. • At least for LOB 1, considerable correlated ups and downs from year to year. • After visual inspection of these graphs, would not be surprised to find greater correlation for LOB 1 than for LOB 2.
In each pairwise product, first and second deviations share common year and LOB 1, but different companies: cross-company, within-LOB correlation. • Pairwise products are not independent; many share a common first or second factor. • Regression line indicates modest positive correlation between first and second deviations, plus considerable noise. • A visual aid only; actual inference not based on this line.
For illustrative purposes only, ignores year effects; measures deviations against decade average, separately by company. • Ignoring long-term trends and patterns, probably predictable, inflates apparent correlations.
Bootstrap Estimates of Standard Errors • Pairwise products of deviations not independent; can’t use the usual sqrt(n) rule. • Don’t bootstrap on pairwise products directly; this destroys two-way structure of data on company and year. • Bootstrap on year, take all companies. Then bootstrap on company, take all years. • Combined standard error is square root of sum of squared standard errors due to year and company separately.
Correlation Parameter Estimates: LOB 1 Between companies: g Estimate: 0.0026 Standard error due to years: 0.0008 Standard error due to companies: 0.0009 Full standard error: 0.0012 Within company: c + g Estimate: 0.0226 Standard error due to years: 0.0048 Standard error due to companies: 0.0078 Full standard error: 0.0092
With respect to g, standard errors due to years and companies are comparable. • Estimate is more than twice the full standard error, so significant. • g is the variance of a frequency multiplier acting in common across companies within LOB 1. Square root of about .05: common underlying effects have the potential to drive frequencies across companies within LOB 1up or down by 5 or 10%. • Contagion is 0.02.
Correlation Parameter Estimates: LOB 2 Between companies: g Estimate: 0.0007 Standard error due to years: 0.0002 Standard error due to companies: 0.0003 Full standard error: 0.0004 Within company: c + g Estimate: 0.0090 Standard error due to years: 0.0007 Standard error due to companies: 0.0022 Full standard error: 0.0023
g just barely significant at two standard errors. • Both g and c smaller than for LOB 1, as expected from graphical evidence.
Correlation Parameter Estimates:LOB 1 vs. LOB 2 Between and within companies: g Estimate: 0.0005 Standard error due to years: 0.0005 Standard error due to companies: 0.0003 Full standard error: 0.0006
What is here labeled g is actually geometric average of gs for LOBs 1 and 2, if in the same covariance group, or 0 otherwise. • Parameter estimate not significantly different from 0: no statistical evidence that LOBs 1 and 2 are in the same covariance group.
Additional Observations • Parameter estimates are pooled across companies, not separate by company size, stock/ mutual, etc. • Correlation in the body of a multivariate distribution vs. “correlation in the tails”: correlation due to parameter uncertainty vs. correlation due to catastrophes.
What Else is in Appendix? • Expected losses derived from expected loss ratio models. We tested several denominators: premium, PPR, exposures. • Adjusted normalized deviations for degrees of freedom. • More thorough treatment of weights in all models: loss ratio, variance, other pairwise products of deviations. • Tested correlation model parameters for dependence on size of company: none found.
Bibliography • Glenn Meyers, “Estimating Between Line Correlations Generated by Parameter Uncertainty,” CAS Forum, Summer 1999. http://www.casact.org/pubs/forum/99sforum/99sf197.pdf • Glenn Meyers, Fred Klinker, and David Lalonde, “The Aggregation and Correlation of Insurance Exposure,” CAS Forum, Summer 2003. http://www.casact.org/pubs/forum/03sforum/03sf015.pdf