Which library do I need?

You need real-looking motor data to test a method but cannot share your book externally — or you want a benchmark with a known data-generating process to validate whether your model is even recovering the right signal.

Your model development team cannot touch production data, but the synthetic data you have does not preserve the multivariate dependence structure of the real book — correlations between vehicle age, driver age, and claim frequency are wrong.

Your book mix has shifted — you wrote more young drivers last year through a new aggregator — and the model fitted on historical data is mispriced for today's portfolio because it was never trained on this mix.

Your GBM outperforms the production GLM on every holdout metric, but the actuarial team and your rating engine both need multiplicative factor tables — and exp(β) from CatBoost is not a thing.

You used random k-fold CV and your held-out Poisson deviance looks fine, but the model was trained on claims that had not yet developed — IBNR from the last six months is in both your training and validation sets.

Your main-effects GLM has unexplained residual structure — one-ways look fine but double-lifts between age and vehicle group show a gap the model cannot explain, and you suspect an interaction you have not captured.

Your point-estimate GBM tells you the expected loss cost per risk, but you have no way to distinguish a low-risk policy with stable claims from a high-risk policy that just happened to be quiet last year — you need per-risk volatility.

Your overdispersion test is significant and the dispersion parameter phi varies by segment — fleet policies have much higher variance than personal lines — but a standard GLM assumes one scalar phi for the whole portfolio.

You want to model dispersion as a function of rating factors without the full complexity of GAMLSS — a simpler double GLM where the mean model and the dispersion model alternate until convergence.

You need interpretability that a GLM factor table gives you, but you also need the predictive power of a neural network — and you want exact Shapley values rather than approximate SHAP for sign-off.

Your vehicle make variable has 500 levels, your occupation code has 350 — the committee wants pricing bands, not a 500-row factor table, and you need a defensible method for collapsing them that respects the underlying risk signal.

Your one-way driver age curve is spiky — the raw relativities jump between adjacent age bands because your data is thin in some cells — and you are eye-balling a smooth curve through it rather than doing this formally.

You have broker, scheme, or fleet as a grouping factor with hundreds of levels — some with thousands of policies, some with ten — and you need credibility-weighted group adjustments rather than either ignoring the grouping or overfitting to it.

Your frequency and severity models are fitted independently, but high-frequency risks also tend to have higher severity — you are ignoring a dependence structure that is inflating the premium in some segments and deflating it in others.

Your severity model uses a single Gamma GLM across all loss sizes, but the body and tail of your loss distribution follow completely different physics — attritional claims behave like Gamma, large losses follow a Pareto tail, and one model cannot fit both.

You need to price for large loss loading or produce an increased limits factor table, but your mean model gives you no handle on the upper tail — you need quantile estimates at the 90th and 99th percentile per risk, not just the expected value.

A segment has 200 policies and your GLM estimates are unstable — the standard errors are wide, the coefficient on vehicle age has flipped sign, and a Bayesian approach requires you to specify priors you do not have strong views on.

Your territory factors are eye-balled from a heat map or computed naively from raw one-ways — thin postcode cells get volatile relativities, adjacent postcodes have no relationship to each other, and the factors do not borrow strength from neighbouring areas.

You have a specialist segment — classic car, high-net-worth, niche occupation — with enough data to price but not enough for a standard GLM to separate signal from noise, and you want to incorporate prior actuarial knowledge formally rather than ad hoc.

Your vehicle value factor looks significant in the GLM, but vehicle value correlates strongly with distribution channel — direct customers buy cheaper cars — and you cannot tell whether it is genuine risk signal or channel confounding.

You put through a rate increase in Q3 and conversion dropped — but you cannot tell how much of that drop was the rate change versus market conditions, because you have no control group and the pre/post comparison is confounded by seasonal effects.

You have raw 1Hz GPS and accelerometer data from a UBI product and need to turn it into a GLM-compatible risk score — but you have no principled method for aggregating trip-level events into driver-level risk that accounts for the mix of journey types.

Claims frequency has been drifting upward for eighteen months and you need to know whether it is a sustained trend, a structural break, or noise — and you need the answer expressed as a trend index in development factor format for the reserving team.

Your NCD schedule is fixed at five years old and the discount scales were set judgementally — you want to fit actuarially correct credibility weights from your own data and incorporate dynamic claim history into an individual policy posterior.

You want to model lapse risk and customer lifetime value, but standard logistic regression ignores the time structure — policies renewing at different durations have different lapse hazards, and a cross-sectional model is missing that entirely.

You need a price elasticity estimate for rate change planning, but conversion data is observational — the prices customers saw were set by a model, not randomly assigned — and a naive regression of conversion on price is picking up selection effects.

Your technical price model and your commercial rate are disconnected — the pricing committee sets rate change by segment without knowing the conversion and retention response, and FCA GIPP requires you to consider demand effects formally.

Your pricing actuary says postcode is a legitimate risk variable, but the compliance team is worried it correlates with ethnicity — and under FCA EP25/2 you need to quantify the indirect discrimination risk, not just assert there is none.

Your model governance committee needs a validation report and you are producing it manually in PowerPoint — bootstrap Gini CI, Poisson A/E CI, double-lift charts, and a renewal cohort test, all formatted to what a model risk function expects.

You want to run a new model in shadow mode before cutting over, but you have no infrastructure for routing quotes deterministically, logging both prices, and eventually running a statistically valid comparison to declare a winner.

Your model has been live for nine months and A/E ratios are creeping up — but you do not know whether to recalibrate the intercept, refit with new data, or whether it is just IBNR noise in the most recent quarter.

Your GBM gives a point estimate per risk but Solvency II internal models need uncertainty bounds — and the standard bootstrap approach is slow and the distributional assumptions are wrong for insurance losses.

You have a technical price per segment, a loss ratio target, and maximum movement caps — and the rate change recommendation is currently done in a spreadsheet where the constraints interact and the solution is not optimal.