Description
Fit a short list of candidate distributions with likelihoods that respect how the data were recorded, then compare them with plots, fit tests and penalised scores to commit to one.
Simulations, forecasts and capacity plans often need a probability model for a skewed quantity such as time, spend or load. Klugman and colleagues set out a disciplined workflow. Restrict candidates to a short, sensible list. Fit each by maximum likelihood with terms that match the recording: grouped data contribute interval probabilities, values capped at a limit contribute survival beyond it, and values recorded only above a threshold contribute density divided by survival at the threshold. Judge fit with the model against the empirical distribution, a difference plot and a p-p plot; with Kolmogorov-Smirnov, Anderson-Darling (tail-weighted) and chi-square tests, using simulated p-values when parameters are estimated; with likelihood-ratio tests for nested models; and with the Schwarz criterion. Judgement settles disagreements, weighting fit where the decision matters. Splicing different distributions over different ranges is a variant.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Latent
- Stage
- Evaluation, Post-launch, Strategy
- Data
- Quantitative
- Effort
- Days
Steps
- Shortlist two to five candidate distributions (for example exponential, Weibull, lognormal, gamma, Pareto), preferring simpler ones.
- Fit each by maximum likelihood with likelihood terms matching the data: interval probabilities for grouped data, survival at the limit for capped values, density over survival at the threshold for truncated values.
- Plot each model against the empirical distribution, a difference plot of empirical minus model, and a p-p plot, stopping at any censoring point.
- Test fit with Kolmogorov-Smirnov, Anderson-Darling or chi-square (expected counts about 5 or more), simulating p-values by refitting to samples from the fitted model because estimated parameters make tabled critical values too lenient.
- For nested models use the likelihood-ratio test, adding a parameter only if the log-likelihood rises by at least 1.92.
- Rank candidates by the Schwarz Bayesian criterion and weigh fit in the region that matters (tail or body) and precedent.
- Commit to one model, document the choice, and keep runners-up for sensitivity analysis.
Sources
- Loss Models: From Data to Decisions (Klugman et al., 2012)
Also known as
- Severity model selection
- Loss distribution fitting
- Model selection for loss models
- Distribution fitting workflow
- Parametric distribution fitting and selection workflow for skewed, censored or truncated measures
Categories
- Analytics & Behaviour
- Statistics & Modelling