Description
Predict how many units, accounts or users of different ages will have an event in the next period, with simulation-calibrated prediction bounds.
Units enter service at different times, so at the analysis date each cohort has its own age, survivors and chance of an event in the coming window. For each cohort the conditional probability of an event in the window, given survival to its current age, comes from the fitted time-to-event distribution; multiplying by the number still at risk and summing across cohorts gives the expected count. The total is a sum of independent binomials with different probabilities, often close to Poisson, which gives naive bounds. Because those bounds ignore uncertainty in the fitted distribution, Meeker and Escobar calibrate them by simulating new datasets from the fitted model, refitting, and finding the nominal level that gives the target coverage. The by-cohort risk table is itself a useful planning artefact.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Unmigrated
- Stage
- Post-launch, Strategy
- Data
- Quantitative
- Effort
- Days
Steps
- Fit a time-to-event distribution to the installed base, treating units still active as censored at their current age.
- Group active units by age (cohort) and record, for each, the number at risk and the expected use or exposure in the coming window.
- For each cohort compute the conditional probability of an event in the window given survival to its current age.
- Multiply by units at risk and sum across cohorts for the point prediction; tabulate the cohorts in a risk table.
- Compute naive prediction bounds from the sum-of-binomials (or Poisson) distribution.
- Calibrate the bounds by simulating datasets from the fitted model with the same censoring pattern, refitting each, and adjusting the nominal level until coverage reaches the target (for example 95 percent).
Sources
- Statistical Methods for Reliability Data (Meeker & Escobar, 1998)
Also known as
- Future-failure risk analysis
- Warranty exposure prediction
- Field failure forecasting
- Within-sample prediction
- Future-failure risk analysis: prediction bounds for the number of failures in an installed base with staggered entry
Categories
- Analytics & Behaviour
- Statistics & Modelling