Does volume tell you which way a breakout will go?

How the tests were run

One simple, mechanical breakout was measured hundreds of thousands of times. The trades were then sorted by how busy volume was, and the busiest were compared with the quietest.

How to read the charts: the luck range

Split any set of trades into two random piles and their averages will differ a little, by chance alone. The luck range is how big that difference gets 95% of the time. If volume really mattered, the gap between busy and quiet trades would land outside it, in both periods. Every chart below shows the gap as a dot and the luck range as a shaded band. The "Could this be luck?" button shows random splits piling up.

Every question, every market, on one screen

Each cell is the gap for the side named in the question (2018-22 and 2023-26). A ▲ or ▼ means the gap was bigger than luck, better or worse. A hollow circle means it was within luck. On the index markets almost every cell is hollow, and the few that aren't don't repeat in the other period. The Bitcoin columns are full of arrows, for the reason explained in the next section. The stocks columns have a few, for a smaller dose of the same reason.

Bitcoin's volume effect was the fees

Every test was run again on Bitcoin, on Binance (2020-2026) and Coinbase (2018-2026). The rules were the same, with a 24-hour day starting at midnight UTC (4 PM PT in winter, 5 PM PT in summer). The written expectation was that nothing would pass. By the pre-set rule, five of the direction tests passed. High-volume breaks lost 0.10 to 0.22R less per trade than low-volume ones, on both exchanges.

The reason is arithmetic, not volume. Crypto fees are a percentage of price, and a quiet setup bar is usually a narrow one. The same fee takes a bigger slice of the risk.

Switch between the two views. After fees, most gaps sit outside the luck range. Before fees, every one falls inside it.

Two more facts keep this from being an edge. Every group lost money. Plain breakouts on Bitcoin averaged −0.17 to −0.29R per trade in 2023-2026 after fees, and even the high-volume third averaged −0.06 to −0.21R. And "skip narrow setups when fees are a percentage" is a rule about bar size and fees. Volume was only standing in for range.

The before-fee check was added after a first run stopped on a data error, before any result had been read. The pre-registration had already flagged that fees work through stop size.

Single stocks: the same answer, and one pattern that held

Volume is discussed most for stocks, so every test was run again on five of the most traded US stocks: AAPL, MSFT, NVDA, AMZN and TSLA. The window is May 2018 to October 2026, regular hours, with a 0.02% round-trip cost (about 4 cents a share on a $200 stock).

The five were pooled, counting each day once, and each stock was also checked on its own. The pass rule was set before the run. After the Bitcoin result, it also required a gap to survive before costs.

The near miss was the cost again

High-volume setup bars did 0.05R better than quiet ones in 2023-2026, on all five stocks. But a percentage cost takes a bigger bite out of a narrow, quiet bar, as on Bitcoin. Before costs, the gap halves and falls inside luck. In 2018-2022 it was only 0.01R. The switch below shows every test both ways.

The lead: busy days and yesterday's high or low

One pattern in the stock data was steady, though it was not one of the pre-set questions. When a stock broke yesterday's high or low on a busy day (session volume high for the time of day), the break held into the close about 64% of the time, against about 56% on quiet days. That held in both periods and on all five stocks. It fits the "stocks in play" idea: on days a stock is busy, its moves carry further. It is about how far a move goes on a busy day, not which way, and it did not show up clearly on NQ, ES or SPY.

Checking it fairly. Because it was found by looking, it did not count until it repeated on stocks it was not found on. The test was written down first (pass: busy days ahead overall, with at least 4 of 5 new stocks agreeing) and run on META, AMD, GOOGL, AVGO and COST. It held. Busy-day breaks of yesterday's high or low closed beyond the level 63% of the time, against 55% on quiet days. That is a gap of 8 points against a luck range of ±2.7 points, and all five new stocks pointed the same way (AMD +10, GOOGL +9, AVGO +10, COST +6 points, and META +2 on data from mid-2022 only, since it traded under the symbol FB before that).

What it is, and what it isn't. It uses only volume known before the break, so it is not hindsight. But it describes how far a move carries on a busy day, the volume-and-volatility link again, not a buy or sell signal. It has not been tested as a trade after costs. It did not show clearly on NQ, ES or SPY, which fits the idea that it is about a single stock being "in play" rather than the whole market being busy.

One limit on the stock data. The volume here is trades on the Nasdaq exchange only, not the whole market. Dark pools and other off-exchange venues handle a large share of US stock volume, and they are not in these numbers. Relative volume is a ratio against the same stock's own history, so a steady venue share cancels out. Shifts in where trades happen do not. See what was not tested.

What volume is good for: how far, not which way

Busy markets move more. That is the oldest result about volume, and it held here. The question is whether volume adds anything to what recent price ranges already tell you, and whether that changes a decision.

Did it change a decision? No.

Exiting on a volume spike cost money

Same breakouts, managed minute by minute. The tests covered exiting on a volume spike (three times normal), with the trade or against it, and using volume to pick a wider or tighter stop.

What was not tested, and why it might matter

This study covers plain bar volume, the number under every candle. A null result here says nothing about richer kinds of volume data. These are the gaps.

Safeguards against false positives

• Written down first. Each round's questions, pass rules and data were committed to a public log before its results were run. Nine rounds: direction, range and exits, a forward check, new symbols, published claims, decisions, Bitcoin, single stocks, and a confirmation on five more stocks.
• One rule to pass, set in advance. The gap had to point the same way in both periods, be well outside the luck range in 2023-2026 after correcting for running nine tests at once, and point the same way on the other markets.
• Real fills. Entry is the next 1-minute open after the signal, never the signal price, and stops and exits are checked minute by minute on real (unadjusted) prices, after costs: 0.5 point round trip on NQ and ES, $0.02 a share on SPY, 0.10% on Bitcoin, 0.02% on stocks.
• Days count together. Trades on the same day tend to win or lose together, so the luck ranges treat each day as one unit. That makes them wider, and harder to beat, than counting every trade separately.
• Ideas found by looking were re-tested. Four patterns that looked steady after round 1 were checked on fresh 2026 data. None held up. The closest, high-volume breaks of yesterday's high or low failing more, then reversed on five new symbols.
• A bug was found and fixed. Rebuilding the code for public release showed that the first SPY runs left a timeframe-continuity value unset for trades in the last half hour (12:30 to 1:00 PM PT). The corrected numbers are shown. No conclusion changed. NQ and ES were not affected and reproduce exactly.

Questions a skeptic would ask

Isn't volume supposed to matter only with a real signal and timeframe continuity?

That is how it was tested. Every trade here is a Strat signal (an in-force break, 2-1-2, 3-1-2, a 2-2 reversal or an inside-bar break), and two tests added full continuity on the 60-minute, daily and weekly. With continuity, high volume did slightly worse, not better. As a heads-up that the day may be wide, volume carries a little information. As a filter on the signal, it added nothing.

Does high volume confirm a breakout?

Not in this data. On NQ and ES futures and SPY from 2018 to 2026, breakouts with heavy volume on the setup bar, the breakout minute or the session finished about the same as breakouts on light volume, with gaps inside what luck alone produces.

Are low-volume breakouts more likely to be fakeouts?

No. The quietest third of breakouts reached +1R before the stop about as often as the busiest third, near 50% on every index market, and averaged the same result.

Does low volume on a pullback mean the trend will continue?

Not in this data. Inside bars where volume dried up broke no better than other inside bars. With timeframe continuity, a quiet setup bar leaned slightly better across 2018-2026 but was within luck everywhere and flipped on fresh 2026 data.

Maybe it works on a different timeframe or market?

The tests covered 15, 30 and 60-minute bars separately, two futures markets, an ETF, Bitcoin on two exchanges and ten large stocks. The answer was the same on each. Daily and weekly setups, small stocks, picking stocks by volume, and overnight futures were not tested. See what was not tested.

Doesn't everyone know volume confirms moves? Some studies found it.

The studies that find volume useful mostly pick among many stocks on a given day, which was not tested here. Two published claims were testable. The first half-hour predicting the last half-hour more on busy days did not hold. Index returns reversing more after high-volume days partly held on daily data, not intraday. Others who measured volume confirmation and found no consistent relationship match this result.

What is trading volume good for, then?

Forecasting how big the move or the day may be. Volume improved range forecasts by about 2-4% on futures, mostly from overnight volume, but that did not change a sizing or premium-selling decision in these tests.

Does volume work on Bitcoin?

No. On Binance and Coinbase, high-volume breaks lost less money after a 0.10% fee, but only because the fee costs narrow, quiet setups more in R. Before fees the gaps vanished, and every group still lost money.

Does volume work on individual stocks?

Not as a direction signal on large stocks. On AAPL, MSFT, NVDA, AMZN and TSLA from 2018 to 2026, 0 of 9 tests passed, and the near miss came from trading costs, not volume. One pattern held up on five other stocks: breaks of yesterday's high or low held into the close more often on busy days (about 63% against 55%), a sign of how far a move carries on a busy day rather than a direction signal.

Doesn't dark-pool volume change this for stocks?

It might. It was not tested. The stock volume here is one exchange's trades, not the full tape, so off-exchange and dark-pool volume is missing. Relative volume compares a stock with its own history, which cancels a steady venue share, but block prints and shifts between venues were not studied.

Reproduce the results

The code is public in 345-strategies/strat-enrichment, folder volume/. The market data is licensed and can't be shared, so bring your own 1-minute bars. They cost a few dollars from a vendor such as Databento, and Bitcoin is free from Binance. Fed the same bars, the code reproduces the NQ and ES tables digit for digit. All results on this page are also available as JSON and as plain text.

1. Install

git clone https://github.com/345-strategies/strat-enrichment
cd strat-enrichment
pip install -e ".[dev]" databento
pytest tests/test_volume.py   # runs the whole study on synthetic bars

2. Get 1-minute bars

python - <<'EOF'
import databento as db
c = db.Historical()   # reads DATABENTO_API_KEY
for root, start in (("NQ", "2018-01-01"), ("ES", "2019-12-01")):
    c.timeseries.get_range(dataset="GLBX.MDP3", schema="ohlcv-1m", stype_in="parent",
        symbols=[f"{root}.FUT"], start=start, end="2026-09-23").to_file(f"{root.lower()}_raw.dbn.zst")
EOF
python volume/prep_futures.py NQ data/nq.parquet nq_raw.dbn.zst --end 2026-09-22
python volume/prep_futures.py ES data/es.parquet es_raw.dbn.zst --start 2020-01-01 --end 2026-05-15
python volume/prep_stocks.py data/spy.parquet spy_arca.dbn.zst --start 2018-05-01 --end 2026-04-21
python volume/prep_stocks.py data/spyfwd.parquet spy_fwd.csv --start 2026-05-08 --end 2026-10-19

SPY from 2018: NYSE Arca bars (ARCX.PILLAR) work for prices and volume. The published runs used consolidated prices with Arca volume. To match, pass consolidated bars as the file and Arca bars with --volume-from. The forward files start May 8, 2026 because relative volume needs 20 prior sessions.

3. Run the index rounds

python volume/run.py data/ results/

4. Bitcoin, free data

python volume/fetch_binance.py BTCUSDT 2020-01 2026-05 raw/binance/
python volume/prep_crypto.py data/btc.parquet raw/binance/*.zip --start 2020-01-01 --end 2026-05-04
python volume/prep_crypto.py data/btccb.parquet coinbase_btcusd_1m.csv --start 2018-01-01 --end 2026-04-11
python volume/volstudy.py data/btc.parquet btc results/work --market crypto
python volume/volstudy.py data/btccb.parquet btccb results/work --market crypto
cd volume
python analyze_btc.py ../results/work ../results/RESULTS_ROUND7_BTC.md
python lit.py ../results/work ../results/RESULTS_ROUND7_BTC_LIT.md btc btccb

5. Stocks

python - <<'EOF'
import databento as db
c = db.Historical()
for s in ("AAPL", "MSFT", "NVDA", "AMZN", "TSLA"):
    c.timeseries.get_range(dataset="XNAS.ITCH", schema="ohlcv-1m", symbols=[s],
        start="2018-05-01", end="2026-10-24").to_file(f"{s.lower()}.dbn.zst")
EOF
python volume/prep_stocks.py data/aapl.parquet aapl.dbn.zst --start 2018-05-01 --end 2026-10-23 --splits 2020-08-31:4
python volume/prep_stocks.py data/msft.parquet msft.dbn.zst --start 2018-05-01 --end 2026-10-23
python volume/prep_stocks.py data/nvda.parquet nvda.dbn.zst --start 2018-05-01 --end 2026-10-23 --splits 2021-07-20:4,2024-06-10:10
python volume/prep_stocks.py data/amzn.parquet amzn.dbn.zst --start 2018-05-01 --end 2026-10-23 --splits 2022-06-06:20
python volume/prep_stocks.py data/tsla.parquet tsla.dbn.zst --start 2018-05-01 --end 2026-10-23 --splits 2020-08-31:5,2022-08-25:3
for s in aapl msft nvda amzn tsla; do
  python volume/volstudy.py data/$s.parquet $s results/work --market stocks --cost-pct 0.0002
done
cd volume
python analyze_stocks.py ../results/work ../results/RESULTS_ROUND8_STOCKS.md aapl msft nvda amzn tsla

Five stocks for eight and a half years cost about $5 in Databento credit. --splits divides earlier prices by each split ratio (and multiplies volume), so a split doesn't look like a crash.

Expect the last digit or two to differ with bars pulled fresh from a vendor. The conclusions should not move. If one does, please open an issue. Not reproducible from public code: the sizing and straddle decisions, because that breakout system's trades and the option quotes are not public.