Do dark pool levels hold?
Off-exchange volume at price, from FINRA's print files, against every other price at the same distance
Key finding. Dark pool levels are sold as the prices where institutions left size and where price comes back to react. Built from 609 million off-exchange prints in FINRA's public files and scored touch by touch against every other price at the same distance, the highest-volume levels hold 38.4% of the time against 37.7% for the rest; once the time price spent at each level is controlled, the top-level flag is worth -0.2 points, with a standard error of 0.7.
Most options-flow platforms now draw dark pool levels on the chart: the prices where large off-exchange trades printed, summed over a window or taken from the biggest single prints. The pitch that comes with them is that price reacts there. In the vendors' own help pages:
- Quant Data: "The larger the level for a particular ticker, the more likely that the specific price level will act like a magnet once that level is reached."
- Cheddar Flow: "These type of orders are know to create support and resistance levels."
- Unusual Whales: "If price holds above a massive print level, it is often bullish."
- BlackBoxStocks: its dark pool volume profile is described as "allowing you to uncover secondary support and resistance levels".
- Tradytics, on a worked example: "when the price came down to the 193 level, it immediately found support at that level and reversed back."
The evidence offered is the same everywhere: a handful of charts where price turned at a line. None of these pages publishes a hold rate, a sample size or a baseline for its levels. One vendor draws the lines and declines the claim: Skylit's Atlas says it "never calls direction: it shows where the levels are, and you judge what price does there." That is the honest position, and this post is the measurement behind it.
The claim is testable, and the data to test it is public. FINRA publishes every off-exchange short sale reported to its three trade reporting facilities, with the time to the second, the price and the size, in monthly files anyone can download. That is the raw material of a dark pool level: where off-exchange volume traded, and how much. We built the levels from those files and asked one question. When price comes back to a high-volume dark pool price, does it turn there more often than at any other price the same distance away?
What a touch is, and what holding means
Prices are cut into bins 5 basis points wide (about 35 cents on SPY at 700). For each symbol and session, the level surface comes from the sessions before it only: the off-exchange volume in each bin, ranked against every bin within 3% of the prior close. The session's own prints never enter its levels. Then the session is replayed minute by minute. The first time price reaches a bin other than the opening one is a touch: from above it is a support test, from below a resistance test. The touch held if price moved 4 bins (20 basis points) back away from the level before it traded 4 bins through it. Touches that resolved neither way by the close are dropped.
Every bin price touches is scored, not only the ones a vendor would draw. That is the control: a level that means something should hold more often than the prices around it, at the same distance from the open, on the same side, at the same time of day.
The result
| level surface | touches | top-decile levels held | all other prices | top-level effect, controlled |
|---|---|---|---|---|
| last 5 sessions, 20 bp test | 134,592 | 38.4% | 37.7% | −0.2 pp (±0.7) |
| last 5 sessions, 40 bp test | 128,426 | 41.3% | 42.2% | −1.3 pp (±0.8) |
| last 20 sessions, 20 bp test | 127,529 | 37.6% | 37.6% | −0.3 pp (±0.7) |
| block prints only ($1M and up), 5 sessions | 134,592 | rank of block volume at price | +1.1 pp (±1.2) | |
"Top-level effect, controlled" is the change in the hold rate for a price in the top tenth of dark pool volume, from a linear probability model with the distance from the open, the side, the time of day, the symbol, the time price had spent at that level over the same window, and the prior session's high, low and close as controls; standard errors are clustered by trading day. In no row does the top level hold better than an ordinary price. The raw split in the first two columns says the same thing before any control is applied.
Blocks, the prints of a million dollars and more that are usually singled out as institutional, add nothing either: +1.1 points with a standard error of 1.2 on the five-session surface, +0.9 (±1.1) on the twenty-session one.
The one trace, and why it is not a level
There is one pattern in the data, and it runs the other way from the pitch. Prices where no off-exchange volume traded over the last five sessions held 32.1% of the time, against 36% to 39% for every other decile, which are flat among themselves. Price moves through air it has not traded in recently; it does not stop harder where the volume piled up. The effect needs the freshest window (it is gone on twenty sessions) and the tightest test (gone at 40 basis points), and the volume-rank slope behind it stands at 1.1 and 2.0 standard errors in the two halves of the year. It describes empty space, not a line to draw.
Why volume at price looks like support
Volume piles up where price spent time. A price the market chopped around for two days carries a tall bar in any volume profile, dark or lit, and price returning to a range it just traded in often chops there again. Without the time-at-price control, the dark pool volume rank does show a small positive slope (+2.9 points from the lowest to the highest rank, t 2.7). That slope is market profile restated, not information about who traded off exchange, and the top-decile flag that a level tool actually draws is flat with or without it.
The index version
A second popular reading of the same files is a daily index: the short share of off-exchange volume, weighted across S&P 500 components, read as a proxy for investor buying and a signal for the index over the following days. We built it from FINRA's daily consolidated files for all 1,108 SPX sessions from 2022-04-14 to 2026-09-23, share-weighted rather than dollar-weighted because the files carry no price, and z-scored against its own trailing year. It does not move with the next day's, week's or month's SPX return (every |t| below 1 for the S&P index, with Newey-West errors and past returns controlled), and its association with the next session's range disappears once past returns and the opening implied volatility are in. The top fifth of readings was followed by a smaller five-day return than the bottom fifth.
What the data can and cannot say
- Short sales only. FINRA's per-print files carry off-exchange short sales, roughly half of off-exchange volume. For a volume-at-price surface that is a large and, on the price axis, representative sample; on the usual reading, a market maker filling a buyer off exchange sells short, so these are the prints a dark pool level is supposed to capture. A full-tape test would settle it; the short half shows no trace of the claimed effect to begin with.
- Fifteen liquid names, one year. SPY, QQQ, IWM, DIA, AAPL, MSFT, NVDA, AMZN, META, GOOGL, TSLA, AMD, AVGO, NFLX, JPM. Thin names were not tested.
- Our construction, not theirs. No vendor publishes its exact level recipe, so this tests the idea every one of them describes, off-exchange volume at price, not any one product's lines. The question for a product is the same one: what is its hold rate against the prices around it?
Reproduce it
Everything here is public. The per-print files are at
https://cdn.finra.org/equity/regsho/monthly/{FAC}sh{YYYYMM}.zip, where FAC is FNYX
(FINRA/NYSE TRF), FNQC (FINRA/Nasdaq TRF Chicago), FNRA (the ADF) and FNSQ (FINRA/Nasdaq TRF
Carteret, split into four parts, _1 to _4); each line is
MarketCenter|Symbol|Date|Time|ShortType|Size|Price|LinkIndicator. The daily consolidated
short-volume file is https://cdn.finra.org/equity/regsho/daily/CNMSshvol{YYYYMMDD}.txt.
Data: FINRA. The minute path of each session is the last print of each minute, so no other price
feed is needed.
Method
Regular-hours prints only (09:30 to 16:00 ET). Bins are 5 basis points in log price. The level surface for a session is built from the N sessions before it (N = 5 or 20), smoothed over the neighbouring bin on each side, and ranked as a percentile among the bins within 3% of the prior close; the top decile is the "dark pool level". Time at price is the number of minutes the minute path sat in each bin over the same window, ranked the same way. A touch is the first minute the path reaches a bin other than the opening bin; a jump across several bins touches each of them. Hold and break are symmetric at R bins (R = 4 or 8), first to occur wins, unresolved touches dropped. The model is a linear probability model with fixed effects for distance band (4 bins), hour of day and symbol, plus side, time-at-price rank and a flag for bins within one bin of the prior session's high, low or close; errors are clustered by trading day (246 days). The daily index uses the CNMS files: short volume over total off-exchange volume, summed across the current S&P 500 components, against forward SPX close-to-close returns at 1, 5 and 20 sessions and the next session's high-low range, with Newey-West errors at a lag equal to the horizon.
Part of FirmTape research. Measured on the free session archive; every session is free to replay.