SMA vs EMA: 672 head-to-head backtests on USD/JPY

Every guide to moving averages eventually reaches the same fork. Simple or exponential? The stock answers are "the EMA reacts faster, so use it for short-term trading" and "the SMA is smoother, so use it for the long term." Neither is usually accompanied by a number.

So we ran the comparison directly: the same period pair, the same timeframe, the same year, with only the averaging method swapped. Fifty-six fast/slow combinations, three timeframes and four windows gives 672 head-to-heads on USD/JPY.

The EMA finished ahead in 318 of them — 47.32%. There were no ties. Split into twelve cells, the EMA's win count ranged from 15 to 39 out of 56, and no cell separates the two methods.

pairs where the EMA finished ahead28 of 56 — the coin-flip lineUSD/JPY, 56 period pairs in each of the 12 rows

That is not the same as saying the choice does not matter. Swapping the method at fixed periods moved net pips by a median of 491.1. The change is real; its direction is what cannot be predicted. That combination is the finding, and the rest of this article is the mechanism behind it.

Two properties turned out to belong to one method and not the other. One is the SMA's, one is the EMA's, and both are arithmetic rather than market behaviour.

What the two averages actually weight

Both are a weighted sum of past closes. The difference is how the weight is handed out.

• Simple (SMA) — add the last n closes and divide by n. Every one of those n bars carries the same weight, and nothing older is read at all.
• Exponential (EMA) — weight the newest bar most and every older bar by a constant fraction less. The weight shrinks but never reaches zero.

The weights can be recovered rather than assumed. Build a flat series, move a single bar up by one unit, and whatever comes out of the average is that bar's weight at each age. Running that through the same calcSMA and calcEMA the backtests use gives this:

SMA 20EMA 20centre of gravity — 9.5 bars back for bothvertical: weight on that bar (%). horizontal: how many bars back

("How far back the EMA reads" is the oldest bar still carrying more than a millionth of the newest bar's weight.)

The quiet row is the middle one. The centre of gravity is identical at every period. For an n-period average of either kind it lands on (n−1)/2 bars back: 9.5 for period 20, 99.5 for period 200, matching to the decimal in all seven periods tested.

So the EMA is not reading newer information on average. The average age of what it weights is the same. What differs is the spread: it doubles the weight on today and pays for that by keeping weight outside its own window. A 20-period EMA holds 86.49% of its weight inside 20 bars and scatters the other 13.51% across older ones, and that 86% figure barely moves with the period (86.47% to 86.83%).

This is why "the EMA is faster" and "the EMA secretly looks back further than its period" are both true. They describe the same weight profile from opposite ends.

Measuring "the EMA reacts faster"

If the centre of gravity is shared, where does the speed show up? Three measurements, three different answers.

Distance from the close. How far the line typically sits from price, hourly USD/JPY, 2025:

The EMA is closer to price in every one of the 42 cells measured (three timeframes × two years × seven periods), by a ratio between 0.8126 and 0.8937. Around 85%, near enough, whatever the period and timeframe. That is one real sense of "faster."

How often the two lines cross. Same periods, one method against the other:

(USD/JPY, 2025)

The supposedly faster method crosses less often. Across 36 cells the simple pair crossed more in 31, fewer in 3, and the same number in 2. The next section explains why.

Which crossover comes first. Matching same-direction crossovers closest-pair-first, so that neither method's extra crossings drag the answer:

(USD/JPY, 2025)

Over all 36 cells the EMA crossover came first between 41.36% and 76.47% of the time, above half in 26 of them. It genuinely leads on long pairs like 50/200. On a 10/20 pair, which method fires first is close to a coin toss.

"The EMA gives you the signal earlier" did not hold on short pairs. On 50/200, five of the six timeframe-and-year cells put the median lead at 3 to 7 bars; the sixth (four-hour, 2025, where the EMA pair crossed 8 times) came in half a bar late.

Only the simple average turns against price

The crossing-count gap has an exact cause. One bar of SMA movement is this, and it is an identity rather than an approximation:

SMA today − SMA yesterday = (close today − close n bars ago) / n

Checked against the shipped function on three timeframes at periods 10, 20, 50 and 200, the largest disagreement over those twelve cells was under 0.0000000000037 pips. Rounding, and nothing else.

Read what it says: whether the SMA turns up or down is decided by comparing today's close to the close n bars ago. That older bar is nowhere on the chart. The bar dropping out of the window has as much say in today's direction as the bar arriving.

The EMA's step is a different identity, verified on the same twelve cells to the same precision:

EMA today − EMA yesterday = k × (close today − EMA yesterday)      k = 2 / (period + 1)

The bracket is "is today's close above or below the line". So the EMA turns up whenever the close is above it and down whenever it is below, always.

That difference is countable. Bars where the line turned one way while the close sat on the other side:

(USD/JPY, 2025)

Zero in all 42 cells for the EMA, and it will be zero on any instrument and any year, because the definition forbids it. The SMA ran between 10.95% and 21.43% — roughly one bar in six.

This drop-off is what inflates the SMA's crossing count. Two simple averages of different lengths shed different old bars at different moments, so they change direction for reasons that have nothing to do with each other, and they meet more often.

Whether that helps or hurts is a separate question, and the backtests below are the answer to it.

One thing that does not follow: the EMA changes direction more often, not less. In 41 of 42 cells it turned more frequently — at period 5 on 15-minute bars, 26.82% of bars against the SMA's 21.2%. The EMA turns more; the EMA pair crosses less. Two different measurements, both true.

The exponential average never forgets its seed

The EMA has an instability the SMA does not. Today's value is built from yesterday's, so wherever the calculation was started leaks into every value that follows.

Two EMAs seeded at different points, measured bar by bar:

The ratio matches the closed form exactly, and it never reaches zero. The SMA is identical from the moment its window clears the join, which is its period, and stays identical after that.

In practical terms, here is what a chart draws on 1 January 2025 depending on how much history it loaded:

An hourly EMA(200) drawn from exactly 200 bars sits 49.083 pips away from the same EMA(200) with adequate history, and takes 620 more bars — over a month of hourly data — to converge. The SMA(200) agrees from its very first value.

Two things follow. Load three to five times the period in history before trusting a long EMA: at 1,000 hourly bars the error is down to 0.008 pips. And when two platforms disagree about where an EMA sits, neither is necessarily wrong — they may simply have started in different places.

This is not unique to the EMA. Anything defined from its own previous value behaves this way; Heikin Ashi candles halve their seed's influence every single bar.

Adding a moving average to your chart

Moving averages ship with every platform. What varies is the list of methods and the default.

The line is drawn over the candles, not in a pane below. The settings dialog holds four fields:

To turn two lines into a rule, the backtest conditions take a trigger and both legs. The fast and the slow leg take their method separately, which is measured later in this article.

The common rule is to buy when the fast line crosses above the slow one and sell when it crosses below — the golden cross and the dead cross. Every number here uses that rule.

Only two fields are worth tuning: the period and the type. The period is the decision that comes first, and it is covered in 280 moving-average crossover settings tested. Here the periods are held and only the type moves.

How this was tested

Fast and slow use the same method except in the mixing section. Holding to the opposite crossover with no stop keeps everything except the method out of the comparison; stops are added separately below.

672 head-to-head runs, and the coin-flip line

Same periods, same timeframe, same window, method swapped. This is the table the chart at the top draws.

(Net pips)

318 of 672 across all twelve cells — 47.32%. No pair produced identical net pips in the two methods.

The two hourly half-year rows are worth reading together. The first half of 2025 gives 39/56 to the EMA; the second half of the same year on the same timeframe gives 15/56 to it. Same instrument, same 56 pairs, opposite answers six months apart.

A different question — how many settings finished both years in profit — gives a different-looking result:

By this count the EMA leaves slightly more standing (52 against 48, 34 against 28, and 2 against 2). "More EMA settings survive both years" and "more EMA settings out-earn their SMA twin" are not the same statement, and only the first one is true here.

Does last year's winner win this year?

For each period pair, take the method that won in 2024 and check whether it won again in 2025.

33.93% on 15-minute bars is below a coin flip. Neither of the other two reaches 50% either, and cutting the year in half gives the same 25-to-30 range.

The pairs people actually type

(Net pips. Seven of the eighteen combinations)

Over all eighteen (six pairs × three timeframes) the EMA was ahead in 11 in 2025, 9 in 2024, and in both years in 3.

The 15-minute 10/20 is the clearest case. In 2025 the EMA beat the SMA by more than 1,000 pips; in 2024 the SMA beat the EMA by more than 400. Same pair, same instrument, opposite sign one year apart.

Is the method a bigger knob than the period?

"The EMA wins 47.32% of the time" is also what you would see if the method changed nothing at all. Those two situations need separating, and separating them takes a control.

So: two ladders of consecutive integers. Fast 5 through 30 with the slow leg pinned at 50, and slow 20 through 60 with the fast leg pinned at 10. At every rung, compare how far swapping the method moved net pips against how far moving the period by a single bar moved it.

one ladder rung, 780 of themthe diagonal — both changes equalabove the line, the method mattered more (593 rungs)

(The six 2025 rows; the totals below cover all 24 cells)

Across 780 rungs the swap moved the result further than the one-bar step in 593 — 76.03%. The medians are 350.7 pips for the swap against 154.2 for the step.

Put in the period's own units: how many bars would the period have to move to shift the result as much as the method did? The median answer is two. One bar sufficed at 268 of 804 rungs, five or more were needed at 208, and at 70 rungs twenty bars still were not enough.

So: swapping SMA for EMA is worth about two bars of period. That is not nothing. Which way it moves you is unknowable in advance.

This is the shape the Fibonacci retracement test needed for a different reason. When there are only two candidates, counting how often one wins proves nothing on its own. A control beside it is what makes the count mean something.

Mixing the two methods across the two lines

The fast and slow legs can take different methods, so six pairs × four combinations × two years = 36 cells.

(Net pips. Six of the 36 cells)

All four combinations landed on the same side of zero in 21 of 36 cells, and the spread between the best and worst of the four ran to a median of 1,399.7 pips, widest 4,517.7. Mixing moves the result substantially.

What it does not do is swap the roles of the two lines. The moving-average crossover article found that a 20-period Hull average sits closer to price than a 10-period simple one, so giving the slow leg a Hull average makes the "slow" line the fast one. An EMA only closes about 15% of the distance, which is not enough:

(USD/JPY, 2025)

In all eighteen combinations the slow EMA stayed further from price than the fast SMA. The inversion needs the periods within 1.3× to 1.45× of each other (holding the fast leg at SMA 20 and lengthening the slow one bar at a time, the boundary fell at 1.3× on M15 and H1, and 1.45× on H4). It is worth checking only for pairs as close together as 20 and 25.

Win rate points one way and net pips the other

(USD/JPY, 2025)

Over all 672 head-to-heads the EMA took fewer trades in 405 and had a lower win rate in 561. Among the eighteen named combinations its win rate was lower in all eighteen.

Net pips still finish level because the EMA's average trade is bigger. On the 15-minute 10/20 the SMA made +1,015.0 pips over 1,402 trades (+0.72 each) and the EMA +2,052.3 over 1,053 (+1.95 each).

Rank the two methods by win rate and the SMA wins. Rank them by result and it is a tie. The same inversion runs through this series: Williams %R wins 55–79% of its trades and loses money.

Here is the whole comparison in one place.

Filters, stops and targets

(Ten of the 24 filter runs)

11 of 24 improved pips per trade. The ADX filter improved 2 of its 8 and turned all four 2025 runs from profit into loss. The Tokyo-session filter improved 4 of 8.

Filters did not behave differently by method. Where ADX hurt, it hurt both, and in the same direction.

(Eight of the 32 exit runs)

6 of 32 beat holding to the opposite crossover — three on the SMA side and three on the EMA side. No difference by method here either.

The spread, and where each method breaks even

Moving the spread from 0 to 2.0 pips costs the 15-minute SMA run 2,804.0 pips, which is exactly 1,402 × 2.0. The EMA loses 2,106.0 (1,053 × 2.0), the hourly SMA 547.9 against 274 × 2.0, and the hourly EMA 408.0 (204 × 2.0). Cost equals trade count times spread holds through the method swap, as it has in every article in this series.

The 15-minute 10/20 breaks even at 1.02 pips as an SMA and 2.25 as an EMA. A third fewer trades buys more than twice the headroom.

This is not an edge in expected value — that part is a tie. It is tolerance: when the spread widens, or trades cluster into expensive hours, the simple version goes underwater first. The EMA took fewer trades in 405 of 672 head-to-heads, so the direction generalises even if the size does not.

What this test can say

1. Result does not separate them. 318 of 672, or 47.32%. Every one of the twelve cells lands between 15 and 39 out of 56
2. The centre of gravity is shared. Both put it at (n−1)/2 bars back, matching at all seven periods tested. "The EMA sees newer data" is not what is going on. It doubles the weight on today and leaves 13.5% of its weight outside its own period to pay for that
3. Only the SMA turns against price. On 10.95% to 21.43% of bars the close sits above the line while the line points down. Zero occurrences for the EMA in all 42 cells, and that is why the simple pair crosses more often (31 of 36 cells)
4. Only the EMA depends on its history. An hourly EMA(200) drawn from 200 bars is 49.083 pips off and needs 620 bars to converge. Load five times the period before trusting one
5. Last year's winner is not a signal. The same method won both years in 33.93% of pairs on M15, 46.43% on H1 and 44.64% on H4

What survives as a practical difference is trade count, and cost tolerance with it. The EMA trades less in 405 of 672 head-to-heads and breaks even at a wider spread. Since the result cannot decide it, the spread you actually pay and how often you want to trade are the only grounds this test supports.

The period is the decision that comes first, so the 280-setting crossover test is worth reading before this one. If you want to run the same comparison on your own pair and dates, it can be done without writing code.

Limits of this test

• One instrument (USD/JPY) and two years. Nothing here promises the same on another pair or another year
• The two years differ in character. 2024 rose 1,632 pips with a 2,237-pip range; 2025 finished 56 pips lower than it started with a 1,900-pip range. Crossovers are a trend-following device, so that difference produces much of what is above
• Only simple and exponential were compared. Weighted, smoothed and Hull are in the crossover article
• The weights, the centre of gravity and the two step formulas follow from arithmetic and hold on any data. The distances, crossing counts and results are USD/JPY over two years
• Entries and exits are taken at the bar close. Real fills differ
• The spread is held at 0.3 pips throughout. Real spreads move with the session and the calendar
• The seeding measurement is read at one moment, 1 January 2025. Another date gives a different size of error, though the decay rate is date-independent
• Volatility varies by hour, which the session-filter rows inherit