Your overnight risk run will not finish before the desks open. What do you cut, and what do you tell the people who use the numbers?

Triage by use, not by cost

The first move is to stop thinking about the batch as one job. It is a set of outputs with different consumers, different tolerances and different legal weight, and the only defensible way to cut it is to rank those.

At the top sit the numbers that constrain trading today. Desk-level exposures against limits, counterparty exposure against lines, and the credit and settlement checks that gate whether a trade can be done at all. These need to exist by the open even if they are approximate, because the alternative is a trading day with no limits, which is a control failure of a different order from an imprecise limit.

In the middle sit the numbers that inform decisions but do not gate them: portfolio-level analytics, attribution, scenario suites beyond the mandated ones, and management reporting. These can be late.

At the top for a different reason sit the numbers with a regulatory or accounting destination — capital calculations, official valuation and disclosure inputs, the mandated stress scenarios. These cannot be approximated, because an approximation submitted as a calculated figure is a misstatement. They can be late, provided the deadline they are late against is your internal one and not the supervisor's, and provided you know which of those it is.

Stating that ordering out loud is most of the answer. Candidates who instead start with which computations are cheapest to skip have inverted the problem: the cost of a calculation is irrelevant to whether you are permitted to omit it.

The thing that failed is probably market data

Before choosing what to cut it is worth being right about why the run is late, because the usual cause is not compute. It is an input that has not arrived or has arrived wrong: a curve missing a tenor, a vendor file delayed, an illiquid instrument with no observable price today, a corporate action unprocessed, a new instrument traded yesterday whose reference data was never set up, a holiday calendar that made a source unexpectedly quiet. The revaluation then stalls, or worse, proceeds on whatever it could find.

That distinction changes the remedy entirely. If compute is the constraint, more capacity or a coarser method fixes it. If an input is missing, no amount of capacity helps and the real decision is what to substitute and how to declare it. That is why a run should validate its inputs for completeness and plausibility before it starts revaluing rather than discovering gaps halfway through: knowing at the start that four curves are missing gives you hours to source them, whereas knowing at the end gives you a choice between a gap and a delay.

Degrade the method, publish the coverage

Where the constraint is genuinely computational, the honest degradations are well established and each has a stated cost. Approximating a position's change in value from its sensitivities rather than revaluing it fully is fast and understates the behaviour of anything with meaningful convexity, so it is acceptable for a linear book and dangerous for an options book. Reducing the number of scenarios or simulation paths widens the error on the tail, which is where the measure you care about lives. Aggregating at a coarser hierarchy gets you a group number sooner while losing the desk-level detail that limit monitoring needs — the wrong trade in this scenario, since the desk numbers are what the open requires.

Whatever you choose, the output carries its own provenance: which portfolios are included and which are missing, which positions were approximated and by what method, the vintage of every market data set used, and the time the run started. This is the whole difference between a degraded number and a misleading one. A risk report that arrives on time with three per cent of the book silently absent looks exactly like a complete one, and every consumer will treat it as complete. A report that says on its face which portfolios are excluded lets a desk head decide whether their own limits are meaningful this morning.

Reusing yesterday's marks is sometimes the only option and is legitimate if it is declared as such and if the population it applies to is small and named. It becomes a serious defect the moment it is the silent default, because stale marks make a portfolio look calm precisely when the market has moved, which is the one morning the number mattered.

Somebody has to be allowed to decide, before the night it happens

Operationally, the failure mode is not the missing number. It is a support engineer at four in the morning deciding on behalf of the firm which parts of the risk framework to omit. That decision belongs to the risk function, it needs to be pre-authorised, and it needs to be written down as a small set of named degradation options with their conditions: what may be approximated, what may use previous-day inputs, what must never be published incomplete, who is told, and by what time.

Alongside it sits an escalation with a clock. If the run has not reached a defined checkpoint by a defined time, you invoke the degraded path rather than continuing to hope, because a full run that completes two hours after the open has already failed its purpose while consuming the window in which the alternative was possible. Checkpointing the run so it can resume rather than restart is what makes that judgement affordable, and it is worth the engineering effort precisely because the alternative is a binary between waiting and starting again.

Reconstructing the day afterwards

Two things close this out. The full run happens when the inputs are there, and its results are compared against the degraded ones — not to embarrass the approximation but because the difference is the only evidence of whether the degradation was safe. A pattern of large divergences on a particular book tells you that book must never be approximated again. That comparison also settles the awkward case where a limit was breached under one method and not the other: the position taken during the day was taken on the information available, so the record must show what was known at the time, and the after-the-fact number is a correction rather than a retrospective breach.

The second is the audit trail, which supervisory expectations for risk data aggregation and reporting — the BCBS 239 principles being the reference point most banks work to — treat as part of the capability rather than as documentation. Accuracy, completeness, timeliness and adaptability are all named there, and a firm that regularly publishes incomplete numbers without being able to say what was missing is failing the completeness principle in a way that is visible in a review. Keeping the inputs, the method flags and the coverage statement for each run is what lets you answer, a year later, why that morning's figure was what it was.

Cut by what the number is used for, never by what it costs to compute, and publish coverage and market data vintage alongside every degraded figure - an incomplete number that does not admit it is worse than no number at all.