YTD-vs-MTD

YTD is not the sum of MTD

The single most common source of wrong numbers in management reporting, and it has nothing to do with AI.

It’s 4:40 on a Thursday. The board pack goes out at six.

Ask for a client’s revenue for the quarter. Somewhere, a query sums a column across three months. If that column holds monthly revenue, the answer is right. If it holds year-to-date revenue, the answer is nearly double — and nothing about it looks wrong.

Flows and positions

Flows and positions
Management reporting mixes two kinds of number, and they behave differently over time.
Flows measure what happened during a period: revenue month-to-date, fees, new lending. They add up across months. January plus February plus March is the quarter.
Positions measure where things stood at a point in time: year-to-date totals, period-end balances, exposures, risk-weighted assets. They are either already cumulative or a snapshot. The quarter’s value is the value at the latest period in scope — not the total of every month in it.
Both kinds can be summed across clients, desks or regions. Only flows can be summed across time. That is what makes positions semi-additive, and it’s why the same SUM that is correct for one column is wrong for the one next to it.

The wrong figure and the right one

One client, one quarter. Illustrative numbers.
Month Revenue MTD Revenue YTD Exposure (period-end)
Jan 120 120 50
Feb 130 250 52
Mar 150 400 55
Summed across the quarter 400 ✓ 770 ✗ 157 ✗
Correct quarter value 400 400 55
Revenue MTD sums correctly to 400. Revenue YTD, summed the same way, reports 770 — 1.9 times the truth. Exposure, summed, reports a balance of 157 that the client has never held; the real position at quarter-end is 55.
None of these wrong figures fails a sanity check at a glance. 770 is a believable quarterly revenue number. It is only wrong in the way that matters.
Ratios are a third case. A return on RWA can’t be summed across months, and it can’t be averaged either — the average of three monthly ratios is not the quarter’s ratio. It has to be recomputed from its components over the whole period.

Why it keeps happening

Every finance team knows the rule. The error persists because the rule lives in people, not in the system. Each new report, each new analyst and each new query rediscovers it — or doesn’t. Column names don’t help: revenue and revenue_ytd sit side by side, and a query that picks the wrong one runs cleanly.
Hand the query to a language model and you’ve given the rule to the participant least able to apply it consistently. It may get it right nine times in ten. The tenth goes to the board.

Put the rule where every question passes

In Equative Insights OS, aggregation behaviour is a property of the measure, not of the query. Each measure in the semantic layer is defined once, with its additivity attached:

Ask Client IQ for a client’s quarterly exposure and the question resolves to the exposure measure, which already knows it is a position. The deterministic computation stage reads the quarter-end value. Nobody needs to remember the rule, because nothing is left to remember.
And the rule is stated back. When an answer uses a semi-additive measure, the reply carries a caveat — for example, that exposure is shown at March, the latest period in the selection, rather than summed across months. The trace shows the same decision at the resolution and computation stages. A reader who expected a total learns in one line why they got a balance.

The test for any system

If you’re evaluating an AI reporting tool — ours or anyone’s — ask it for quarterly year-to-date revenue, then ask where the decision “should this be summed?” lives. If it lives in a prompt, a model or an analyst’s memory, it will eventually be wrong. If it lives in the definition of the measure, it is right the same way every time.
If the answers get vague, you’ve spotted the pitfall before it spots you.
Take the rules with you. Download the measures one-pager [link] — one page on which measures sum across time, which read at period end, and which must be recomputed. Useful for anyone who builds a report, whether or not they use Equative.

Recent Posts

Learn why enterprise AI costs rise at rollout, how hidden model calls inflate cost per answer, and how to measure and optimize AI spend.
enterprise-AI-access-control
RoRWA