Journal · Desk notes

Composites, not case studies with a client’s name rubbed out.

These three notes are constructed teaching stories. They use public market structure and the failure modes we write about. They are not engagements, not track records, and not a description of any identifiable firm.

A green banker's lamp and stacked cream folders on a dark desk

If you need the underlying argument without the narrative, start with forecasting, risk and execution. The notes below exist so the failure has a clock and a room, which is how desks actually meet it.

Note 01 · 21 July 2026

The fiscal week the nowcast arrived on Wednesday

A composite. No performance is claimed. No gilt trade is suggested.

Imagine a small macro pod that had spent a year teaching a model to nowcast the next gilt-market relevant data print from timely series: card spending, job adverts, a wholesale energy proxy, a survey. The model was dull in the way good models are dull. It updated a distribution. It did not speak in certainties.

Then a fiscal announcement arrived mid-week, larger than the pod’s scenarios, and the long end moved before any of the timely series had anything new to say. The nowcast, loyal to its inputs, still looked like the previous Friday. By Wednesday afternoon a junior asked why the model had “missed the sell-off.” It had not missed a print. The print had not happened. The market was repricing an announcement the feature set treated as a residual.

The repair was not a larger network. It was an event flag with a human owner, a scenario that widened the long-end distribution when a fiscal document crossed a size threshold, and a sentence on the morning note: the nowcast is about the data, not about the auction of government paper in the next hour. The pod kept the model. They stopped asking it to be a politician.

The lesson sits beside the collateral discussion. A yield can move because the economy changed, or because the people who hold the hedge need cash. A single nowcast is not obliged to cover both, but the desk is obliged to know which one it built.

Note 02 · 12 October 2026

The earnings model that cheated

A composite about leakage. The accuracy in the first draft was never a result.

Printed chart blurred into texture beside a pen, standing in for a result that should not be read too closely

A research group fitted a language model to UK trading statements and asked it to classify whether the next reported earnings number would sit above or below the prevailing analyst mean. The first evaluation was splendid. The second evaluation, done by someone who had not built it, joined each document to the timestamp it became public.

A slice of the splendour was mechanical. Some “statements” in the vendor file were time-stamped to the period they described, not the morning they were released. The label, the earnings surprise, was therefore partly visible in documents that, in the real week, did not yet exist. Another slice came from the embedding itself, fit once on the full corpus, so later vocabulary helped earlier rows. Neither mistake requires malice. Both require a calendar.

Rebuilt with publication times, a frozen embedding trained only on the past, and a purge around reporting clusters, the classifier looked ordinary. On some sectors it was no better than a model that ignored the text and used the last revision to consensus. The group published the ordinary version internally and shelved the splendid one. That is the whole success of the story.

We retell it because the splendid chart is what usually escapes into a slide. The timestamp table does not. Clarsachaxa would rather be the timestamp table. The method is set out without the fiction on the forecasting page.

Note 03 · 19 August 2026

A thin name into the close

A composite about execution. No symbol is named because the symbol is not the point.

Picture a systematic book that rebalances monthly into UK mid-caps. The research treated the closing auction as a convenient print. A scheduler, trained mostly on FTSE 100 days, learned to send a large child order into the last call because, in those names, the close absorbed it.

On a thinner name the same fraction of the order was no longer a participant. It was the auction. The fill was official, the benchmark was met in a narrow sense, and the next morning’s print gave back a move that the backtest had recorded as implementation edge. The model had not “failed to predict.” It had been shown a room with more chairs than the room it was later locked in.

The fix the composite desk adopted is the one we argue for on the execution page: separate schedules by liquidity bucket, a hard cap on participation in the call, and a comparison against a patient schedule that is allowed to finish the next day. Finishing tomorrow is not a moral failure. Forcing a thin close in order to match a research timestamp is.

Again, nothing here is a claim about a live book’s profit. It is a claim about a category error: liquidity is not a single UK equity fact. It is a fact about this name, this day, and this size.

If one of these failure modes is the one you are currently living with, the desk can look at the shape of it. We still will not sell you a signal.

Write to the desk