Divergences, absorption, CVD by trade size

A divergence is when over the same stretch the price goes one way and CVD the other. For example, over an hour the price rose 1.2% while the futures delta over that hour was −$3M: more was sold at market on futures than bought, and the price rose anyway.

How is that possible if market buys move the price up? The price is the result of everything happening in the book, while CVD reflects only market orders on the chosen venues. The price could have been lifted by:

• pulled asks — sellers removed their limit orders, and the price “found it easier” to go up even without buying;
• buying on another market — e.g. the coin was bought at market on spot, which futures CVD does not see;
• a few large buys in a thin book alongside many small sells into dense bids — delta negative, price up;
• the order of trades within the stretch — sells came early in the hour, buys late; the total delta is negative, yet the hour closed higher.

A divergence says exactly one thing: the move was not made by market orders on these venues. That is a useful fact — it sends you to look at who did move it: the book, spot, other venues. But it does not tell you where the price will go.

EVA marks divergences in two ways. The «Deltas» menu has a «Highlight divergences» item — an add-on over the CVD pane (it is not enabled for the other panes). It takes the latest pair of neighbouring price highs (or lows) and compares what CVD did on the same bars; the points are labelled A and B and described in words. Classic: the price is higher at the second high while CVD is lower — the move is not confirmed by market buying. Hidden is the mirror: price lower, the indicator higher. Differences that are too small are not counted at all. The second way is the «Find» layer: one of its features is “divergence between price and flow”; it highlights bars that fall into the top shares by that feature within the window and the coin.

Why this is not a signal. 54 reversals out of 107 is 50.5% — a coin flip. Splitting by side changes nothing: 41 vs 43 on rises, 12 vs 11 on falls. Per-coin skews look tempting — “BNB continued 5 out of 7”, “DOGE reversed 23 out of 40” — but these are small samples: a 5-to-2 result on seven cases easily happens by chance, and BTC has a single case. Picking a coin from such a breakdown in hindsight means fitting the rule to the past.

There is a second check, from the EVA help: a divergence between futures and spot (archive 04–24.09, against random minutes of the same coin and hour, on two halves of the data) gave small coins a move back within the hour about 4 percentage points more often, and BTC, ETH, BNB nothing. A weak effect on a narrow group is not grounds for a decision by itself.

Most importantly — the one-picture trap. On a chart it is easy to find a divergence after which the price reversed beautifully and present it as proof. For every such picture our archive has a matching one where the price kept going. An honest check looks like this: the rule is written down in advance, all cases are taken, the shares of all outcomes are counted.

Absorption is when a strong market flow in one direction does not move the price because limit orders on the other side meet it. Market buys take the asks at a level, and the seller there keeps placing new ones — or had a very large order there from the start. The delta is large, and the candle is almost a dot.

The mechanics in the book. A market buy is filled against the best ask. While that level still has volume, the price cannot rise above it: the next ask opens only when the current one is used up. If the seller refills the order faster than it is eaten, the level holds. You recognise the picture by the combination: a strong delta or a steep CVD slope — and a price that stands still. Part of a limit order's volume may be hidden and not visible in the book in advance (some orders show only their “tip”) — then absorption can be seen only in the trades.

Let us count. An illustrative example (it is also in the figure). A limit sell of 1,500 ETH rests at 100.0. Over 15 minutes 1,100 ETH were bought at market at that level; the seller, seeing the order being eaten, added another 600 ETH. Left: 1,500 − 1,100 + 600 = 1,000 ETH — more than before the last refill, and the price never got above 100.0.

The bar's delta: in total 1,250 ETH were bought at market (1,100 at that level and 150 below) and 400 ETH sold — delta 1,250 − 400 = +850 ETH. The price opened at 99.95 and closed at 99.95. At a price of about $100 that is roughly +$85,000 of net market buying that did not move the price a single step.

What can be said: market buying was met by a seller with limit orders, and over this bar the seller withstood it. What cannot be said: that the seller has plenty left (1,000 ETH is visible, any hidden part is not), that the order will not be pulled a minute later, or that the price will now go down.

How absorption ends. In two ways, and both are common. Either the limit side runs out — the seller pulls the order or it gets eaten — and the price passes the level, often quickly: continuation. Or the market side runs out of steam — buyers stop hitting the asks — and the price may move in the limit player's direction. Which outcome will happen cannot be seen from the absorption picture itself.

In EVA such places are marked in the «Walls · Events» feed in «Book · Flow»: the kind Δ “large delta with a flat price” — over 15 minutes aggression that is large by the coin's standards went one way while the price barely moved (the move is given in ATR). The help says plainly: this is a fact about the flow, not a forecast — a check on the archive for 04–24.09 did not confirm a reversal after such events. The level where absorption happened is also visible inside the candle — in the clusters layer, covered in section 9.

The ordinary CVD adds all market orders into one line. But the pressure may come from a crowd of small orders or from a handful of large ones, and those are different pictures. CVD by trade size splits one line into three: small, medium and large orders, each with its own cumulative delta in dollars. The three lines add up to the same ordinary CVD.

Where the bucket boundaries lie is decided by the coin itself: the thresholds are computed from its own orders. Small is the bulk of this coin's orders, large is the rare top of the distribution, medium is in between. So a “large” order in BTC and in a small coin are amounts of different orders of magnitude, and that is right: you compare with what is usual for the coin. The current boundaries are written in the line's name.

In EVA this is the «CVD by trade size» pane under the chart. Its starting point is shared with the ordinary CVD — over the period, since the start of the day (UTC) or from an anchor — and is labelled the same way. If the bucket boundaries changed within the visible window (for example, the window spans several days), the pane says so plainly.

Why merging is essential. The exchange sends not orders but trades — fills at each book level. A $600,000 market buy that walked six levels arrives as six trades of about $100,000 each. If you sort the trades themselves by size, one large order lands in the “medium” or even “small” bucket six times and does not appear in “large” at all.

So CVD by size is computed from merged trades: prints of one side on one venue for one coin that arrive almost simultaneously are collected into one order — just as in the tape (section 5). Merging is an approximation: two different orders on the same side at the same instant cannot be told apart. How much it matters is shown by the live measurement from the tape section: over 67 minutes 957,298 trades merged into 295,364 orders, and for 582 of 894 large orders not a single individual trade passed the large threshold — without merging they would simply be invisible.

How to read it. The large line rising and the small one falling — large orders net-bought at market, small ones sold. The reverse — large orders sold while the crowd bought. All three the same way — everyone is pressing.

What is not here. “Large” is order size, not identity and not “smart money”: the bucket contains many different participants, including position closes, hedges, market makers' algorithms and mistakes. A large order does not know the future. And the bucket boundary itself is arbitrary: an order just below the threshold lands in medium although it is essentially no different.