London is a deep market and a local one at the same time. A global model of “equities” will treat a FTSE multinational and a domestic services name as cousins. They share a listing venue and, on many days, not much else. The use case starts by refusing that cousinhood until the data earns it.
FTSE 100, FTSE 250, AIM
The FTSE 100 is liquid enough for intraday questions and international enough that a sterling-only story will often be the wrong one. Many constituents earn and report outside the UK. A model that explains them with only domestic macro is fitting the postcode of the listing, not the business. The FTSE 250 is where domestic exposure is richer and depth is less forgiving. AIM is where a beautiful feature set goes to discover that the spread was the feature.
Cross-sectional equity models, the ranking of names within a universe on a slow signal, have a fairer chance than heroics on the next five minutes. They still require point-in-time membership, a cost model that knows about stamp duty reserve tax, and a willingness to keep sector structure explicit. The portfolio note is the companion to this paragraph. The auction note is what happens when the cross-section has to be implemented rather than admired.
Gilts and index-linked gilts
The gilt curve is a small number of instruments with a large amount of meaning. That is a gift and a trap. A gift, because the object is coherent: issuance, the DMO’s calendar, Bank of England holdings, inflation expectations, and a dealer network you can at least name. A trap, because flexible models have more parameters than genuinely distinct regimes. You can fit a neural net to a curve and discover that it has redrawn a spline.
Useful machine-learning work here is usually state classification and scenario completion, not a tick-level price prophet. Is depth normal. Is the long end moving with or without the short end. Are linkers and conventionals telling the same inflation story. The 2022 stress, discussed on the risk page, remains the reminder that positioning and collateral can dominate the next print of the curve. No public yield history contains the full position set. Pretending otherwise is how backtests become short stories.
SONIA replaced a habit, not a dataset
Sterling overnight index average is the spine of the sterling rates complex after LIBOR. Models trained across the transition that do not mark the break will mix two contracts and call the mixture a cycle. The short end is also a policy instrument. A forecast that “beats” the forward curve over a short window may be beating noise, or it may be taking a view on the Bank that should be labelled as a view. Machine learning does not convert a macro opinion into a measurement. The honest use case is often nowcasting the inputs the committee itself watches, with the lags written down, and then stopping before a trading instruction is implied.
Beating the forwards for a month is an anecdote. An anecdote is not a desk.
Sterling is liquid and still local
Cable and the other sterling crosses are deep enough for serious microstructure and for slower macro models. They are also a concentrated expression of UK political and fiscal news in a way a global equity index is not. Text models on that news are tempting and easy to leak: timestamps on headlines, revisions, and the difference between a rumour and a Budget document. The same point-in-time discipline as on the forecasting page applies, with less mercy, because the move can be over before a sloppy join is noticed.
Power, gas, and the property that listens to rates
UK power and gas are where weather, outages, interconnectors and storage meet a market design that is not a textbook geometric Brownian motion. Machine learning is genuinely useful for short-horizon load and renewable generation, which then feed a trader’s view of the prompt curve. It is much less useful as an unconstrained forecaster of the far curve, where policy and capital dominate. Keep the horizon shorter than the story you are tempted to tell.
Listed REITs and housebuilders sit at the junction of that rates complex and ordinary equity microstructure. A model that treats them as generic mid-caps will miss the duration. A model that treats them as pure rate instruments will miss the balance sheet. Interaction, stated explicitly, beats another hidden layer that may have found the same thing and will not admit it.
The shrug list
We would rather not build, and we will not applaud, models whose sample is a handful of regime days dressed as a large number of ticks. Single-name options with wide markets. Tiny AIM prints used as if they were a continuous price. Any strategy whose backtest requires yesterday’s index membership and tomorrow’s corporate action. Any accuracy figure quoted without the decision it was meant to serve.
The map is allowed to have blank counties. A journal about machine learning in markets that cannot say “not here” is a brochure.