do-order-blocks-work

Do order blocks predict where crypto prices go next?

Order blocks are a popular trading idea: certain candles supposedly mark zones where price will react, in a predictable direction. This project tests that on two years of 15-minute Binance data for 76 altcoins (Sep 2024 - Sep 2026).

Answer: not here. The direction an order block points to is no better than the direction of a random candle picked to look similar, and no way of trading it that was tested makes money after costs. The one real effect is that order blocks mark slightly more volatile moments. It is small, it says nothing about direction, and nothing here turns it into a profitable trade.

The results in one table

Two definitions of an order block were tested, each through the same pipeline:

• Round 1 uses the detector from a live alerting bot: high-volume, long-wick candles near a swing high or low, pointing the way the candle closed.
• Round 2 uses the textbook ("smart money concepts") definition: the last opposite-coloured candle before a break of structure, pointing the way of the break.

	Round 1	Round 2
Zones found	152,266	218,217
Direction vs random look-alike candle, 4h ahead	−0.31 bps [−3.38, +2.69]	+0.50 bps [−2.17, +3.63]
Extra volatility vs random look-alike candle, 4h ahead	+2.35 bps [+1.88, +2.87]	+0.77 bps [+0.34, +1.34]
Best average trade after costs	−8.86 bps	−7.6 bps
Months profitable after costs (walk-forward)	0 of 21	3 of 21
Deflated Sharpe ratio (corrected for configurations tried)	0.00	0.00

bps = basis points (0.01%). Brackets are 95% intervals; an interval that contains zero means no detectable difference. Costs are modelled on MEXC futures taker fees.

How it was tested

• Against look-alike candles, not against zero. Every order block is compared with random candles from the same coin, within 24 hours, with similar volume and volatility. This separates "order blocks predict direction" from "order blocks happen at busy moments".
• The test can find an edge when there is one. A planted +25 bps edge is detected, and random trade directions are flagged as significant 4 times in 60 at a nominal 5%.
• No peeking at the future. A leakage audit checks that no feature or label uses data from after the moment it describes.
• Every configuration tried is counted. A trial ledger records every variant evaluated, and the headline statistics are corrected for that number.

What this does and does not show

• Shown, for this data: for these 76 coins, 15-minute candles, these two years and these cost assumptions, neither definition's direction beats a look-alike candle, and no tested configuration is profitable after costs.
• Not shown: that order blocks never work. Other timeframes, higher-timeframe structure, combining with other signals, other markets and limit-order execution were not tested.
• Only observed, not explained: the volatility effect, and the improvement from waiting for price to come back to the zone before entering.
• Exploratory: only one further test (Family A, under "Read more") was pre-registered; rounds 1 and 2 were not (docs/deviations.md).

Known limitations: costs are modelled rather than measured, the coin list was chosen by recent volume, some thresholds in round 1 were set by hand, and when one 15-minute bar hits both the target and the stop, the lower level is assumed to be hit first: the stop for a long, but the target for a short, which flatters short trades. Each is discussed, with the direction of its bias, in docs/methodology.md and docs/PROVENANCE.md.

Try it (about 5 minutes, no download)

Requires Python 3.11-3.14.

python -m venv .venv && source .venv/bin/activate      # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
pytest -m "not integration"     # unit tests, under a minute
hawk-research smoke              # the full pipeline on a synthetic market, about 2 minutes

The smoke run builds a fake market with planted order-block shapes and runs every stage end to end. Its numbers mean nothing about real markets; it shows that the pipeline runs and gives the same answer on every machine. (The command is called hawk-research after the alerting bot the study started from.)

To reproduce the published numbers from real Binance data (about 200 MB to download, then about 8 minutes on a 12-core laptop):

make download reproduce

Full instructions, the pipeline stages, configuration and the repository layout are in docs/REPRODUCING.md.

Read more

• docs/OVERVIEW.md: a two-page summary of everything tested.
• docs/methodology.md: how each test works.
• docs/archive/round1_findings.md and docs/archive/round2_smc_findings.md: full results.
• docs/archive/s1_s5_results.md: Family A, two refined systems tested against a pre-registration; both failed.
• experiments/: three side studies (zone straddles, a daily fade, market-neutral pairs); none profitable.
• docs/deviations.md: every place the work departs from its own protocol.
• docs/PROVENANCE.md: how the project came about, what cannot be verified, and the use of AI in building it.

License

MIT; see LICENSE.