CVD Explained — Cumulative Volume Delta

What is CVD (cumulative volume delta)?

Cumulative volume delta is the running sum of signed trade volume: for every executed trade, the size is added if a buyer crossed the spread and subtracted if a seller did. Plotted as a line under the chart, it shows which side has been paying the spread to get filled over a stretch of time, independently of where price actually ended up. It is the most compact summary of order flow that still fits on a price chart.

The key word is signed. Ordinary volume is unsigned — it tells you a transaction occurred but not who initiated it, and since every trade has a buyer and a seller, unsigned volume can never answer "who was in a hurry". CVD answers exactly that, and only that.

What makes it useful is the comparison. Price is the outcome of an auction; CVD is the effort spent in that auction. When the two move together, the read is boring and coherent: buyers are lifting offers and price is rising, or sellers are hitting bids and price is falling. When they come apart — CVD grinding higher while price stalls or slides — the effort is not converting into displacement, and something on the passive side is eating it. That is the whole reason the indicator exists.

CVD is not a filter, an oscillator or a threshold system. It has no overbought level and no standard setting to optimise. It is a reconstruction of a fact the exchange already published and the candle threw away.

How is delta computed from the raw trade tape?

Delta is computed by classifying every trade by aggressor side and summing with sign, bar by bar. The classification is not a guess on venues that publish it: exchanges report which side was the maker, and the aggressor is the other one. On Binance's trade stream a boolean field marks "the buyer was the market maker" — when true, a seller crossed into a resting bid; when false, a buyer lifted a resting ask.

The rest is bookkeeping, and the bookkeeping is where implementations diverge:

1. Bucket trades into bars. Each trade goes to the bar whose interval contains its timestamp. Trades landing in a still-forming bar are held back, not merged into the last closed one.
2. Sum with sign per bucket. Aggressive buy size minus aggressive sell size gives that bar's delta.
3. Accumulate. CVD at bar n is the sum of every delta up to and including n. The starting point is arbitrary, so the absolute level carries no meaning — only the shape does.
4. Stamp when it became knowable. A bar's delta only exists once that bar has closed. In the Confluence Engine each CVD point carries the bar's open time as its drawing anchor and the bar's close time as the moment the value became knowable, so nothing is ever plotted earlier than it could have been observed.

That last step sounds pedantic and is not. A great deal of order flow analysis published online is drawn with hindsight — marks placed on the bar that created them, using information that only existed several bars later. Separating "where this is drawn" from "when this was known" is the difference between a chart you can audit and a chart that flatters itself.

Delta versus cumulative delta

Delta and cumulative delta answer different questions and fail in different ways. Delta is a per-bar measurement with no memory and a meaningful absolute size; cumulative delta is a path with memory and an arbitrary origin. Reading one as if it were the other is the most common CVD mistake.

A useful refinement is the delta candle: instead of a single delta number per bar, plot delta's own open, high, low and close within the bar. Two bars with the same closing delta can have completely different intrabar paths — one that pushed positive and gave it all back, one that ground steadily up — and the delta candle restores that distinction. The Confluence Engine ships this as its own subpanel for the same reason it stores trades rather than candles: the reduction is the thing that destroys the information.

Why does CVD diverge from price?

Because aggression and displacement are separate quantities, and someone can supply liquidity against one without conceding the other. A CVD divergence is the arithmetic statement that a lot of signed volume arrived and price did not move accordingly. Mechanically, that means resting limit orders on the opposite side absorbed the flow.

The canonical construction compares swing pivots rather than eyeballing the two lines. Price makes a lower low while CVD makes a higher low: sellers pushed harder into the second low and got less for it — bullish divergence. Price makes a higher high while CVD makes a lower high: buyers spent more and got less — bearish divergence. In the Confluence Engine both pivots must be confirmed by a symmetric window of closed bars before the divergence is emitted at all, so a divergence never appears on the pivot bar itself; it appears on the bar that confirmed it.

Three warnings, because this indicator is over-read more than any other:

• A divergence is an observation of absorption, not a forecast. Absorbed flow can hold and reverse price, or the absorber can be overwhelmed and the level can break violently. Absorption covers how to tell whether the defence is holding.
• Long-horizon CVD trends drift for structural reasons that have nothing to do with conviction — persistent hedging flow, market-maker inventory patterns, venue mix. The shorter the window, the more the divergence means.
• CVD is venue-specific. A divergence that exists on one exchange's tape and nowhere else is a fact about that exchange.

Why do spot CVD and perpetual CVD disagree?

Because they are different instruments traded by different populations for different reasons, and neither is the "true" one. A perpetual swap is leveraged, funded and liquidatable, so its tape carries directional leverage, hedging and forced closes. A spot tape carries actual transfers of the asset. Persistent disagreement between the two is a positioning read, not a data error.

Reading this properly requires the same aggressor convention on both feeds, which is a real implementation trap: the maker flag has the same meaning on spot and futures streams, and inverting it on one of them produces a permanent phantom divergence that looks like signal. The Confluence Engine records both tapes with a single shared convention and tracks sustained slope disagreement between them as its own layer, alongside funding and open interest, because the three together describe positioning far better than any one of them.

What CVD cannot do

CVD cannot see hidden liquidity, cannot tell you why anyone traded, and cannot be compared across sources on its absolute value. It is one reduction of the tape, and a lossy one — it collapses the entire price dimension of a bar into a single signed number.

That last point is the reason footprint exists. CVD says a bar had positive delta; it cannot say whether that buying happened at the high of the bar (chasing) or at the low (accumulating into weakness), and those are different events with the same delta. Footprint charts restore the price axis that CVD integrates away, which is why the two are normally read together rather than as alternatives.

Finally: no order flow measurement, CVD included, generates instructions. On The Confluence Show, CVD is one of more than forty layers NAIRO reads live to build and stress-test a thesis about the current auction — and every thesis is published with the condition that would invalidate it, stated in advance. See how the engine works, or watch the show free on a delay. It is educational market analysis, not financial advice.