This is a bit stream of consciousness for now. It'll need cleaning up.

Framework

A tree of model components. Each component will be a ModelBit. Each ModelBit may have children. The order of the children is significant.

When updating the model, we'll do a depth-first traversal of the tree. This means that any ModelBit can use the results of another ModelBit as follows:

• if the user precedes the other, it can use only the results from the previous period
• if the other precedes the user, the results from the current period can be used too.

It's important that, for a given configuration file, the children of each ModelBit are stored in the same order, so that the same random numbers (from a given seed) are applied to the same things.

We are not going to try to work out what random numbers (shocks) are needed in advance. This is because when we come to things like large claims, the number of shocks we'll need for claim amounts will depend on the actual shocks generated for claim frequencies.

The top-level configuration will be fixed.

The root of the model will be a ModelBit, as usual, and will have the following children:

• External (interest rates, and other things that come from outside. Maybe cats, for example?)
• Liabilities
• Assets
• Corporate level (tax, and other whole picture things)
• Business logic (management actions, that affect only the following period)

The liabilities ModelBit will have one child for each line of business, plus children for reinsurance.