Mindencap

Insights / AI Operations

Reconciliation Breaks: Most Aren't Errors, and That's the Problem

Most breaks between a fund's books and its administrator's are timing, price-source and mapping differences, not mistakes. Sorting them is where automation earns its place.

· Mindencap

Two columns of fine ivory lines on a navy ground, nearly aligned and joined by a single gold line

It is the third working day of the month. The administrator's draft NAV is in your inbox, your own position file is open beside it, and a column of lines is highlighted red. Someone on your side is about to spend the afternoon discovering that most of them are not mistakes.

That afternoon is the problem worth fixing. Not because the red lines are wrong to exist, but because a process that treats every difference as equally alarming teaches people to stop reading them.

What a break actually is

A break is a difference between two records that are supposed to agree: your books against the administrator's, the administrator's against the custodian's, a position file against a cash statement. It tells you the two records disagree. It does not tell you which one is wrong, or whether either is.

Treat every break as an error and the review becomes a queue of identical-looking red cells. Treat none of them as errors and the one that matters slips through unnoticed. The real work is classification: deciding, for each difference, what kind of difference it is.

Where breaks come from

The causes fall into a small number of families. Knowing them is most of the job, because each family has a different fix.

  • Timing. A trade is booked the day it is agreed; cash and settlement arrive later. Two systems can each be right, on different dates. Income accruals and fee debits behave the same way.
  • Price source. Two parties value the same holding using different vendors, different cut-off times or different rules for thinly traded and illiquid assets. Both prices are defensible. They are not equal.
  • Identifier mapping. The same instrument carries different identifiers in different systems, or a record is renamed after a corporate action. A position appears to vanish from one side and reappear as a new line on the other.
  • Currency. Multi-currency holdings pick up differences from the rate used, the moment it was struck and the rounding applied at each step.
  • Rounding and fractions. Small residuals from fractional units, reinvested income and conversions, repeating every period.
  • Genuine errors. A wrong quantity, a missed corporate action, a duplicated posting, a fee booked to the wrong period. These are the reason the exercise exists, and they are the hardest to spot when they sit among the other five.

Why the red-cell approach fails

Matching on exact equality turns every one of those families into the same signal. Reviewers quickly learn which lines are always like that, and stop looking at them. The genuine error then arrives looking like the nineteenth routine break of the day, and gets the same glance.

A flat tolerance is the usual patch, and it is a blunt one. Set it wide enough to silence rounding and it also silences a real error on a small holding. Set it tight enough to catch that error and rounding floods the list again. A tolerance only means something when it is set per break type.

What the automation should actually do

Software that merely finds more differences, faster, makes the afternoon longer. The useful version does three things before a person sees anything.

Match on more than one key

Identifier, quantity and date together, with a maintained mapping table for the cases where identifiers differ between systems. Most phantom position breaks disappear here, because the instrument was never missing. It was filed under a different name.

Classify before routing

Each break is tagged with its most likely family and carries the evidence for it. A timing break shows the expected clearing date. A price break shows both sources and both timestamps. A mapping break points at the identifier that has no counterpart. A person opening the item starts from a hypothesis rather than a blank cell.

Clear only with a recorded reason

A timing break can clear itself when the other side lands on the expected date, and the log says so. If the date passes and it has not landed, it escalates on its own. The rule is never "ignore differences under a threshold". It is "clear this kind of difference when this specific thing happens".

Position differences are the exception. We would leave them with no tolerance by default, because a missing or surplus unit is exactly the kind of difference a tolerance exists to hide.

What to leave with a person

Resolving a price difference means deciding whose price stands. That is a valuation judgement with the manager's name on it, and no matching rule should make it silently. Automation can assemble both prices, the valuation policy that applies and the history of the same holding, so the decision takes minutes instead of an afternoon. The decision itself stays human.

The same goes for anything that touches NAV sign-off, and for any break the classifier cannot place. An unclassified break is not a failure of the system. It is the system correctly admitting it does not know. For the wider question of where that line sits across a fund's back office, we set out our view in what to automate and what to leave alone.

Making it auditable

Every cleared break should leave a record: what was compared, which rule cleared it, when, and what had changed since the break first appeared. When an auditor or an LP asks how a particular NAV was verified, the answer should come from the log and not from someone's memory of a busy week.

The log also keeps the rules honest. A rule that clears a lot and never escalates deserves a second look, because either it is excellent or it is hiding something.

Where to start

Pick one reconciliation where the data already arrives in a stable format, and take a month of history. Sort its breaks by hand into the families above. The proportions will tell you which rule to build first, and which family is generating most of your afternoon.

If you cannot classify a break by hand, a system cannot classify it for you. If you can, the rule is usually a few lines long.

This is the kind of back-office rebuild we do alongside board and advisory work. If you would like to talk it through from your side of the table, you can reach us through our contact page. We can describe how we would approach it. What we cannot promise is that your administrator's data will cooperate.

Frequently asked questions

What is a reconciliation break?

A difference between two records that should agree, such as a fund's own books and its administrator's, or a position file and a cash statement. It shows the records disagree, not which one is wrong.

Are most breaks errors?

No. Many come from timing, different price sources, identifier mapping, currency rates and rounding. Genuine errors are a separate family that has to be found among those explainable differences.

Can reconciliation be fully automated?

The matching, classifying and clearing of explainable differences can be. Valuation judgements, NAV sign-off and breaks the system cannot classify should stay with a person, with the evidence assembled for them.

What tolerance should we set?

There is no universal figure. Tolerances work better set by break type: wide for rounding, tighter for prices, and none for positions. Agree the rules with your administrator and review them regularly.

How do we show an auditor a break was handled?

Keep a log for each break: what was compared, the rule or person that cleared it, when, and what changed since it first appeared. The record should answer the question without anyone reconstructing events from memory.

Mindencap

Capital markets, AI operations and cross-border build

We take board seats and advisory mandates, each bundled with a real operational rebuild.