Experience Rating: Advanced Credibility and NCD Systems

No-Claims Discount Systems

No-claims discount (NCD) or bonus-malus systems adjust premiums based on claim history. These systems are widely used in auto insurance across many countries. The basic structure assigns each policyholder to a level (state) that determines their premium discount or surcharge. Policyholders move between levels based on whether they file claims during each policy period. A claim-free year typically moves the policyholder to a more favorable level (higher discount), while filing a claim moves them to a less favorable level (lower discount or surcharge).

The system is naturally modeled as a Markov chain where the states represent discount levels and the transition probabilities depend on the claim frequency. For a policyholder with annual claim probability q, the probability of moving up one level (claim-free) is approximately (1 minus q) and the probability of dropping back (one or more claims) is approximately q (simplified for the case where the transition depends only on whether any claim was filed, not the number of claims).

Markov Chain Analysis of NCD Systems

The transition matrix P of the NCD Markov chain has entries p_ij representing the probability of moving from level i to level j in one period. For a system with L levels, this is an L-by-L matrix. The transition rules are deterministic given the claim outcome: if no claim, move to level min(i+1, L); if a claim, move to level max(i minus k, 1), where k is the penalty (number of levels dropped per claim). The randomness comes from the claim frequency, which determines the probability of each outcome.

The stationary distribution pi satisfies pi = pi * P and sum of pi_i = 1. It gives the long-run proportion of time a policyholder with a given claim frequency spends at each level. The stationary distribution is important for pricing: the average premium collected from a policyholder in the long run equals the sum of pi_i * r_i * base premium, where r_i is the relativity at level i. For the NCD system to be financially sound, this weighted average premium should cover the expected losses plus expenses.

Optimal (Bayesian) NCD Systems

An optimal NCD system sets the premium at each level equal to the expected claim frequency given the claim history implied by occupying that level. This requires Bayesian updating. Assume the population has heterogeneous claim frequencies, with each policyholder's frequency lambda following a prior distribution (typically Gamma with shape alpha and rate beta). Given that a policyholder has filed k claims in n years, the posterior distribution of lambda is Gamma with shape alpha + k and rate beta + n.

The predictive expected frequency is E[lambda | k claims in n years] = (alpha + k) / (beta + n). The Bayesian premium is proportional to this expected frequency. The key insight is that two policyholders at the same NCD level may have different claim histories (different paths to the same level), and a truly optimal system would charge different premiums based on the full history, not just the current level. The standard NCD system approximates this by using the level as a sufficient summary of the relevant history.

Efficiency of NCD Systems

The efficiency of an NCD system measures how well it separates good risks from bad risks. One measure compares the variance of premiums under the NCD system to the variance of the true expected losses across policyholders. A perfectly efficient system would assign each policyholder a premium exactly proportional to their true expected frequency, achieving maximum variance in premiums. A system that charges everyone the same premium has zero efficiency.

Formally, the efficiency can be defined as Var(NCD premium) / Var(true expected loss), where the variance is taken over the population of policyholders in stationarity. Efficiency depends on the number of levels, the transition rules, and the heterogeneity of the insured population. More levels generally increase efficiency, but with diminishing returns. The penalty structure (how many levels a claim drops you) also affects efficiency: harsher penalties increase discrimination but may encourage claim suppression.

Advanced Credibility in Experience Rating

Beyond standard Buhlmann credibility, advanced methods accommodate features of real experience rating data. Serial correlation credibility addresses the case where observations for a given risk are correlated over time (for example, because a risk's underlying frequency drifts rather than being constant). The standard Buhlmann model assumes observations are conditionally independent given the risk parameter, which may not hold if the risk's characteristics are changing.

Trend credibility extends the framework to handle systematic trends in a risk's experience. If losses are trending upward or downward, the standard credibility estimate (which is a level average of past experience) may lag behind the true current risk level. The Hachemeister regression credibility model addresses this by fitting a credibility-weighted regression to each risk's data, allowing the estimated trend to vary by risk. This produces predictions that extrapolate each risk's individual trend, blended with the collective trend.

Bayesian Updating in Practice

Bayesian updating for bonus-malus systems uses conjugate prior-posterior pairs to maintain analytical tractability. The Poisson-Gamma conjugate pair is the most important: if the claim frequency lambda has a Gamma(alpha, beta) prior and the policyholder generates k claims in n exposure units, the posterior is Gamma(alpha + k, beta + n). The posterior mean (alpha + k) / (beta + n) is the Bayesian estimate of the claim frequency, and the predictive premium is proportional to this quantity.

The credibility interpretation connects Bayesian and frequentist perspectives. The posterior mean can be written as Z * (k/n) + (1 minus Z) * (alpha/beta), where Z = n / (n + beta) is the Buhlmann credibility factor and alpha/beta is the prior mean. This shows that the Bayesian update is exactly a credibility-weighted blend of the individual experience (k/n) and the prior expectation (alpha/beta), with credibility increasing as more data accumulates.

Common Pitfalls

Students often confuse the transition matrix for a specific policyholder (which depends on their individual claim frequency) with the population-level transition matrix (which averages over all frequencies in the population). The stationary distribution depends on which transition matrix is used. For pricing purposes, you typically need the stationary distribution for a specific claim frequency to determine what premium that type of policyholder pays on average in the long run.

Another common error is forgetting that the NCD level is not a sufficient statistic for the policyholder's entire claim history. Two policyholders at the same level may have very different numbers of years of experience and total claims. The optimal Bayesian premium accounts for the full history, while the standard NCD system uses only the current level. This gap between the NCD premium and the optimal premium is the source of inefficiency in the system.