MACD: A Lagging Filter Dressed as a Signal

If RSI is folklore with a formula, MACD is something subtler: a genuinely respectable piece of signal-processing machinery, applied to a problem it cannot solve. That makes it the more instructive defendant, because its failure is not in the mathematics — it is in the claim bolted onto the mathematics.

What MACD actually computes

Moving Average Convergence Divergence is built from exponential moving averages of price. You met the EMA in Course 102: a weighted mean whose weights decay geometrically into the past. MACD subtracts a slow one from a fast one, then smooths the result once more into a "signal line":

$\text{MACD}_t = \text{EMA}_{12}(P)_t - \text{EMA}_{26}(P)_t, \qquad \text{signal}_t = \text{EMA}_9(\text{MACD})_t$

To an engineer, this object has a precise name: a band-pass filter. Subtracting a slow smooth from a fast smooth cancels the long-run level and the highest-frequency noise, leaving the medium-frequency wiggles. Filters like this are legitimate tools — they describe what has happened at a chosen frequency band, and they do it with an unavoidable cost: lag. An EMA's output reflects the past by construction; a difference of two EMAs, smoothed a third time, reflects it three times over. Every MACD crossover is an announcement that something already occurred.

So the honest description of MACD is: a thrice-smoothed, lagging summary of recent price movement. The folklore's claim is that its crossings predict what comes next — that when the MACD line crosses its signal line, momentum is "confirmed" and will continue. Once again, the claim is a conditional probability dressed in chart language, and once again it is testable.

The rise and fall of the evidence

Here the story has a genuinely interesting twist, and intellectual honesty requires telling it. In 1992, Brock, Lakonishok and LeBaron published a careful study finding that simple moving-average crossover rules did show statistically unusual returns on ninety years of Dow data. It is the best evidence technical trading ever had, and it is the paper every defender cites.

It did not survive. Sullivan, Timmermann and White re-examined the same rules inside a universe of nearly eight thousand variants, applying White's Reality Check for data snooping — and showed that the 1992 results lose their significance once you account for how many rules the research community had implicitly searched over the decades. Worse, the rules' apparent edge fails out of sample: extend the data past 1986 and the performance degrades toward noise. Bessembinder and Chan, sympathetic readers of the original evidence, added the final blow — once realistic transaction costs are charged on the frequent trades the rules generate, the profits are gone even before the snooping correction. Survived multiple testing: no. Survived out of sample: no. Survived costs: no. That is not a mixed record. That is a demolition, delivered in the field's own top journals.

The general lesson is the one this institute keeps returning to. A backtest is a sample statistic and carries sampling error like any other; selecting the best-looking rule from a large searched space and reporting it as if it were a single pre-registered hypothesis is the multiple-testing sin from Course 201, performed at industrial scale.

The five-gate verdict — and the honest alternative

Scale: MACD operates on price levels, not returns, so its magnitude is meaningless across assets. Shape and spread: it carries no model of the distribution and no awareness of volatility clustering — crossovers fire constantly in turbulence and rarely in calm, for reasons that have nothing to do with opportunity. Base rate: its conditional hit rate, honestly tested, does not survive. Uncertainty: it reports no error bars. Zero gates of five.

And the honest alternative is precisely the machinery of Course 202. If you believe price movements have exploitable serial structure, that belief has a proper home: an autoregressive or ARMA model, with coefficients estimated from data rather than fixed at twelve, twenty-six and nine by convention; with standard errors that tell you whether the structure is distinguishable from noise; with residual diagnostics that tell you when the model is wrong. A fitted, tested AR model and a MACD crossover are reaching for the same intuition — yesterday's movement says something about tomorrow's — but only one of them measures whether the intuition is actually true, and reports its own uncertainty when it answers.