Value at risk is a quantile. That is all it is. A machine can estimate the quantile with a historical window, a parametric guess, or a flexible model that has been shown more days than a human would willingly plot. None of those choices turns the quantile into a description of what the book does on the day the quantile is breached. The breach is the use case. The rest is housekeeping.
Flexible models help when the shape of the tail changes with measurable state: the level of implied volatility, the tightness of dealer balance sheets, the distance of a rates move from the last period in which margin calls were boring. They hurt when they are so flexible that the backtest can be massaged into a green cell. A risk model should be slightly embarrassing in calm times. If it never is, it has been fitted to the embarrassment.
September 2022 was not a volatility surprise alone
The gilt market stress of autumn 2022 is public history, and it is still the cleanest UK lesson available to a model builder. Long-dated yields moved hard. Liability-driven investment pools, structured around leverage and collateral, had to sell into the move. The loop was not “volatility was high.” The loop was “the hedge demanded cash, the cash required sales, the sales moved the hedge.” A model of daily gilt returns that stops at the return distribution will describe the week badly, because the distribution was partly produced by the positioning the model never saw.
Machine learning does not repeal that. It can, used modestly, help in three places. First, as a classifier of market state: is today’s move still inside the cushion of two-way interest, or has depth left the long end. Second, as a map from yield moves to plausible collateral calls, if and only if the desk has the position data and is allowed to use it. Third, as a generator of coherent scenarios rather than a single shocked number: a parallel shift is a cartoon; a steepening with patchy depth and a sterling move beside it is closer to a week a human remembers.
If the model cannot see the margin agreement, it should not pretend to have seen the crisis.
We do not publish a gilt “early warning” score. Anyone offering one without talking about who is leveraged, on what documents, should be treated as a writer of weather, not of risk.
Crowding is a correlation that arrives late
Equity factor books in the UK, as elsewhere, can look diversified on a covariance estimated in a quiet sample and singular on the day a style is unwound. Machine learning is sometimes asked to detect crowding from returns alone: residual co-movement, falling dispersion, the sort of pattern that appears when many books hold the same cheap quality name. Returns are a shadow of the holdings. Shadows are usable. They are not the object.
A practical use case is monitoring, not prophecy. Cluster residual returns of a liquid UK universe. Watch the clusters tighten. Hand the picture to someone who knows whether the book is in the cluster. The model’s job is to make the tightening visible earlier than a monthly factor tear-sheet. It is not to emit a sell ticket.
Credit migration is a panel, not a vibe
For lenders and for anyone holding sterling credit, the machine-learning use case that has aged better than most is old-fashioned: rank obligors, with a model that is monotonic in the features a credit committee already believes, validated on a cycle that includes stress, and challenged for fairness and stability where consumer outcomes are involved. Gradient-boosted trees can beat a scorecard on a leaderboard and then fail a conversation about why a borrower moved two notches. If you cannot explain the move, you do not have a use case a UK credit committee will sign.
On traded credit the sample is thinner and the stamps of liquidity are larger. A model that predicts spread change without a feature for “could you have sold the bond” is predicting a mid that may not have been a market. We would rather read a short note on when sterling investment-grade depth vanishes than a long note on a neural net’s hit rate against that mid.
A small library beats a large simulation nobody opens
Scenario generation is where modern models are genuinely helpful, provided the scenarios stay nameable. A library we would actually keep on a shelf:
- A long-end gilt move with rising yields, patchy depth, and a collateral spiral, informed by 2022 rather than copied from it.
- A sterling gap against the dollar that is not matched by a gentle equity drift, because the currency is sometimes the whole story.
- A UK power and gas spike that rewrites near-term inflation nowcasts and the equity sectors tied to them, without assuming the spike mean-reverts on a researcher’s preferred half-life.
- A calm tape in which correlations are low and the danger is leverage quietly rebuilding. Models trained only on drama miss the loading of the spring.
Generative models can fill paths between those anchors. They should not be allowed to invent a fifth anchor that no one on the desk can describe to the board. If the path cannot be said aloud, it is not a control. It is a screensaver.