511 SMC and Dow Theory rules from books. Two were testable. The engines are a different story.

August 14, 2026

We extracted 358 Smart Money Concepts rules and 153 Dow Theory rules from trading books. Two ever reached twenty trades. Meanwhile the deterministic engines built from the same methods raise our buy list from +4.66% to +5.87%. Both facts are real.

Smart Money Concepts is the most popular framework in retail trading right now. Dow Theory is the oldest. We tested both, twice, in two different forms — and got opposite answers depending on the form.

That contradiction is the most useful thing we have learned about technical frameworks.

Form one: rules extracted from books

Five hundred and eleven rules. Two ever accumulated twenty trades.

This is the same attrition we found across seven other frameworks, and the cause is the same. A rule dies before it reaches a backtest because it cites something no function can compute, or because it contains no mechanical trigger at all — roughly 30% of our entire 21,191-rule library is structurally untestable before performance is even in question.

Read that table on its own and the conclusion writes itself: SMC is fashionable noise, Dow is a century-old relic, neither survives contact with data.

That conclusion is wrong, and we can show it is wrong with our own money.

Form two: the same methods as running engines

Separately from the rule library, we built deterministic engines for both — pure functions over daily OHLCV, no LLM, no I/O, clean-room implementations of the public-domain methodology.

The SMC engine computes market structure (break of structure and change of character), order blocks as the last opposing candle before a displacement leg, fair value gaps as three-candle imbalances with mitigation tracking, premium and discount zones around the dealing range equilibrium, displacement as an ATR-multiple impulse, and liquidity pools at equal highs and lows with sweep detection.

The Dow engine computes trend structure from pivots (higher highs and higher lows versus lower highs and lower lows), separates primary from secondary trend using a large-versus-small ZigZag threshold on the same close series rather than calendar heuristics, and — this is the part that matters — reverses a confirmed primary trend only on two breaks: first an at-risk break of the prior secondary swing, then a confirmed break of the prior primary extreme.

That hysteresis is not decoration. It is what stops a single bar flipping the trend, and it is the difference between Dow's actual tenet and the caricature of it that a book rule usually encodes.

What the engines measured

Both run as a gate on our daily buy list. A candidate must show a current bullish read from at least one of them before it can be ranked. Measured on the same oversold population and the same 10-day forward horizon:

And the mechanism is specific. The gate earned its place by raising the win rate from 62% to 65%, not by finding more names. It is a falling-knife filter: its job is to reject the oversold reading that is a collapse in progress rather than a base.

That is a smaller claim than "SMC predicts price." It is also a claim that survived measurement, which the rules did not.

Why the same method gives two answers

The gap is not that one test was kinder. It is that the two forms are not the same method.

A book rule says "buy the retest of a bullish order block in discount." To mechanise it you must decide, in code, which leg counts as displacement, which candle was the last opposing one, where the dealing range starts and ends, and what counts as a retest. Each of those is a judgement the book leaves to the reader, and a rule that hard-codes one arbitrary answer to each is not the method — it is one analyst's habits, frozen.

The engine makes those same decisions, but it makes them once, explicitly, consistently, and on top of a shared pivot substrate — and then it is used as a filter over a population rather than as a standalone entry trigger.

The rules failed because mechanising a framework as an entry signal throws away

the judgement. The engines worked because they mechanise the same framework as

a context read and let something else make the entry decision.

What we would not claim

Not that SMC or Dow Theory "works." The gate improved a specific mean-reversion population on a specific horizon by 1.21 points of forward return and 3 points of win rate. That is a filter contribution measured in one context. It is not evidence that order blocks predict price generally, and we have not tested that.

Not that our attrition table condemns the frameworks. Two testable rules out of 511 tells you about our extraction and about what compiles, not about whether the methods have merit in a human's hands.

And not that the gate is proven. It sits inside a strategy whose unscreened form backtests at profit factor 0.90. A filter that improves a losing baseline has demonstrated it improves the baseline, not that the result is an edge.

The transferable version

If you are evaluating any framework — SMC, Dow, Wyckoff, or the next one — the question that separates the two answers above is:

Are you testing the method, or one frozen instance of the judgement the method asks for?

Almost every published backtest of a discretionary framework is testing the second thing and reporting it as the first. Ours did too, until we built the engines and got a different answer from the same books.