Cumulative delta

Cumulative delta is a running total of volume that traded at the ask minus volume that traded at the bid, accumulated across a session or across an unbroken series of sessions.

What it measures

Every completed trade has two sides, so volume itself is symmetrical. One party wanted the trade badly enough to cross the spread; the other was already sitting in the book, waiting. Delta is an attempt to separate those two roles and count only the first.

When a trade prints at the offer, the buyer is treated as the aggressor and the volume is added. When a trade prints at the bid, the seller is treated as the aggressor and the volume is subtracted. A single bar's delta is the net of those two numbers. Cumulative delta simply refuses to reset it, carrying the running total forward so the line reflects the whole session rather than one bar.

What you are looking at, then, is a measure of net aggression over time. Not net buying, which is impossible, since every contract bought was also sold. Net impatience.

One bar of a footprint chart, broken out by price level. The left column counts contracts that traded at the bid, the right column those that traded at the ask. Summing the level deltas gives the bar delta. Cumulative delta is what you get when you never reset that total.

How it is calculated

The arithmetic is trivial. The classification is not.

An exchange feed does not label a trade "buyer-initiated". It publishes a print: a price, a size, a timestamp. Whether that print counts as a buy or a sell is inferred, by comparing the trade price against the prevailing best bid and offer at the moment the print arrived. Prints at or above the offer are classified as buys, prints at or below the bid as sells. Prints that land between the two, which happens more often than people assume, are handled by a fallback rule, usually the tick rule: compare against the previous trade price and inherit its direction.

This inference is the entire foundation of every delta number you will ever see, and it is a reconstruction, not a measurement.

Why your delta does not match someone else's

Two traders watching the same instrument in the same minute routinely see different cumulative delta values. The usual causes are all mechanical: a different data feed, a different tick filter, a different session start time, unfiltered versus filtered ticks, a platform that aggregates prints occurring in the same millisecond, or one platform reading market-by-order data while another reads market-by-price. None of this is a bug. Delta is a derived figure, and derivations differ.

Reading the line

Cumulative delta is plotted either as its own subgraph or as a candle series in its own right, with an open, high, low and close built from the running total. The candle form exists so that the delta series can be examined for its own internal structure rather than only its level.

The observation most often attached to it is divergence: cumulative delta and price moving in opposite directions over some window. Price makes a higher high while cumulative delta makes a lower high, or the reverse. Practitioners describe this as aggressive activity failing to move price, and interpret it as evidence that resting liquidity on the other side is large enough to absorb it. That is the claim. Whether it holds in any particular market, timeframe or session is an empirical question, and the honest answer is that it depends heavily on the instrument.

A cumulative delta line across a single session. The downward steps are periods where sellers were the more aggressive side; the total still ends positive. Note that the line says nothing at all about price, which is the point and also the trap.

Limitations

These are not caveats to skim. They are the reason cumulative delta is misread so often.

It is inferred, not reported. Every classification rests on matching a print to a quote that was moving at the time. In fast conditions the quote and the print can be milliseconds apart, and the classification degrades exactly when the market is most active.
Aggression is not conviction. A market order can come from a discretionary trader with a view, a fund rebalancing on a schedule, a market maker hedging an options position, or a stop being triggered. Delta counts all of them identically. It cannot tell you why anyone crossed the spread.
It has no memory of size distribution. Ten thousand contracts of delta from one participant and from four thousand participants produce the same number.
Its level is arbitrary. Where you start the accumulation determines every subsequent value. Session-reset, weekly and continuous cumulative delta will disagree about whether the current reading is high or low, and none of them is more correct than the others.
In fragmented markets it is partial. For instruments trading across several venues, delta computed from one feed describes one venue. Centralised futures markets are cleaner in this respect than equities, which is why the tool is more common in futures.
Divergence is common. It appears constantly, including in stretches that resolve in no particular direction. Counting only the instances you remember is not evidence.

Provenance

The idea behind delta is much older than the calculation. Tape readers in the early twentieth century were doing something structurally similar by eye: watching the ticker and forming a judgement about which side was pressing. Richard Wyckoff wrote about reading the tape for the relationship between effort and result, which is the same question delta divergence asks in numerical form.

What changed was data. Once electronic order books published quotes and prints with timestamps, the classification could be automated and the running total could be plotted. Footprint and cluster charting, which display delta at each individual price rather than per bar, followed the arrival of that data in retail platforms rather than preceding it.