Course syllabus
AI, Finance and Central Banking: the course outline
One outline for the whole course. Part I, Finance and AI, runs four teaching days on bank capital, liquidity, and what prediction and language models add to financial judgement. Part II, Central Banking, runs four lectures on communication, expectations, policy surprises, and text as data. Materials for every session are on the slides page.
Part I · Finance and AI
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.
Aims
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.
The four days form one argument, in the order mechanism, machinery, application, accountable judgement. The part closes on the sentence it has spent four days earning: “Every number in this course has a date on it and will go out of date. The four questions will not: who measured it, against what, which errors matter, and who profits if you believe it.”
Learning outcomes
On successful completion of Part I, you will be able to:
- 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;
Teaching method
Four three-hour lectures, one each day. The first two open on a real institution and a puzzle it poses, develop the mechanism that resolves it, and close on a conditional answer rather than a slogan. Their decks carry a technical appendix with the derivations in full, hyperlinked from the lecture frames, so the argument can be followed at two levels of detail. Days 3 and 4 change register: they are self-contained and carry no appendix.
Numerical examples are worked on the slides. Where a number comes from an institution it is sourced; where it is generated from the equations on the slide it is labelled as an illustration.
Sessions
| Day | Topic |
|---|---|
| Day 1 | When does an asset loss become a credit contraction? |
| Day 2 | Banking: lending capacity and runs |
| Day 3 | How modern models learn, represent, predict, and generate |
| Day 4 | Twenty cases, the models behind them, and what to do next |
Session overview
When does an asset loss become a credit contraction?
The day opens on the crisis of 2007 to 2008 and on the question the Queen asked at the LSE in 2008, which was why nobody saw it coming. Standard models treat financial markets as perfect, so they had no room for the answer, and macro-finance research grew out of that gap. The lecture then turns to a puzzle. Bear Stearns reported $395bn of assets against $384bn of liabilities, so a loss of about three per cent would exhaust its equity. Leverage turns a small asset mark into a large equity loss, and duration decides how large that mark is. But accounting alone does not determine lending. The lecture builds an agency model in which the intermediary must stay credible to its own funders, derives a debt limit from an incentive constraint, and then closes the model with a funding-supply schedule so that lower net worth reduces equilibrium credit rather than merely reducing a ratio. It ends on capital repair: shrink, issue, or bail in, and which diagnosis each instrument actually fits.
Running case: Bear Stearns, 2007 to 2008. Sections: Questions and Introduction; Losses and the equity buffer; Why intermediary net worth matters; The incentive constraint and debt limit; Equilibrium amplification; Capital repair and policy.
Banking: lending capacity and runs
Day 1 traced how a loss to intermediary net worth becomes less lending. Day 2 opens the bank itself and asks three questions in turn. First, how much can banks lend: the incentive constraint of the Gertler and Kiyotaki setup turns net worth into a lending limit, so capacity is a multiple of equity rather than a matter of will. Second, why can a bank die in a day: in Diamond and Dybvig, deposit contracts insure depositors who need money early, and the same contracts create a second equilibrium in which everyone queues; Silicon Valley Bank, where 94 per cent of domestic deposits sat above the insurance limit, is read as that queue at modern speed. Third, what kills a run: suspension, deposit insurance and the lender of last resort are compared on the margin each repairs, and on who bears the loss.
Running case: Silicon Valley Bank, March 2023. Sections: How much can banks lend?; Why can a bank die in a day?; What kills a run?.
How modern models learn, represent, predict, and generate
Days 1 and 2 built the finance: amplification from thin equity, and runs from demandable deposits. Day 3 changes register and opens the machines, for an audience assumed to have no technical prerequisites. It begins with what an AI system produces and how the field arrived at learning from data, then with where the business value has actually appeared. The core mechanism is built once: a model is a rule fitted by making errors expensive and judged only on data it has never seen. Classical model families come first, then the lecture goes inside a neural network layer by layer, and on to transformers and language models. It closes on trust and on the risks of AI-assisted decisions: what these systems still get wrong, and why verification is the bridge to the finance applications of the final day.
Running case: From ImageNet to ChatGPT. Sections: Artificial intelligence: history and context; Business applications of AI; How learning from data works; Classical model families; Inside a neural network; Transformers and language models; Trust and the bridge to finance; Risks in AI-assisted decisions.
Twenty cases, the models behind them, and what to do next
Day 3 closed on four questions and a promise. Day 4 pays it with twenty real deployments, each on one slide in the same five fields: the task, the learning type, the model family, the data and where its labels came from, and the measured result. They run from discriminant analysis on instalment loans in 1941 to card fraud decided in milliseconds, machine learning for stock returns, insurance pricing that must stay readable to a regulator, demand forecasting, contract review and firmwide language assistants. The middle of the day draws the general lesson: the label structure chooses the model, which is why fraud gets a deep network and consumer lending keeps a logistic regression. The day then turns to what fails, and finds that only one of the seven common failures is a modelling problem at all: the rest are wrong labels, wrong wiring, absent governance, a comparison that was never run, regime change, or an answer key that never existed. Every figure carries an evidence tier saying who measured it. The last third asks what a finance graduate should do about all this, and answers with an order rather than a list: the free and irreversible things first, and the expensive ones only after cheap practice has tested the direction.
Running case: Twenty deployments, from credit scoring in 1941 to firmwide language models. Sections: Part 1. Map; Part 2. Cases; Part 3. Models; Part 4. Limits; Part 5. Potential; Part 6. Suggestions.
How far Part I goes
Each day works one mechanism to the end rather than surveying several. This table separates what you derive or compute yourself from what you learn to read.
| Day | You derive or compute | You learn to read |
|---|---|---|
| Day 1 | Compute leverage, a duration mark, capital-ratio repair, a debt limit and an equilibrium response | Read a 10-K balance sheet and a stress-test result |
| Day 2 | Solve the planner allocation, derive a withdrawal threshold, compute repo capacity and contagion losses | Read a supervisory review and a resolution announcement |
| Day 3 | Build a dated bank-quarter row, fit a logistic probability rule, compute log loss and one gradient step, and score capture and precision at a fixed review capacity | Read model-risk validation guidance |
| Day 4 | Reproduce a scope-qualified ratio from a filing, build a retrieval and counterevidence plan, and score an assisted evidence packet claim by claim | Trace an API contract and a bounded controller |
What Part I leaves out
Days 1 and 2. One cumulative argument, not a survey. There is no general history of banking crises, no catalogue of every 2023 failure, no complete global-games derivation, no full network-clearing model, and no legal treatment of resolution regimes.
Days 3 and 4. One structured-data application and one research workflow, developed from beginning to end. There is no survey of model families, no neural network architectures or backpropagation, no reinforcement learning or high-frequency trading, and no generic prompt-engineering taxonomy.
Throughout. The lectures contain no lab instructions. Numerical examples are generated from the equations on the slides and are labelled as illustrations unless an institutional source is named.
Sources and reading
The lectures cite their sources on the slide that uses them. The items below are the principal ones; the public documents link directly.
- Diamond, Douglas W. and Philip H. Dybvig, Bank Runs, Deposit Insurance, and Liquidity. Journal of Political Economy 91(3), 1983. The run model Day 2 derives.
- Holmström, Bengt and Jean Tirole, Financial Intermediation, Loanable Funds, and the Real Sector. Quarterly Journal of Economics 112(3), 1997. The agency structure behind the Day 1 debt limit.
- Bernanke, Ben, Mark Gertler and Simon Gilchrist, The Financial Accelerator in a Quantitative Business Cycle Framework. Handbook of Macroeconomics, 1999. Amplification from net worth to activity.
- Baron, Matthew, Emil Verner and Wei Xiong, Banking Crises Without Panics. Quarterly Journal of Economics 136(1), 2021. Bank-equity declines predict contractions across 46 countries, 1870 to 2016.
- Board of Governors of the Federal Reserve System, Review of the Federal Reserve's Supervision and Regulation of Silicon Valley Bank. April 2023. The supervisory record behind the Day 2 case. Read online
- Federal Deposit Insurance Corporation, Dissecting Depositor Flight: An Analysis of the Spring 2023 Bank Failures. May 2026. The deposit-flow evidence Day 2 uses. Read online
- Federal Reserve Board, Bank Holding Company Supervision Manual, section 4090.0. Surveillance and Risk Assessment. The decision context Day 3 works inside. Read online
- Gaul, Lewis, Jonathan Jones and Pinar Uysal, Forecasting High-Risk Composite CAMELS Ratings. Federal Reserve International Finance Discussion Papers 1252, 2019. Why one transparent logit is enough for the Day 3 task.
- SVB Financial Group, Annual Report on Form 10-K for the year ended 31 December 2022. Filed 24 February 2023. The filing Day 4 reads by hand. Read online
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative AI Profile, NIST AI 600-1. July 2024. The governance vocabulary behind the Day 4 controls.
Part II · Central Banking
Aim. To understand how central banks influence beliefs and markets, and how economists measure those effects without confusing expected policy with genuine news.
Learning outcomes
On successful completion of Part II, you will be able to:
- explain why credibility, independence, and communication affect monetary-policy transmission;
- compare survey, market-based, and textual measures of expectations;
- construct and interpret high-frequency monetary-policy surprises;
- distinguish target, path, and information effects;
- evaluate text-based measures of tone, attention, and readability;
Sessions
| Lecture | Topic |
|---|---|
| Lecture 1 | Central-bank communication |
| Lecture 2 | Measuring expectations |
| Lecture 3 | Monetary-policy surprises |
| Lecture 4 | Communication as text data |
Teaching approach
Each lecture begins with one empirical question, develops the measurement or identification problem, and closes by stating what the evidence can and cannot support. Open the lecture before class and note the main identification question; use the study guide after class. Two core papers per lecture are on the readings page, and the mechanisms are drawn as twenty-five illustrations.
