← Back to Blog

Tested, not trusted: the order block

About this artifact

essayon-the-recordmaintained

markets · since 2026

Hobbyist research, not investment advice.

A glossary card defining the order block collected thirteen thousand likes in the inspiration batch I keep for this series. It had clean typography, a tidy candle diagram, and no evidence of any kind. That is the genre: definitional confidence, zero base rates. This series republishes the format with the one column the genre never ships. Here is the definition, and here is what it tested at.

Two glossary cards side by side: the left card carries only a definition and no evidence, the right card carries the same definition plus a highlighted measured-verdict strip showing a net t-statistic of minus 0.91 across 3,558 trades.

The definition, made executable#

Prose definitions of the order block do not commit to anything a computer can check, so the first step was writing one down that does. The last opposing candle before a displacement move, where displacement is required and measured against the recent swing structure with a 5-bar lookback. The tradeable rule: enter in the block's direction at the close of the first bar whose range retests the zone, searching from two bars after the block so the displacement is fully observable, and exit after a fixed N days. No discretion anywhere. Every parameter is in the artifact.

The test#

Thirty large-cap US names, daily bars from 2021-07-23 through 2026-07-21, 37,560 pooled periods. Horizons of 3, 5, and 10 days, a small declared sweep over the swing lookback, and 10 basis points of cost per side. Nine variants total, declared up front, so the overfitting check has honest inputs.

The verdict#

Metric (default variant, 5-day exit) Value
Trades 3,558
Hit rate 48.2%
Annualized Sharpe, net -0.07
Net t-statistic -0.91
Probability of backtest overfitting (9 variants) 0.57

The strict, displacement-required order block retest earned nothing net of costs in this window. The t-statistic is not just small, it is slightly negative. And the PBO of 0.57 says the best-looking variant in the sweep is more likely than not an artifact of picking the best-looking variant. If the edge is real, it is not in the daily-bar retest rule the glossary cards describe.

The honest boundaries: one five-year window, one universe of liquid large-caps, daily bars, fixed exits. A definition this popular deserves retests on other intervals and universes, and the harness is built to run them. What it does not deserve is to be taught as an edge before anyone has measured it.

The number is pinned to its run. Detector v1.0.0 at git f9f4f50, parameter hash a702f152...d99d3, artifact order_block.json from the evaluation run of 2026-07-22. The recorded config reproduces this table, and the hash is what keeps me honest about which config that was.

This is historical analysis of a public trading concept, not investment advice, and not a forward claim about any market. Nothing here is a recommendation to trade anything.

The glossary card asks you to trust the diagram. The better ask is smaller: trust nothing, test the definition, and let the base rate do the talking.

Questions this post answers

What is an order block?
In the smart-money-concepts vocabulary, an order block is the last opposing candle before a strong displacement move, treated as a zone where institutions supposedly left unfilled orders. The tradeable version: when price later retests that zone, enter in the direction of the original move.
Does the order block retest entry work?
Not in this test. Across 30 large-cap US names, 2021 to 2026, daily bars, the strict displacement-required definition with a 5-day exit produced 3,558 trades, a 48.2% hit rate, and an annualized Sharpe of -0.07 net of 10 basis points per side. The net t-statistic is -0.91: indistinguishable from zero, leaning negative.

Keep reading

Tool

Market's Best

The top-graded stocks from the latest market scan. No sign-in needed.

Read
Demo

Grade my portfolio

Run a sample portfolio through the investor committee.

Read
Case study

Factor-First AI Investment Platform Narrated by a Six-Persona Committee

Grew a single-ticker grader into a full investment platform: a four-factor composite (Quality, Valuation, Momentum, Health) narrated by a six-persona committee, a nightly scan of 600+ large caps, portfolio and net-worth tracking, and daily measurement of historical ranking behavior.

Read
Post

Autopsy: the 5/10 strategy that never lost

A viral post claimed a simple 5/10 moving-average strategy that has not lost in eight years. I codified the most charitable readings of the claim, ran each one over SPY with costs, and checked the claim against its own success metric. It fails under every interpretation, and the receipts are hashed.

Read
Post

How to backtest without fooling yourself

A backtest's job is not to find an edge. It is to stop you from believing in one that is not there. The toolkit I used to test my own trading engine, and the part where it killed my single best signal.

Read
Post

Collect Everything, Trust Almost Nothing

My market tracker now listens to ten different crowds: message boards, raw forum archives, search trends, encyclopedia traffic, insider filings, video chatter, news tone. The interesting design decision is not the collection. It is the wall that keeps every one of those feeds away from the number that actually grades a stock, and the mechanics that make the wall structural instead of remembered.

Read

Follow the work

New tools and writing as they ship — pick a channel.

Written by Eric Caskey. I build AI tools you can actually use. Explore the Tools or see the case studies.