Four-day course
Finance and AI
The course follows one argument across four days: a financial mechanism is established, then measured, then used to predict, then required to produce evidence that a responsible person can check. Days 1 and 2 are finance. They ask when an asset loss becomes a credit contraction, and how an institution with valuable assets can still fail to pay. Days 3 and 4 take those mechanisms and ask what they license: one dated probability that ranks cases for review, and one evidence packet in which every claim carries its source and its qualification.
Course materials
Course overview
Finance comes first. The course does not open with a model or a library. It opens with a balance sheet and a week in March, and asks what has to be true for a loss to reach lending, and for a payment to fail while the assets are still sound. Only once those mechanisms are on the table does it ask what a machine could usefully add.
What it adds is deliberately narrow. There is one prediction problem and one transparent probability rule, specified completely before it is fitted and tested at genuinely later dates. There is one evidence workflow, built against a filing read by hand so that the automated version has a benchmark to match. The recurring discipline is that a prediction is not a diagnosis, a fluent answer is not a sourced answer, and neither a model nor a language model holds the authority to act.
Learning outcomes
By the end of the course you can:
- write a bank balance-sheet identity, and compute how far an asset loss travels into equity through leverage and duration;
- distinguish a capital shortfall from a liquidity shortfall, and match each to the instrument that repairs it and the party that bears the loss;
- derive a debt limit from an incentive constraint, and explain why lower net worth can reduce equilibrium lending rather than only a ratio;
- derive why demandable claims provide liquidity insurance and create a run equilibrium, and compute the withdrawal threshold;
- separate a common shock from contagion, and name the transmission margin any contagion claim requires;
- specify a prediction problem completely, including population, information cutoff, target, measurements, evaluation and permitted use, before any model is fitted;
- fit and read one transparent probability rule, and test it in financial time without moving future information backwards;
- turn a frozen risk signal into a dated evidence packet in which every claim carries its support, its calculation and its qualification;
- state what a model or a language model may and may not authorise, and who retains responsibility for a consequential decision;
The four days
| Day | Topic |
|---|---|
| Monday | When does an asset loss become a credit contraction? |
| Tuesday | Liquidity, runs, and systemic fragility |
| Wednesday | From financial mechanisms to one defensible risk signal |
| Thursday | From a frozen risk signal to verifiable financial evidence |
