Agricultural Statistics & Data Science · Time series, forecasting and modelling · also Actuarial Statistics, Statistics of Agricultural Insurance
The numbers behind crop and livestock insurance: loss distributions, premium principles, credibility theory and ruin probability, then their agricultural use in area-yield index design, threshold yield and indemnity levels, premium rating from yield and weather histories, basis risk in weather-index insurance, and the sampling of crop-cutting experiments for claims. Scheme-level insurance management is in the Agribusiness track.
In the exams: ICAR MSc elective (actuarial statistics); light in national exams; crop-insurance analytics roles (applied)
Crop and livestock insurance stands or falls on numbers. A premium rate that is too low bankrupts the insurer or the public budget that backs it; one that is too high drives farmers away; a threshold yield set from a short, trending yield series pays claims in normal years and fails in bad ones; a weather index that does not track losses leaves farmers unpaid when their crop fails. Actuarial statistics provides the methods for these numbers: models of claim frequency and size, principles for setting premiums with a margin for risk, credibility methods for blending a small unit's own experience with wider data, and ruin theory for the reserves an insurer needs.
This subject treats those general methods first and then applies them to agriculture: area-yield index insurance with its insurance units and threshold yields, detrending of historical yields and premium rating, weather-index insurance and basis risk, livestock mortality insurance with life-table methods, the number of crop-cutting experiments needed to settle claims with acceptable precision, and the growing use of remote sensing and technology in yield estimation for claims. The emphasis throughout is statistical. The economics of insurance, the history of crop-insurance schemes and the management of national programmes such as India's crop insurance schemes are treated in the Agribusiness track's Agri Risk Management and Insurance subject; regression and trend fitting are in Linear Models and Regression; and crop-cutting surveys are in Crop Estimation Surveys.
In national examinations the subject appears lightly; it is an elective in master's programmes in agricultural statistics and a working toolkit in insurance companies, reinsurers, government insurance cells and agricultural finance institutions. Rules of specific schemes (indemnity levels, the years used for threshold yields, minimum numbers of crop-cutting experiments) change with official guidelines, so this manual explains the statistical logic and asks the reader to confirm current scheme rules from official sources.
Insurance losses are usually modelled in two parts: the number of claims in a period (frequency) and the size of each claim (severity). Common frequency models are the Poisson distribution, whose variance equals its mean; the negative binomial, whose variance exceeds its mean and which suits portfolios where risk varies between policyholders (overdispersion); and the binomial, for a fixed number of exposures such as animals insured. Common severity models are the exponential, gamma, lognormal, Weibull and Pareto distributions, ordered roughly from light to heavy tails; the Pareto and lognormal suit losses where a few very large claims dominate. The aggregate loss in a period is S = X(1) + ... + X(N), a compound distribution. If N and the X(i) are independent and the X(i) identically distributed, E(S) = E(N) E(X) and Var(S) = E(N) Var(X) + Var(N) [E(X)] squared. For a compound Poisson with mean lambda, Var(S) = lambda E(X squared).
Parameters are estimated by the method of moments, which is quick, or by maximum likelihood, which is more efficient and handles incomplete data properly. Insurance data are often incomplete in characteristic ways. With a deductible d, losses below d are not reported, so the data are left-truncated: the likelihood for each observed loss uses the density divided by the probability of exceeding d. With a policy limit u, losses above u are recorded only as u, so the data are right-censored: those observations contribute the probability of exceeding u. Ignoring truncation or censoring biases the fitted distribution. Goodness of fit is assessed by comparing fitted and empirical distribution functions (the Kolmogorov-Smirnov statistic, the Anderson-Darling statistic, which gives more weight to the tails), chi-square tests on grouped data, quantile-quantile plots and information criteria such as AIC and BIC for comparing families. Because the tail drives large-loss costs, tail fit matters more than fit near the centre.
Agricultural losses differ from typical motor or fire losses in two ways. First, many policyholders suffer losses at once: a drought, flood or pest outbreak affects a whole region, so losses are strongly correlated (systemic risk), and the aggregate loss is far more variable than independence would imply. Second, the loss for an area-yield or weather-index policy is a function of an underlying continuous variable (area yield, rainfall), so the natural approach is to model that variable's distribution and compute the payout from it, rather than to fit a severity distribution to past claims. Loss experience is often summarised by the loss cost ratio, indemnities divided by liability (sum insured), for each unit and year; its distribution typically has many zeros (years with no claim) and a long right tail, which zero-inflated or mixture models describe. The burning cost, the average historical loss cost ratio, is the simplest estimate of the pure premium rate.
The pure (risk) premium is the expected loss, E(S). An insurer that charged only this would, by the law of large numbers, break even on average across a large, independent portfolio, but would face ruin with high probability in any finite period because losses fluctuate. The premium therefore includes a risk loading for this variability, together with loadings for administrative expenses, the cost of capital held against losses, reinsurance and profit. In crop insurance, the premium rate is usually expressed as a percentage of the sum insured; the pure premium rate is the expected loss cost ratio. When expenses are a proportion e of the gross premium, gross premium = (pure premium + risk loading) / (1 - e). Where governments subsidise premiums, the actuarially fair rate and the farmer's share are separate quantities, and the subsidy is the difference.
A premium principle is a rule that maps a loss distribution to a premium. The expected value principle charges P = (1 + theta) E(X), with a safety loading theta; it is simple but ignores how variable the loss is. The variance principle charges P = E(X) + alpha Var(X), and the standard deviation principle P = E(X) + beta SD(X); both load more heavily for more variable risks. The exponential (zero-utility) principle charges P = (1/a) ln E[exp(aX)], derived from an insurer with exponential utility and risk aversion a; it is additive for independent risks. The percentile principle sets P so that the probability of loss exceeding P is small. Desirable properties include non-negative loading (P at least E(X)), no rip-off (P no more than the maximum possible loss), additivity for independent risks, and consistency (adding a constant to the loss adds it to the premium). No principle has every property; the expected value principle is not sensitive to variability, while the variance principle is not consistent with changes of currency unit.
The relative variability of a portfolio's total loss shrinks as more independent risks are added: if each of n independent policies has loss mean mu and standard deviation sigma, the coefficient of variation of the total is sigma/(mu sqrt(n)), falling towards zero as n grows. This is why a risk loading per policy can be small for large portfolios of independent risks. When losses are correlated with a common pairwise correlation rho, the variance of the total is n sigma squared [1 + (n - 1) rho], and as n grows the coefficient of variation approaches sigma sqrt(rho)/mu rather than zero: part of the risk cannot be diversified away within the portfolio. In crop insurance, correlations between units in the same agro-climatic zone can be high in drought years, so geographical spread across zones, reinsurance and government backing are the means of absorbing the undiversifiable part. Estimating these correlations, and especially their behaviour in extreme years, is a central statistical task in pricing catastrophe loadings.
In agricultural insurance, risk loadings must reflect systemic risk, which the law of large numbers does not diversify within a region; uncertainty in the estimated pure premium itself, which is large when only 10 to 20 years of yield data exist; and catastrophe years, which may be absent from a short record. Loadings for parameter uncertainty can be based on the standard error of the estimated loss cost, or on credibility weighting with wider data (Chapter 3). Catastrophe loadings may use longer weather records, simulation from crop models, or the cost of reinsurance. Rates are often capped or smoothed by policy, so that the premium charged and the actuarially fair premium diverge; the statistician's job is to state the fair rate clearly and quantify the gap, which is then a matter of policy.
An insurance unit, such as a village, block or herd, has its own loss history, but it is short and noisy. A wider group (the district or state) has more data but may not reflect the unit's own risk. Credibility theory blends the two: credibility premium = Z x (unit's own mean loss) + (1 - Z) x (group or manual rate), where the credibility factor Z, between 0 and 1, grows with the amount and reliability of the unit's own data. A unit with a long, stable history gets a rate close to its own experience; a unit with two years of data gets a rate close to the group's. This is the actuarial form of shrinkage, closely related to empirical Bayes estimation and to best linear unbiased prediction (BLUP) of random effects.
The older limited-fluctuation (classical) approach asks how much data is needed for the unit's experience to be given full weight (Z = 1). The standard requires that the observed mean be within a proportion k of its expected value with probability p. For claim counts that are Poisson, the number of claims needed for full credibility is n(F) = (z/k) squared, where z is the standard normal quantile for (1 + p)/2; with p = 0.90 and k = 0.05, z = 1.645 and n(F) = (1.645/0.05) squared, about 1,082 claims. Standards for aggregate losses also include the coefficient of variation of claim size. With fewer data, partial credibility is given by the square-root rule, Z = square root of (n/n(F)), capped at 1. The approach is simple and widely used but has no strong theoretical justification for the square-root rule.
Suppose a block has five years of loss cost ratios averaging 12 per cent of liability, while the district rate is 8 per cent. From the experience of many blocks in the state, the expected process variance (year-to-year variance within a block) is estimated as 0.0100 and the variance of the hypothetical means (between blocks) as 0.0005, both in squared loss-cost-ratio units. Then K = EPV/VHM = 20, and with n = 5 years Z = 5/(5 + 20) = 0.2. The credibility rate is 0.2 x 12 + 0.8 x 8 = 8.8 per cent: the block's own history moves its rate only a little above the district's, because five noisy years carry little information relative to the real but modest differences between blocks. With 20 years, Z would be 20/40 = 0.5 and the rate 10 per cent. The example shows why short yield histories should not be rated on their own, and why the variance components behind Z must themselves be estimated from a wide enough base. In practice the estimated VHM can be negative in small samples; it is then set to zero, giving Z = 0 and the group rate.
Buhlmann's greatest-accuracy credibility derives Z from a model in which each unit has a risk parameter drawn from a population. Two variance components are needed: the expected value of the process variance (EPV), the average within-unit year-to-year variance, and the variance of the hypothetical means (VHM), the variance between units' true mean losses. With n years of data, Z = n/(n + K), where K = EPV/VHM. When units differ greatly (large VHM), K is small and the unit's own data get high weight; when year-to-year noise dominates (large EPV), K is large and the group rate dominates. The Buhlmann-Straub extension allows different exposures (for example different insured areas or numbers of animals) in each year, with Z = m/(m + K), m being the total exposure. EPV and VHM are estimated from the data of many units, using formulae analogous to within- and between-group mean squares in one-way analysis of variance. In crop insurance, credibility weighting is used to rate units with short yield histories by borrowing strength from neighbouring units or the district.
Ruin theory studies the insurer's surplus over time. In the classical Cramer-Lundberg model, U(t) = u + c t - S(t), where u is the initial surplus (capital), c the premium income per unit time and S(t) the aggregate claims up to time t, a compound Poisson process with claim rate lambda and claim size X. Ruin occurs if the surplus ever falls below zero. Premiums must exceed expected claims, c = (1 + theta) lambda E(X) with theta > 0, or ruin is certain in the long run. The probability of ultimate ruin, psi(u), decreases as initial capital u and the loading theta increase. For exponential claims with mean mu, there is an exact formula: psi(u) = [1/(1 + theta)] exp[-theta u / ((1 + theta) mu)]. With zero initial capital, psi(0) = 1/(1 + theta) for any claim distribution in this model. More generally, Lundberg's inequality bounds psi(u) by exp(-R u), where the adjustment coefficient R is the positive solution of an equation involving the claim-size distribution; for light-tailed claims it shows that ruin probability falls exponentially with capital.
Reinsurance transfers part of the risk to another insurer. Under quota-share reinsurance, the reinsurer takes a fixed proportion of every premium and every loss, reducing the volume of risk but not its shape. Under excess-of-loss reinsurance, the reinsurer pays the part of each claim above a retention. Stop-loss reinsurance pays the part of aggregate annual losses (or the loss ratio) above a threshold, which directly addresses the systemic years that dominate crop insurance. Governments often act as reinsurers of last resort for agricultural insurance, for example by paying claims above a stated loss ratio. The statistical question is how retention levels change the variance of retained losses and the probability of ruin, and what the reinsurance costs relative to the risk transferred.
Reserves are amounts set aside for claims. Claims reserves cover claims that have occurred but are not yet paid, including those incurred but not reported (IBNR). In crop insurance, the delay between the end of a season and the settlement of yield-based claims (while crop-cutting data are collected and processed) creates large outstanding liabilities at the end of an accounting period. A classical method for estimating them is the chain ladder: claims are arranged in a triangle by origin period and development period, average development factors are estimated from past ratios of cumulative claims in successive development periods, and these factors project the immature periods to their ultimate values. Its assumption that past development patterns continue can fail when procedures or the timing of yield estimation change. Premium reserves cover unexpired risk, and capital or solvency requirements, set by regulators, protect against adverse deviations.
Individual-farm yield insurance (multiple-peril crop insurance based on each farm's own yield) suffers from asymmetric information. Adverse selection arises because farmers know their own risk better than the insurer; moral hazard because insured farmers may reduce effort or inputs when losses are compensated; and measuring each farm's yield for every claim is costly in smallholder systems. Area-yield insurance, proposed by Halcrow in 1949 and implemented, for example, in the United States Group Risk Plan and in India's national crop insurance schemes, pays indemnities based on the average yield of a defined area (the insurance unit), not the individual farm. Because no single farmer can influence the area yield much, moral hazard and adverse selection are greatly reduced, and the area yield can be estimated by crop-cutting experiments in a sample of fields. The cost is basis risk: a farmer can suffer a loss when the area yield is normal, and be paid when his own crop was fine.
The insurance unit is the area whose yield triggers claims: a district, block, revenue circle, village or village cluster. Smaller units bring the area yield closer to farm yields, reducing basis risk, but need more crop-cutting experiments in total and give noisier unit-yield estimates. The statistical trade-off is between basis risk, which falls as units shrink, and sampling error and cost, which rise. Indian schemes moved over time from large to smaller units for major crops, for example to the village or village-panchayat level, with larger units for minor crops; current unit definitions should be read from the official guidelines. The correlation between farm yields and the unit yield within a unit is a direct measure of how well the area index protects individual farmers.
The threshold yield is the yield below which claims are paid. It is usually computed as an average of past yields of the unit multiplied by an indemnity (coverage) level, such as 70, 80 or 90 per cent, which reflects the riskiness of the unit or the farmer's choice. In India's current scheme guidelines, the average has been defined as the average of the best five of the past seven years; this and the indemnity levels are scheme rules that change and must be confirmed in the current guidelines. The claim, when the actual (estimated) yield of the unit falls below the threshold, is indemnity = (threshold yield - actual yield)/threshold yield x sum insured. For example, with a threshold of 2,000 kg per hectare, an actual yield of 1,500 and a sum insured of 40,000 rupees per hectare, the shortfall is 25 per cent and the claim is 10,000 rupees per hectare. Using an average of past yields ignores trend: if yields are rising, a simple average understates the expected yield and, after multiplying by the indemnity level, protects less than intended; detrending (Chapter 6) addresses this.
Historical yields reflect the technology of their time. Yields from fifteen years ago, before newer varieties, irrigation or fertiliser use, are not draws from this year's yield distribution. To estimate this year's distribution, each historical yield is adjusted to the current technology level. The usual steps are: fit a trend to the yield series (linear, quadratic, log-linear, or a robust or locally weighted smoother such as LOESS, which resists the influence of disaster years); compute each year's deviation from its trend value; and add the deviations to the trend value for the coming year. With an additive model, adjusted yield(t) = trend(T+1) + (y(t) - trend(t)). If the variability of yields grows with the yield level, which is common, a multiplicative adjustment is better: adjusted yield(t) = trend(T+1) x y(t)/trend(t). The choice matters: additive adjustment of a rising, heteroscedastic series makes old deviations too large relative to the current level. Testing for trend and its form, and for structural breaks, uses the methods of Time Series Analysis and Linear Models and Regression.
The simplest rating method applies the insurance contract to each year of the detrended yield history. For each year, the indemnity that would have been paid is computed from the detrended yield and the current threshold, and the loss cost ratio (indemnity/liability) is calculated. The average across years is the empirical (burning-cost) pure premium rate. It needs no distributional assumption, but with 15 to 25 years of data it is very imprecise, and it gives a zero rate to any unit that happened to have no loss year in its record, even if its true risk is substantial. Rates are therefore often smoothed across neighbouring units, given credibility weights (Chapter 3), or floored.
An alternative fits a distribution to the detrended yields and computes the expected indemnity from it: the pure premium rate is E[max(0, threshold - Y)] / threshold, when liability is the threshold yield times price. Yield distributions are often negatively skewed (a long lower tail of disaster years, a ceiling set by biological potential), so the normal distribution may understate the risk of severe losses; the beta, Weibull and skewed families are common choices, compared by likelihood and tail fit. For a normal yield with mean mu and standard deviation sigma and a threshold c, the expected shortfall is E[max(0, c - Y)] = (c - mu) Phi(z) + sigma phi(z), with z = (c - mu)/sigma, Phi and phi being the standard normal distribution and density functions. Non-parametric kernel density estimates avoid choosing a family but need bandwidth choice and are unreliable in the tails with few data; semi-parametric approaches combine a parametric start with non-parametric correction. Spatial pooling of yield deviations from neighbouring units, under the assumption that their shapes are similar after standardisation, greatly enlarges the sample for the tails.
Several practical issues affect rating. Area-yield histories are themselves estimates from crop-cutting surveys, with sampling error that inflates apparent variability. Changes in unit boundaries break the continuity of the series. Strategic behaviour, such as shifts in which crops are insured in which units, changes the risk pool. Climate change may make even detrended history a poor guide to future risk, especially in the tails. Rating should therefore report the uncertainty of the estimated rate, test sensitivity to the trend model and distribution, and be updated as new years are observed.
Weather-index insurance pays according to a measured weather variable, such as cumulative rainfall in a crop phase, the number of consecutive dry days, the number of days above a temperature, or excess rainfall, recorded at a reference weather station or from gridded or satellite products. No farm visit or yield measurement is needed, so claims can be paid quickly and cheaply, and moral hazard and adverse selection on the weather itself are negligible. A typical deficit-rainfall contract specifies a trigger (payouts begin when the index falls below it), an exit (the index level at which the maximum payout is reached) and a payout rate (tick) per unit of index. With a trigger of 300 mm, an exit of 100 mm and a sum insured of 20,000 rupees, the payout rate is 100 rupees per mm, and a season with 220 mm pays (300 - 220) x 100 = 8,000 rupees. Contracts often divide the season into phases (sowing, vegetative, flowering, grain filling) with separate triggers, because the same rainfall shortfall matters more at some stages.
A good index must be strongly related to the losses farmers suffer. Design starts from the crop's physiology and agronomic knowledge of when it is sensitive, and from historical data relating the weather variable to area yields. Statistical tools include correlation and regression of detrended yields on candidate indices, phase-wise weather indices similar to those of crop forecasting, quantile regression to focus on the lower tail of yields, and crop simulation models to translate weather into yield losses where yield records are short. Triggers and exits are chosen to optimise a criterion such as the reduction in the variance (or the downside risk) of farm income with insurance relative to without, for a given premium. Over-fitting is a real danger: with 20 to 30 years of data and many possible indices, phases and triggers, a design can match past losses well and perform poorly in the future. Designs should be validated on held-out years or simulated data and kept simple enough for farmers to understand.
Basis risk is the mismatch between index payouts and the farmer's actual losses. Spatial basis risk arises because the weather at the reference station differs from that at the farm; it grows with distance and is large for convective monsoon rainfall, which varies over short distances. Design (or model) basis risk arises because the index captures only part of what causes losses (pests, floods, heat at flowering may be missing). Temporal basis risk arises when the phases or calendar dates in the contract do not match actual sowing and crop stages. Basis risk is measured by the correlation between payouts and losses, by the probability of a false negative (a significant loss with no payout, the most damaging outcome for trust), by the probability of a false positive (a payout without a loss), and by the remaining variance or downside risk of income with insurance. Denser station networks, satellite rainfall and soil-moisture products, indices combining weather and vegetation information, and hybrid products that add an area-yield or audit trigger are ways to reduce it.
Weather-index products are rated by applying the contract to historical weather (historical burn analysis): the payout that would have occurred in each of, say, 30 years is computed, and the average payout is the burning cost. Longer weather records than yield records are usually available, which is an advantage. Missing station data must be filled, trends and changes in the station or its surroundings checked, and climate change considered, since a 30-year record may not reflect current risk. Fitting a distribution to the index, or simulating weather with stochastic weather generators, gives smoother estimates of the tail probabilities. Loadings for uncertainty and for correlated payouts across a portfolio of stations, which may all pay in a widespread drought, are added as in Chapter 2.
Livestock insurance most commonly covers death of the animal from disease, accident or specified natural disasters, with a sum insured based on the animal's market value, sometimes with cover for permanent disability. The pure premium is the probability of death during the policy period times the sum insured, adjusted for partial payments and salvage. The key statistical input is therefore the mortality rate, which depends on species, breed, age, management, region and disease environment. Mortality must be estimated from data such as previous insurance claims, herd records, veterinary surveys and mortality surveys, each with its own biases: insured animals may be healthier (selection at entry) or riskier (adverse selection), and claims data omit uninsured deaths.
Life tables summarise mortality by age. Starting with a cohort of l(0) animals, l(x) is the number surviving to exact age x; d(x) = l(x) - l(x + 1) is the number dying between ages x and x + 1; q(x) = d(x)/l(x) is the probability that an animal alive at age x dies before x + 1; and p(x) = 1 - q(x) is the probability of surviving the year. The probability of surviving n years from age x is the product of the yearly p values. When animals enter and leave observation at different times, deaths are related to exposure, the total animal-years at risk: the central death rate m(x) = d(x)/L(x), where L(x) is the animal-years lived between x and x + 1. Under the assumption that deaths are spread evenly through the year, q(x) is approximately m(x)/(1 + m(x)/2). Survival analysis methods, such as the Kaplan-Meier estimator for censored data and proportional-hazards regression for the effect of covariates like breed and housing, refine these estimates.
Age-specific mortality gives age-specific premium rates; because mortality usually rises at older ages and is high in very young animals, flat rates over-charge some groups and under-charge others. Correlated losses matter here too: an outbreak of a contagious disease or a flood can kill many insured animals at once, so the aggregate loss distribution is not that of independent deaths, and catastrophe loadings or reinsurance are needed. Claims verification relies on reliable identification of the animal (ear tags, and increasingly electronic identification) and on post-mortem or veterinary certification; fraudulent claims, such as substituting an uninsured carcass, are a known problem, and the statistical monitoring of claim rates by area, agent and season helps to detect anomalies.
In area-yield insurance, the actual yield of each insurance unit is estimated from crop-cutting experiments (CCEs): randomly selected fields in the unit in which a plot of fixed size is harvested and weighed (see Crop Estimation Surveys). The estimated unit yield is the mean of the n experiments, and its standard error is roughly s/sqrt(n), where s is the standard deviation of plot yields within the unit. Because claims are paid in full on the estimated yield, sampling error translates directly into errors in claims: a unit whose true yield is just above the threshold may be paid because of a low estimate, and one with a true loss may receive nothing. The number of experiments per unit is therefore a decision about how much sampling error in payouts is acceptable.
Under simple random sampling, the number of experiments needed for the estimated mean to lie within a relative margin e of the true mean with confidence 1 - alpha is approximately n = (z times CV / e) squared, where CV is the coefficient of variation of plot yields (as a percentage) and z the standard normal quantile. With a CV of 30 per cent, a margin of 10 per cent and 95 per cent confidence, n = (1.96 x 30/10) squared = 34.6, so 35 experiments; with a margin of 20 per cent, about 9. A finite population correction reduces n when the number of fields is small. Because within-unit CVs for yields commonly lie between 20 and 50 per cent, depending on crop and region, small numbers of experiments per unit, as often used in practice, give wide margins; this is a fundamental statistical limitation of area-yield insurance at small unit sizes. Stratifying fields within a unit (by irrigation, soil or satellite-based vigour) and using auxiliary information in regression estimators reduce the required number.
Insurance guidelines set a minimum number of experiments per insurance unit and per crop, and the total across a state can run into very large numbers each season, straining staff and causing delays in claims. Choices about unit size and the number of experiments per unit should be made together: halving unit size roughly doubles the number of units and, for a fixed number per unit, the total number of experiments. Non-sampling errors often outweigh sampling error: non-random choice of fields, wrong plot size, harvesting at the wrong time, and errors or manipulation in recording weights. Measures that reduce them include random selection of fields by an independent agency, recording with geo-tagged and time-stamped mobile applications, photographs of the plot and the weighing, and audits by independent observers. Disputes over claims often turn on the quality of these experiments, so documented procedures matter as much as sample size.
Crop-cutting experiments are expensive, slow and vulnerable to error and manipulation, and they cannot be scaled to very small insurance units. Remote sensing and related technologies offer complete spatial coverage, objectivity and speed. Satellite data can verify the area sown and the crop reported (detecting over-insurance of area or wrong crops), monitor crop condition through the season with vegetation indices, map the extent of floods with radar imagery, and estimate yields with statistical or crop-model approaches (see Crop Forecasting and Growth Models and Spatial Statistics and Remote Sensing). Drones give high-resolution images for localised damage such as hail or lodging. Several countries now use technology-based yield estimates alongside or partly in place of crop-cutting experiments; India introduced a technology-based yield estimation system for its crop insurance scheme in the 2020s, with weights for technology-based and experiment-based yields that are set in scheme rules.
Satellite-based yield estimates are model predictions with errors, not measurements. The statistical task is to combine them with ground data so that the result is both precise and unbiased. Approaches include regression estimators that use satellite proxies as auxiliary variables with a reduced number of crop-cutting experiments; small-area estimation models that shrink each unit's direct estimate towards a satellite-based prediction according to their precisions; smart sampling that stratifies or selects fields for crop-cutting experiments by satellite-derived vigour, improving efficiency; and technology yields used to trigger targeted crop-cutting experiments only where losses appear likely. Each model must be validated against independent ground data from the same season and region, with errors reported at the level at which claims are paid, the insurance unit, not only at district or state level, where errors partly cancel.
Because claims are money, technology-based assessment must be transparent and contestable. Farmers and insurers must be able to understand how a yield estimate was produced, and there must be a procedure for disputes, for example by additional crop-cutting experiments when the technology estimate and other evidence disagree. Systematic biases matter more than random error: a model that underestimates losses in rainfed or smallholder areas, where mixed pixels and cloud cover degrade satellite signals, would shift money away from the most vulnerable farmers. Statisticians should therefore test models for bias across farm types and regions, report uncertainty, document data and methods, and keep a record of performance against ground truth season after season. The guiding principle is the same as elsewhere in this track: technology improves estimates when it is combined with sound sampling, not when it replaces it without validation.
A student must be able to compute compound-distribution moments, premiums under the main principles, full and partial credibility and Buhlmann credibility premiums, the classical ruin probability for exponential claims, a threshold yield and claim, detrended yields, a burning-cost rate, a weather-index payout, life-table quantities, and the number of crop-cutting experiments for a given precision, and explain the assumptions behind each.
Five years into work at an insurer, a reinsurer or a government insurance cell, the analyst prepares rate filings from yield and weather histories, detrends and pools data, designs and back-tests weather indices, checks claims data for anomalies, plans crop-cutting experiment numbers with state departments, and integrates satellite data into yield estimation, explaining results to underwriters and officials.
A senior actuary or programme head with 25 years of experience sets rating policy, negotiates reinsurance and government support for catastrophe years, decides unit sizes and the balance between technology and crop-cutting experiments, manages disputes and audits, and judges whether a scheme's premiums, subsidies and payouts are sustainable over decades of climate change.
A professor or research leader develops methods for systemic agricultural risk: spatial and copula models of correlated losses, Bayesian and credibility rating for short yield histories, optimal index design under basis risk, integration of remote sensing and crop models into yield estimation, and evaluation of insurance impacts on farm welfare. The research leader advises regulators and governments on statistically sound scheme design.
This manual does not derive the adjustment coefficient or treat ruin theory beyond the classical model, does not cover copula models of dependence, generalised linear models for ratemaking, or the full theory of claims reserving beyond the chain ladder. Scheme-specific rules (indemnity levels, years used for threshold yields, minimum crop-cutting experiments, technology weights, subsidy shares) are described only in outline because they change; current official guidelines must be consulted. Numerical examples use illustrative values, not actual rates.
Original practice questions written to the syllabus and pattern of the exams named in each question's examTags. They are not past papers and do not reproduce any board's questions. Every numerical answer was computed with a script; the values needed are stated in the question.
These notes are free to read here, with no account.
What the Laisha app adds:
Free for ten days, then a subscription. We would rather write that here than have you find it later.