Several models of the underlying stochastic system drivers are postulated, e.g. the way event probabilities change with time could take different forms.  The available historical data is used to determine which of these models are consistent with the observations.  This is achieved through the calculation of likelihood for each model-parameter combination within a Bayesian framework.  It would be possible simply to pick the best combination, but when the aim is to understand uncertainty it is better to work with several models that are consistent with the data but may extrapolate differently.  The relative likelihoods for the models and parameter distributions can be established.

An important step in cases where new data is continuously becoming available is to test the predictive power of the models by making blind predictions ahead of the new data being obtained.  These predictions will be in terms of probabilities of the various possible outcomes, which may be discrete or continuous.  It is useful to include other models, e.g. process-based models, within such an exercise too.  Once the new data is obtained, the models can be judged on how well they performed.  Poorly performing models can be rejected or revised.  Good models will consistently make good predictions and confidence in forecasts made with them will improve.

The figure shows a statistical prediction of the likelihood of possible inspection results for the number of additional doubly-cracked bricks (x axis) and additional singly-cracked bricks (y axis), with the measured result (ringed) lying in the high likelihood region.

The development of CoreStats used pre-existing Quintessa software components together with specific code for the domain of interest, enabling a functioning code to be developed rapidly and efficiently.