A forecast can be right about the direction of a price and wrong about the trade, because the trade moves the price. Implementation shortfall, the gap between the price when you decided and the price you actually obtained, is the honest ledger. It contains delay, spread, impact and the opportunities you missed by waiting. Machine learning enters as a way to predict that gap, conditional on size, time of day, the name’s recent depth and how urgently the portfolio actually needs the fill.
Urgency is the feature researchers forget. A risk-reducing sale into a gilt-linked equity move is not the same object as a slow accumulation of a mid-cap the book intends to hold for a year. A model that treats both as “buy or sell N shares” has already thrown away the decision. The use case begins with a deadline and a regret, then asks the tape how expensive those are today.
Child orders are a sequence, so sequence models are tempted
Splitting a parent order across the day is an old problem with a new vocabulary. Volume curves, square-root impact sketches, and reinforcement-style schedulers all try to answer when to show size. The data are plentiful at the event level and poor at the decision level: you observe the path you took, not the path you declined. That counterfactual is the whole difficulty. A model praised for matching yesterday’s fills may simply have copied yesterday’s timetable.
We like approaches that stay close to a baseline a human can audit. Start from a participation schedule. Let the model deviate when measured depth, spread and a short-horizon volatility state justify it. Cap the deviation. Log the reason. If the reason cannot be stored in a sentence, the deviation does not ship. This is less glamorous than an agent “learning to trade.” It is also how you explain a bad afternoon to someone who was not in the room.
Venue choice is a dataset with a short memory
UK cash equities print across lit books, periodic auctions, dark mechanisms where the rules allow, and systematic internalisers. Market share moves. Fee schedules move. A classifier trained in one microstructure regime will recite that regime after it has ended. The use case that survives is comparative and refreshed: given this size and this urgency, where has similar flow recently been completed without becoming the price. “Recently” is the word that does the work. A five-year average of venue quality is a history essay.
Order-book features, imbalance among them, decay quickly. They can be real for seconds and folklore by the time a slide is formatted. Any published claim that a particular imbalance pattern “predicts” a UK mid-price minutes ahead should be read with the costs and the latency attached. If the author does not have the latency, they do not have the claim.
The closing auction is not the leftovers
A substantial share of UK lit volume gathers at the close. Benchmarks, futures hedges and ordinary human preference for a reference price all meet there. For a machine, the auction is attractive because it is a repeated, rules-based event. It is dangerous for the same reason: everyone who can count has noticed, and the statistics of a crowded close are not the statistics of the afternoon.
A serious use case asks a narrow question. Given what is already public in the auction call, and given the parent order’s constraint, is participation likely to reduce shortfall relative to a patient afternoon schedule. The answer depends on the name. In the most liquid FTSE names the close is deep. In parts of the FTSE 250 it is a room with fewer chairs than the backtest assumed. A composite of that thin-name problem is in the desk notes. It is labelled as a composite because a single war story is not a sample.
Do not train a close model on days when you were not trying, and then act surprised by the days when you are.
Market abuse is not an edgy feature set
UK desks operate under the market abuse regime. Layering, spoofing, and any pattern of orders whose purpose is to give a false or misleading impression are illegal, not clever. This journal will not discuss techniques for painting a book, nudging a close, or inferring information you are not entitled to have. Features belong to information you may use. If a data vendor’s feed would make a compliance officer close their laptop, it is not a feature. It is a problem you have not had yet.
The constructive machine-learning work sits next door to surveillance, not opposite it. Models that reconstruct a desk’s own intent, that flag child orders which no longer match the parent, and that explain a shortfall in ordinary language, are use cases a firm can defend. Models that treat the rest of the market as an opponent to be confused are a different industry, and not this one.