MACD: Read the Spreads Before You Read the Signals

MACD is usually introduced as a bundle of crossovers: buy here, sell there, watch the bars change color. That shortcut hides the most useful idea.

MACD is a measurement system built from two spreads:

1. the distance between a fast and a slow exponential moving average;
2. the distance between that first spread and its own smoothed history.

Once those two distances are clear, the zero line, signal line, and histogram stop being separate pieces of chart folklore. They become simple algebra that you can calculate, test, and reason about.

This tutorial builds that understanding from the inside out. We will calculate a complete example by hand, connect every crossover to an equality, expose the two-stage warm-up, test scale behavior, and finish with production-safe implementations.

Using the tables on a narrow screen: wide tables scroll horizontally inside their own table frame. Swipe within the table to inspect later columns; the article page itself should remain fixed.

The three outputs answer three different questions

Let $F_t$ be the fast EMA and $S_t$ the slow EMA.

The MACD line $M_t$ asks:

Is the shorter-horizon average above or below the longer-horizon average, and by how much?

Now smooth the MACD line with another EMA:

The signal line $G_t$ asks:

What is the recent smoothed level of the fast-minus-slow spread?

Finally:

The histogram $H_t$ asks:

Is the current spread above or below its own smoothed history, and by how much?

The word “momentum” is often attached to these outputs. That can be a useful interpretation, but the equations give a narrower and safer statement: MACD measures relationships among recursively smoothed levels. It does not directly measure return, velocity, acceleration, probability, or expected profit.

First visual key: the sign is just the EMA gap

This identity is exact:

So “MACD crossed zero” and “the fast EMA crossed the slow EMA” are not two confirmations. They are two drawings of the same event.

That distinction matters. Counting the same equality twice can create false confidence. A dashboard that shows both a moving-average crossover badge and a MACD-zero-cross badge may be presenting duplicate information, not independent evidence.

The EMA convention must be named

“Use a 12-period EMA” is not yet a complete computational specification. An EMA needs both a recurrence and an initial value.

For span $n$, this package uses:

The first defined EMA is the simple mean of the first $n$ observations:

Every later value is:

This is the classic SMA-seeded recursive convention. The fast and slow EMAs are seeded independently from the source. The signal EMA is then seeded from the first signal_span defined MACD values.

When two charting packages disagree near the beginning of a series, compare these choices before assuming that either one has an arithmetic bug.

Fix the timestamp and price basis before calculating

The numeric kernel receives only an ordered array, so it cannot discover whether two adjacent values are daily closes, five-minute bar closes, mixed sessions, or differently adjusted prices. The caller must attach a unique timestamp to every value and document:

• time zone and whether a bar timestamp denotes its opening or closing instant;
• frequency, trading calendar, and the treatment of missing sessions or bars;
• instrument identifier, currency, and source field;
• adjusted or unadjusted price basis and the upstream corporate-action policy;
• finalized versus provisional status.

Use one convention for the entire series. Splicing adjusted history into an unadjusted suffix can create a discontinuity that both EMAs faithfully process even though the discontinuity came from data construction rather than the market path. Likewise, a provisional latest bar can change before finalization and recursively change its MACD row.

Two warm-ups, not one

The slow EMA controls when MACD first exists. Then the signal needs its own history of MACD values.

Counting observations from one:

Before observation 26 in the default case, there is no canonical slow EMA under this seed convention, so there is no MACD. At observations 26 through 33, MACD exists but the nine-period signal is still accumulating its seed. Observation 34 is the first row with all three plotted outputs.

Returning those stages explicitly is safer than dropping rows. An aligned result can attach each component to the exact source timestamp and label its state:

Unavailable is represented by None or null. It is never represented by zero, because zero is a meaningful equality.

A complete 3/5/3 worked example

Short spans make every step visible. Use:

• fast span $3$, so $\alpha_f=2/(3+1)=1/2$;
• slow span $5$, so $\alpha_s=2/(5+1)=1/3$;
• signal span $3$, so $\alpha_g=1/2$.

The synthetic source is:

10, 11, 12, 13, 14, 15, 16, 15, 14, 13, 12, 13, 15, 18

The path rises, reverses, and recovers. It is deliberately educational rather than representative of a particular instrument.

Step 1: seed the price EMAs

The first fast EMA occurs at observation 3:

The first slow EMA occurs at observation 5:

The fast EMA has already updated twice:

Therefore the first MACD is:

The signal is still unavailable. It needs three MACD values.

Step 2: seed the signal

At observations 5, 6, and 7, the MACD values are $1,1,1$. Therefore:

Observation 7 is the first fully ready row.

Step 3: see why MACD and histogram can disagree in sign

At observation 8:

MACD remains positive because the fast EMA remains above the slow EMA.

But the signal updates from $1$:

So:

The two signs now say different, perfectly compatible things:

A negative histogram does not require a negative MACD. The first spread can be positive while the residual spread is negative.

Step 4: cross the zero line

At observation 10:

The negative MACD means the fast EMA is now below the slow EMA. This is not a nearby confirmation of that crossover—it is the same subtraction reaching a negative result.

Step 5: recover

At observation 13:

Both spreads are positive again in this path. That simultaneity is not a general rule. A MACD zero cross and a signal cross answer different equalities and can occur at different observations.

Full audit table

Rounded values are shown for readability; the shared fixture stores the verified path.

Mobile table cue: this eight-column audit table is intentionally preserved. Swipe left or right inside its table frame to follow the same observation across all components.

Two crossover identities worth memorizing

MACD zero cross

The direction of the cross is the direction of the ordering change. A cross above zero means the fast EMA moved from at-or-below the slow EMA to above it, subject to the application’s equality and numeric-tolerance policy.

Histogram zero cross

A signal-line crossover and a histogram zero cross are likewise the same event drawn as lines and bars.

For event detectors, equality needs a policy. Exact IEEE-754 equality is often too brittle for independently computed series. A separate crossover component should define:

• whether touching zero counts;
• absolute or relative tolerance;
• how repeated equal values are handled;
• whether an event is emitted intrabar or only after finalization;
• whether the first defined value can create an event.

Those are event semantics, not part of continuous MACD calculation.

Four tiny experiments that reveal the structure

1. Constant input

If every source value is $42$, each seeded and updated price EMA is $42$. Therefore:

This is a useful implementation test because any nonzero ready output reveals an alignment or seed defect.

2. Add a constant

Add $100$ to every source observation. Both price EMAs also increase by $100$, so their difference is unchanged:

The signal and histogram are unchanged too. MACD is translation invariant.

3. Multiply the series

Multiply every source value by $10$. Every linear EMA component and difference is multiplied by $10$:

MACD is scale equivariant, not scale invariant.

This is why “MACD is 5” has no portable meaning without the instrument’s price scale and history. The next catalog topic, Percentage Price Oscillator, addresses that normalization problem.

4. Make the signal span one

With a signal span of one, the signal EMA equals the current MACD immediately:

Therefore:

This edge case is valid and highly diagnostic. It confirms that the histogram is derived, not independent.

Why range-bound markets produce whipsaw

A fast EMA reacts more quickly than a slow EMA. If the source oscillates back and forth without developing a durable directional path, the fast EMA can repeatedly move above and below the slow EMA. MACD crosses zero each time. The MACD line can also repeatedly cross its signal.

Every event can be arithmetically correct while a naïve crossover strategy performs poorly after costs.

The practical response is not to claim that MACD “fails.” It is to keep the claim boundary honest: MACD describes selected averages. Strategy quality depends on regime selection, event rules, execution timing, costs, and risk controls outside this indicator.

Use the causal lab

Open the interactive MACD causal lab.

Try this sequence:

1. Keep spans at 3/5/3 and step through the reversal path.
2. Stop at observation 5. Notice that MACD exists while signal and histogram do not.
3. Stop at observation 8. Compare positive MACD with negative histogram.
4. Switch to Range and whipsaw and play the complete path.
5. Switch to Same shape × 10 and compare magnitudes.
6. Set signal span to 1; every ready histogram bar should be zero.
7. Attempt fast 5, slow 5; the lab rejects the ambiguous configuration.

The lab calculates only the visible prefix. Nothing at observation $t$ uses observation $t+1$, so it also demonstrates causality.

Every lab scenario is synthetic. It demonstrates calculation, warm-up, scale, and whipsaw mechanics; it is not evidence that a crossover predicts a later observation or earns a return.

Implementation blueprint

The core algorithm is small:

fast = SMA-seeded EMA(source, fast_span)
slow = SMA-seeded EMA(source, slow_span)
macd = fast - slow wherever both exist
signal = SMA-seeded EMA(defined macd values, signal_span)
histogram = macd - signal wherever signal exists
return all rows in source alignment

What makes an implementation production-safe is the surrounding contract:

The Python implementation returns immutable dataclass rows. The TypeScript implementation returns typed component objects. Both read the same fixture, test algebraic identities, and test translation and scaling behavior.

Streaming state

After warm-up, a streaming system only needs:

• previous fast EMA;
• previous slow EMA;
• previous signal EMA;
• three coefficients;
• current lifecycle status.

One appended finalized observation then takes constant time:

fast   ← fast   + α_fast   × (value − fast)
slow   ← slow   + α_slow   × (value − slow)
macd   ← fast − slow
signal ← signal + α_signal × (macd − signal)
hist   ← macd − signal

During seeding, keep running sums and counts for the price EMAs and later the signal EMA. A revised historical observation is different: because the recurrence carries state forward, every later dependent value must be recomputed.

Common mistakes

What MACD can and cannot tell you

MACD can tell you, under a named source, span, seed, and finality convention:

• the ordering and distance of two EMAs;
• the position of that spread relative to its EMA;
• when either exact equality changes sign;
• how those measurements evolved over the observed history.

MACD alone cannot tell you:

• the probability of the next return;
• whether a crossover will persist;
• whether expected profit exceeds transaction costs;
• the correct position size;
• whether a divergence rule has valid pivots;
• whether the current observation is final;
• whether raw magnitude is comparable to another instrument.

That boundary does not make the indicator less valuable. It makes it testable.

Final mental model

When you see a MACD panel, read it in this order:

1. MACD sign: Which price EMA is on top?
2. MACD magnitude: How large is their gap in source units?
3. Histogram sign: Is that gap above or below its own signal EMA?
4. Histogram magnitude: How large is the residual?
5. Lifecycle: Are all components actually warmed and finalized?
6. Context: Is the source trending, ranging, scaled differently, or being revised?

Then—and only then—attach a strategy interpretation.

MACD is important not because three plotted shapes predict the future. It is important because a compact chain of recursive averages exposes trend relationships in a form that is visual, causal, and mechanically auditable.

Continue the track

Next is D07-F02-A02 — Percentage Price Oscillator. It keeps the fast-versus-slow idea but divides by the slow average, turning an absolute price-unit spread into a relative percentage spread.

For formulas, implementation-source analysis, and claim provenance, see the reference ledger.