Materials for each session

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.

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.

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.

Sections. How much can banks lend?; Why can a bank die in a day?; What kills a run?.

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.

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.

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.

Sections. Part 1. Map; Part 2. Cases; Part 3. Models; Part 4. Limits; Part 5. Potential; Part 6. Suggestions.

When communication moves expectations, and why anchoring and accountability make it believable.

Surveys, prices, and text reveal different traces, each with its own measurement wedge.

Separate expected actions from target news, path news, and cross-border responses.

Measure attention, interpretation, disclosure, and readability without confusing release with impact.