Credibility theory and general insurance pricing

In short

Credibility theory tells you how much weight to put on a risk's own claims experience and how much on a wider book or manual rate, when the risk's own data are too thin to be trusted alone. The credibility premium is always a weighted average, Z × (own experience) + (1 − Z) × (collateral estimate), and the whole subject is about choosing Z defensibly. It underpins experience rating of fleets, group health and large commercial accounts, and the smoothing of small rating cells inside motor and health tariffs built with GLMs. It is examined in IAI/IFoA CS1 (Bayesian statistics and credibility theory, EBCT Models 1 and 2) and applied in SP8 General Insurance Pricing, and in B.Sc Actuarial Science papers on general insurance; employers who test it include ICICI Lombard, Bajaj Allianz, HDFC ERGO, Tata AIG, GIC Re, Swiss Re/Munich Re India, and the actuarial and pricing teams of broking and consulting firms.

Explained

Why credibility is needed at all

• A single garment factory's fire claims over 3 years, or one taxi fleet's own-damage claims, carry huge sampling error: observed pure premium swings wildly even if the true risk is unchanged.
• Ignoring the risk's own data (pure exposure rating off the manual tariff) is stable but unresponsive and loses good risks to competitors.
• Using only the risk's own data (pure experience rating) is responsive but unstable and invites selection against you.
• Credibility is the formal compromise: shrink the noisy individual estimate towards a stable collective estimate, by an amount that depends on how noisy the individual data are and how genuinely different risks are from one another.
• Vocabulary: pure premium / risk premium = expected claim cost per unit exposure; office / gross premium = risk premium plus expenses, commission, reinsurance cost, profit and cost of capital; exposure = vehicle-years, sum insured × years, wage roll, turnover, bed-days.

Limited fluctuation (classical, American) credibility

• Question asked: how much data are needed before the risk's own experience can be used *in full*?
• Full credibility standard: the standard is met if the observed quantity is within ±k of its mean with probability p. For claim *numbers* under a Poisson assumption, the standard is λ_full = (z/k)², z = the standard normal point for (1+p)/2.
• Classic benchmark: p = 90% (z = 1.645), k = 5% → λ_full = 1,082 claims. With p = 95% (z = 1.96), k = 5% → 1,537 claims.
• For aggregate claims the standard inflates by the severity variability: n_full = (z/k)² (1 + CV_Y²), where CV_Y is the coefficient of variation of individual claim size.
• Partial credibility (square-root rule): Z = √(n / n_full), capped at 1, where n is the expected (or observed) number of claims.
• Advantages: simple, transparent, needs only the risk's own volume; still used in US workers' compensation style plans and in quick commercial-lines experience-rating worksheets.
• Limitations: the ±k target and p are arbitrary; it never uses information about how much risks actually differ (no VHM), so it under-credits homogeneous books and over-credits heterogeneous ones; the square-root rule has no probabilistic justification.

Greatest accuracy credibility — Bühlmann and Bühlmann-Straub

• Set-up: each risk i has an unknown risk parameter θ_i drawn from a prior; given θ_i, yearly results X_i1 … X_in are i.i.d. with mean μ(θ_i) and variance s²(θ_i).
• Three structural quantities:
• μ = E[μ(θ)] — the collective (overall) mean.
• EPV = s² = E[s²(θ)] — expected process variance, the *within*-risk noise.
• VHM = a = Var[μ(θ)] — variance of the hypothetical means, the *between*-risk heterogeneity.
• Bühlmann (EBCT Model 1): Z = n/(n + k) with k = EPV/VHM = s²/a; credibility premium = Z X̄_i + (1 − Z) μ. This is the best *linear* estimator of μ(θ_i) in the least-squares sense.
• Behaviour of Z: increases with n (more own data), increases with VHM (risks really differ, so own data are informative), decreases with EPV (own data are noisy). Z → 1 as n → ∞, and Z = 0 if all risks are identical (a = 0).
• Bühlmann-Straub (EBCT Model 2): risks have unequal exposure P_ij; Y_ij = claims per unit exposure. Then Z_i = P_i/(P_i + k), P_i = Σ_j P_ij. This is the version used in practice, because fleets and group schemes differ in size.
• Empirical Bayes: μ, s² and a are unknown, so they are estimated from the data of the whole collective (N risks × n years) by unbiased moment estimators — hence "empirical Bayes credibility theory (EBCT)".
• Important property: the credibility estimates are balanced — the exposure-weighted average of the credibility premiums equals the overall mean, so the exercise redistributes premium without changing the total (useful when a regulator or a reinsurer asks whether the plan is revenue-neutral).

Bayesian credibility and exact credibility

• Full Bayesian answer: the premium is the posterior mean E[μ(θ) | data]. For some conjugate pairs this posterior mean is *exactly* linear in the data — then Bayes and Bühlmann coincide, called exact credibility.
• Poisson–Gamma: claim numbers X | λ ~ Poisson(λ) per unit exposure, λ ~ Gamma(α, β) (β = rate). Posterior is Gamma(α + Σx, β + n); posterior mean = (α + Σx)/(β + n) = Z X̄ + (1 − Z)(α/β) with Z = n/(n + β).
• Binomial–Beta (claim propensity) and Normal–Normal (claim amounts, known variance) are the other examined conjugate cases; the Normal–Normal gives Z = n/(n + σ²/τ²).
• Loss functions: under squared-error loss the Bayes estimate is the posterior mean; under absolute error it is the posterior median; under all-or-nothing loss it is the mode. Exam questions often test this distinction.
• Advantages: uses the whole posterior, handles small samples honestly, extends to hierarchical models fitted by MCMC. Limitations: needs a prior, is not linear in general, and is harder to explain to an underwriter or a board.

How a general insurance price is actually built in India

1. Data: earned exposure, claims incurred (paid + outstanding), development to ultimate, large-loss capping and spreading, catastrophe loading, trending of frequency and severity to the future policy period, "on-levelling" past premiums to current rates.
2. Burning cost / pure premium method: risk premium = ultimate claims ÷ exposure. Loss ratio method: new rate = current rate × (experience loss ratio ÷ target loss ratio) — used when exposure is a poor measure.
3. Risk classification: rating factors (vehicle cubic capacity, age, geographical zone, IDV, occupation, age of insured, sum insured band, construction class, claims history).
4. Multiplicative tariff by GLM: frequency as Poisson/quasi-Poisson with log link and log(exposure) offset; severity as Gamma or inverse-Gaussian with log link; or a single Tweedie GLM on pure premium. Credibility then smooths thin cells (small districts, rare vehicle classes) towards the parent level — an explicit link between the two halves of this topic.
5. Experience rating on top of the tariff for large accounts: credibility-weighted own loss ratio, plus no-claim bonus on motor own-damage (IRDAI's standard slabs 20/25/35/45/50% after 1 to 5 consecutive claim-free years), deductibles and profit-sharing.
6. Loadings: commission, management expenses, reinsurance cost, GST collected separately, and a margin for profit and capital.
7. Regulatory frame: the Insurance Act, 1938 and the IRDAI Act, 1999; general insurance was de-tariffed from 1 January 2007, so fire, engineering, motor own-damage and most commercial rates are filed by insurers rather than fixed by a tariff; product filing now runs under the IRDAI (Insurance Products) Regulations, 2024 with a largely Use-and-File route for general insurance. Motor third-party premium remains administratively fixed and is notified by the Ministry of Road Transport and Highways in consultation with IRDAI (rates have been carried over in some years — check the current notification before quoting a figure). The Insurance Information Bureau of India (IIB) publishes industry-wide motor and health loss-cost data that insurers use as the collateral estimate in exactly the credibility sense. Pricing must be signed off by the Appointed Actuary; IRDAI's file-and-use/use-and-file discipline requires that rates be justified by data and not discriminate unfairly.
8. GST: general insurance such as motor and commercial lines is taxed at 18%; individual life and health insurance policies were exempted from GST from 22 September 2025, so no GST is loaded on those retail premiums.

Advantages, limitations and comparisons

• Credibility is not a forecast of a risk's future claims; it is the best compromise estimate of its *expected* cost.
• Strengths: stabilises rates, defensible to regulators, revenue-neutral, improves on both extremes, and needs no distributional assumption in the Bühlmann form (only first two moments).
• Weaknesses: assumes risks are exchangeable draws from one prior; ignores rating factors unless you apply it to GLM residuals; EPV/VHM estimates are themselves unstable when N or n is small; negative â estimates must be floored at zero (then Z = 0); linear form cannot capture a risk that has genuinely changed (a fleet that fitted telematics).

Key points

• Credibility premium - Z × own experience + (1 − Z) × collateral estimate; Z ∈ [0, 1].
• Full credibility standard - the data volume at which Z = 1; λ_full = (z/k)² for Poisson counts, 1,082 claims at p = 90%, k = 5%.
• Partial credibility - Z = √(n/n_full) in the limited-fluctuation approach.
• EPV (s²) - expected process variance, the random noise within a single risk; large EPV lowers Z.
• VHM (a) - variance of hypothetical means, genuine difference between risks; large VHM raises Z.
• Bühlmann k - k = EPV/VHM, measured in years of exposure; Z = n/(n + k).
• Bühlmann-Straub - the exposure-weighted version, Z = P/(P + k), used for unequal-sized risks.
• EBCT Model 1 vs 2 - Model 1 assumes equal exposure per risk-year; Model 2 allows differing exposures.
• Empirical Bayes - estimating μ, EPV and VHM from the collective's own data instead of assuming them.
• Exact credibility - Poisson-Gamma, Binomial-Beta, Normal-Normal: Bayes posterior mean equals the linear credibility premium.
• Balance property - exposure-weighted credibility premiums reproduce the overall mean; the plan redistributes, not inflates.
• Pure premium - frequency × severity per unit exposure; the base of every general insurance rate.
• Loss ratio method - adjust the existing rate by experience LR ÷ target LR; used where exposure is unreliable.
• Exposure vs experience rating - manual tariff versus the risk's own record; credibility blends them.
• GLM tariff - multiplicative rates from Poisson frequency and Gamma severity models, or one Tweedie model, with credibility smoothing thin cells.
• De-tariffing - Indian general insurance rates have been free since 1 January 2007, except motor third-party, which is notified administratively.
• NCB - IRDAI's motor own-damage discount ladder 20/25/35/45/50% for 1-5 claim-free years; a crude, regulator-fixed experience rating.
• IIB - the industry data pool whose loss costs serve as the collective estimate in Indian motor and health pricing.

Formulas and examples

Example 1 — Limited fluctuation credibility. A transport fleet is expected to produce 400 claims next year. Standard: p = 90%, k = 5% → λ_full = (1.645/0.05)² = 32.9² = 1,082.41, so 1,082 claims.

• Z = √(400/1,082.41) = √0.3696 = 0.608.
• Fleet's own indicated pure premium ₹3,500 per vehicle-year; manual rate ₹3,000.
• Credibility premium = 0.608 × 3,500 + 0.392 × 3,000 = 2,128 + 1,176 = ₹3,304 per vehicle-year.

Example 2 — Bühlmann with given structure. EPV = 20,00,000 (₹²), VHM = 5,00,000 (₹²), so k = 20,00,000/5,00,000 = 4 years. A group health scheme has 6 years of data, own mean claim cost ₹5,200 per member, book mean ₹4,500.

• Z = 6/(6 + 4) = 0.60.
• Premium = 0.60 × 5,200 + 0.40 × 4,500 = 3,120 + 1,800 = ₹4,920 per member.
• If only 2 years were available: Z = 2/6 = 0.333 → 0.333 × 5,200 + 0.667 × 4,500 = 1,733 + 3,000 = ₹4,733 — less movement from the book rate, as expected.

Example 3 — EBCT Model 1 from raw data. Three taxi fleets, claim frequency per vehicle-year for 2 years:

• μ̂ = (0.12 + 0.23 + 0.35)/3 = 0.70/3 = 0.23333.
• ŝ² (EPV) = (0.0008 + 0.0018 + 0.0050)/3 = 0.0076/3 = 0.0025333.
• Variance of the three means: deviations −0.11333, −0.00333, +0.11667 → squares 0.0128444 + 0.0000111 + 0.0136111 = 0.0264667; ÷ (3 − 1) = 0.0132333.
• â (VHM) = 0.0132333 − 0.0025333/2 = 0.0132333 − 0.0012667 = 0.0119667.
• k = 0.0025333/0.0119667 = 0.2117; Z = 2/(2 + 0.2117) = 0.9043.
• Credibility frequencies: A = 0.9043 × 0.12 + 0.0957 × 0.23333 = 0.10851 + 0.02233 = 0.1308; B = 0.9043 × 0.23 + 0.02233 = 0.2303; C = 0.9043 × 0.35 + 0.02233 = 0.3388.
• Check the balance property: (0.1308 + 0.2303 + 0.3388)/3 = 0.6999/3 ≈ 0.2333 = μ̂ ✓. Z is high because the fleets clearly differ (large VHM) relative to year-to-year noise.

Example 4 — Bayesian Poisson-Gamma (exact credibility). Prior on a factory's fire claim rate: Gamma(α = 4, β = 20), prior mean 4/20 = 0.20 claims per unit-year. Observed: 15 claims in 50 unit-years.

• Posterior = Gamma(4 + 15, 20 + 50) = Gamma(19, 70); posterior mean = 19/70 = 0.27143.
• Credibility form: Z = 50/(50 + 20) = 0.71429, X̄ = 15/50 = 0.30 → 0.71429 × 0.30 + 0.28571 × 0.20 = 0.21429 + 0.05714 = 0.27143 ✓ identical, which is what exact credibility means.

Example 5 — Private car premium build-up. Frequency 0.085 claims per vehicle-year, average claim ₹32,000; fixed expense ₹150 per policy; commission plus management plus profit = 30% of gross; NCB 25%; GST 18%.

• Risk premium = 0.085 × 32,000 = ₹2,720.
• Gross before discount = (2,720 + 150)/(1 − 0.30) = 2,870/0.70 = ₹4,100.
• After 25% NCB (applied to own-damage premium in motor practice) = 4,100 × 0.75 = ₹3,075.
• Premium payable with 18% GST = 3,075 × 1.18 = ₹3,628.50.
• Implied loss ratio on the discounted premium = 2,720/3,075 = 88.5% — i.e. the NCB has consumed the expense and profit margin, exactly the test an actuary runs before agreeing a discount ladder.

Where it is used

• Motor insurance pricing at ICICI Lombard, Bajaj Allianz, HDFC ERGO, Tata AIG, New India Assurance: GLM tariffs on IIB and own data, with credibility smoothing of RTO-level and vehicle-model-level relativities, plus NCB and fleet experience rating.
• Group health and group personal accident: renewal pricing of corporate schemes blends the scheme's own burning cost with the book rate by Bühlmann-Straub on member-years — the single most common commercial use in India.
• Commercial lines after de-tariffing: fire, engineering and marine cargo accounts where only a few years of a client's own losses exist; large-loss capping plus credibility on the capped experience.
• Reinsurance and GIC Re treaty pricing: experience rating of a cedant's layer blended with exposure rating from a market curve, weighted by credibility judgement.
• Crop and parametric covers (PMFBY-type schemes): district yield or weather experience credibility-weighted towards state or agro-climatic zone averages where a district's series is short.
• Job roles: actuarial pricing analyst, Appointed Actuary's team, pricing/portfolio management, data science teams building Tweedie and GBM models, reinsurance pricing, and IRDAI/IIB analytics roles; examined in IAI CS1 and SP8 and in placement tests for actuarial trainees.

Interview questions

• Q: In one sentence, what is credibility theory? A: A method for weighting a risk's own claims experience against a broader collective estimate, so that the premium is responsive without being unstable.
• Q: What do EPV and VHM mean, and how do they drive Z? A: EPV is the random year-to-year variation within a risk and VHM is the genuine variation between risks; Z = n/(n + EPV/VHM), so Z rises with exposure and with VHM and falls with EPV.
• Q: If all risks in a portfolio were identical, what would Z be? A: Zero, because VHM = 0 means a risk's own data tell you nothing beyond the collective mean, so everyone pays the book rate.
• Q: Why is 1,082 a famous number? A: It is the full credibility standard in claim counts for 90% probability of being within ±5% under a Poisson assumption, (1.645/0.05)².
• Q: What is the main criticism of limited fluctuation credibility? A: The confidence level and tolerance are arbitrary and it ignores how much risks actually differ, so Z depends only on your own volume, not on the information content of the data.
• Q: When do Bayesian and Bühlmann credibility give the same answer? A: Under exact credibility — conjugate pairs such as Poisson-Gamma, Binomial-Beta and Normal-Normal, where the posterior mean is already linear in the observations.
• Q: Difference between EBCT Model 1 and Model 2? A: Model 1 assumes every risk-year has the same exposure; Model 2 (Bühlmann-Straub) works with claims per unit exposure and weights by exposure, which is what real fleets and group schemes need.
• Q: What do you do if the estimate of VHM comes out negative? A: Set it to zero (so Z = 0 and everyone gets the collective mean) and report it — a negative estimate signals that between-risk differences are indistinguishable from sampling noise.
• Q: How does credibility sit alongside a GLM tariff? A: The GLM handles rating factors; credibility is applied to thin levels of a factor or to residual experience of an individual account, shrinking them towards the fitted or parent value.
• Q: Is motor insurance pricing free in India? A: Own-damage and most commercial lines have been de-tariffed since 1 January 2007 and filed by insurers, but motor third-party premium is notified administratively by MoRTH in consultation with IRDAI.
• Q: Why is the balance property useful commercially? A: Because the exposure-weighted credibility premiums reproduce the overall mean, you can show management or the regulator that the experience-rating plan redistributes premium rather than changing total income.
• Q: A fleet fitted telematics last year and its frequency halved. Would you credibility-weight the last five years? A: No — credibility assumes a stable risk parameter; a known structural change means re-weighting towards the post-change period or re-rating off the new exposure characteristics.

Common mistakes

• Treating Z as a probability or a confidence level rather than a weight.
• Inverting k: writing Z = n/(n + VHM/EPV) instead of k = EPV/VHM.
• Using the Poisson full-credibility standard for aggregate claims and forgetting the (1 + CV_Y²) severity inflation.
• Applying EBCT Model 1 to risks of very different size — the small fleet gets the same Z as the large one, which is wrong.
• Using the within-risk variance divided by n as the VHM estimate; â is the variance of the risk means *minus* ŝ²/n.
• Mixing up exposure units between numerator and denominator (claims per policy vs per vehicle-year vs per ₹1,000 sum insured), or loading GST before the NCB discount.
• Quoting a Bayes estimate as a "credibility premium" without checking the loss function, and quoting old tariff rates as if general insurance were still tariffed.

Related topics

Insurance basics · Risk theory and loss distributions · Enterprise risk management and solvency under IRDAI · Generalised linear models and ratemaking · Claims reserving and IBNR · Bayesian statistics and conjugate priors