Moving Average Calculator

The lag is exact, not approximate

“Moving averages lag” is said constantly and quantified almost never. It is quantifiable, and the answer is clean.

A trailing n-period simple average lags a linear trend by exactly (n − 1)/2 periods. The mean of the last n values of a straight line is the line’s value at the middle of that window, which is (n − 1)/2 periods ago.

So a 50-day average is 24.5 days behind, and a 200-day average is 99.5. Those are not rules of thumb. The verification suite measures the lag on a linear series at seven window lengths and finds agreement to 10⁻⁹.

A linearly weighted average lags by (n − 1)/3, two thirds as much, because weighting recent points more heavily pulls the effective centre of the window forward.

A centred average lags by exactly zero, and cannot be computed for the last (n − 1)/2 points. That is the trade, and it is the reason trailing averages exist despite lagging.

On real data the measured lag differs from the formula, because a real series is not a line. The tool prints both: the formula is what the average does on a trend, which is the figure worth quoting, and the measured one shows what it did on your last few points.

Why α = 2/(n + 1)

Every EMA formula contains it and almost no source explains it. The explanation is the lag from the previous section.

An exponential average lags by (1 − α)/α periods. That is the mean age of the weights — each observation gets weight α(1 − α)ᵢ, and the weighted mean of the ages is (1 − α)/α.

Setting that equal to (n − 1)/2 gives α = 2/(n + 1). One line of algebra. The convention exists so that a “10-period EMA” and a “10-period SMA” are comparable — same lag, different weighting.

The tool prints both lags side by side, and they match: 4.5 for a 10-period window either way. The verification suite checks it on a 500-point linear series across four windows.

Equal lag does not mean equal behaviour. An EMA reacts faster to a sudden jump, because it puts the most weight on the newest point rather than equal weight on all n. The equal lag is a statement about a steady trend, not about a shock.

And an EMA never fully forgets. Every past observation retains a small weight forever, where a simple average drops each point cleanly after n periods. That makes the EMA smoother and means an old outlier leaves a trace long after it has fallen out of any finite window.

A crossover cannot precede the turn

The most common use of two moving averages is the crossover, and the lag makes its timing a matter of arithmetic rather than of skill.

Both averages report the past, so their crossing reports the past. A fast average crosses a slow one only after the underlying series has already turned, by roughly the difference in their lags.

For a 50/200 pair that difference is 75 periods. (200 − 1)/2 minus (50 − 1)/2. A golden cross is, structurally, a report that something changed about seventy-five days ago.

The tool’s turning-point preset makes it concrete. The series turns at period 11; the crossover arrives several periods later, and the suite asserts it can never arrive before the turn.

On a series with no trend, every crossover is noise. The noisy-and-flat preset produces several, all of them false. Shorter windows produce more crossings and more false ones; longer windows produce fewer and later.

That is the trade, and it has no optimum. Any window choice is a position on the responsiveness-versus-noise curve, and it cannot be tuned from the same data it will be evaluated on without overfitting. A crossover rule tested on the history it was chosen from will always look good.

Centred averages, and deseasonalising

If you are describing what has already happened rather than reacting to the present, the trailing average is simply the wrong tool.

A centred average has no lag at all. Its window straddles each point, so the smoothed value sits where the data sits, and there is no correction to apply.

The cost is the end of the series. The last (n − 1)/2 points have no value, because the window would need data that does not exist yet. For historical analysis that is not a cost.

An even window needs the 2×n convention. A 12-period window has no middle point, so the standard fix averages two adjacent windows — half weight on each end of a 13-wide span. That is what a 12-month centred average of monthly data actually is, and it is what the tool computes.

Which is the first step in classical decomposition. A centred average over exactly one seasonal cycle removes the seasonality, because every season appears once in the window. Dividing the series by that average leaves the seasonal factors, which is how a seasonal index is built.

The window must match the cycle exactly. Twelve for monthly, four for quarterly, seven for daily data with a weekly pattern. A 10-month average of monthly data does not remove seasonality; it smears it.

Choosing a window, and what the choice costs

There is no correct window length, and the ways of choosing one are all statements about what you are willing to give up.

Longer windows smooth more and lag more, in a fixed relationship. Noise falls with √n and the lag rises with n, so doubling the window halves the noise by a factor of 1.41 and doubles the lag exactly.

Match the window to a real cycle where one exists. Seven days for weekly seasonality, twelve months for annual. A window chosen for a reason is defensible in a way that 50 and 200 are not.

The round numbers in finance are historical. 50 and 200 trading days are roughly ten weeks and ten months, and they persist because everyone uses them rather than because anything selected them. Their popularity is self-reinforcing and is itself a reason they sometimes work.

Optimising a window on past data overfits, always. With enough candidate windows one will have worked beautifully on any history, and that is the same problem the multiple comparisons page covers. Testing on data the choice did not see is the only check worth anything.

And a moving average is not a forecast. It estimates the current level from recent history. Projecting it forward assumes the level is stable, which is exactly the assumption the series violates whenever it matters — and a fitted trend is the honest tool when you need to extrapolate.

Where a moving average misleads

Smoothing changes the data, and three of the changes are routinely mistaken for findings.

It manufactures autocorrelation. Adjacent values of a moving average share n − 1 of their inputs, so they are correlated by construction. Running a test that assumes independent observations on smoothed data inflates significance badly, and the smoothing is the cause.

It can create cycles that are not there. The Slutsky-Yule effect: averaging pure noise produces a series with apparent periodicity, and the period depends on the window rather than on the data. A cycle visible only after smoothing is a candidate for this.

It hides the variability you may have wanted. A smoothed series understates the range of the underlying values, so a capability or risk figure computed from averages rather than from individual readings is systematically optimistic — which is the mistake the capability page is about.

The endpoints are the least reliable part. The most recent value of a trailing average is built from the fewest settled observations and will be revised as more arrive. Charts rarely say so, and decisions are usually made at exactly that point.

And a change in the window changes the picture. As with histogram bins, a feature that disappears when the window changes was a property of the smoothing rather than of the series. Trying two or three windows costs nothing and settles it.

Sources and methodology

References for the methods and their known artefacts.

• Hyndman & Athanasopoulos, Forecasting: Principles and Practice — moving averages and classical decomposition
• Slutsky, “The Summation of Random Causes as the Source of Cyclic Processes”, Econometrica 5(2), 1937
• NIST/SEMATECH e-Handbook — moving average and exponential smoothing
• Frequency Distribution Calculator — the same “the parameter decides the picture” problem, for histograms
• Exponential Regression Calculator — a fitted trend, for when you need to extrapolate
• How we calculate — methodology

Method. Each average is computed directly rather than through a shared abstraction, so the lag each one carries can be stated as its own property rather than inferred. The lag figures are checked against the series itself: on a linear trend the implied lag — (value − average)/slope — must equal (n − 1)/2 for a simple average, (1 − α)/α for an exponential one, (n − 1)/3 for a weighted one and zero for a centred one, and the suite verifies all four to 10⁻⁹ across seven window lengths. It separately confirms that α = 2/(n + 1) makes the exponential lag equal the simple one, that every method reproduces a constant series exactly, that shifting the series shifts the average by the same amount, and that a crossover never precedes the turn that caused it. That engine is verified on every change against 54 assertions. The count and the per-case breakdown are published on the formula verification page.

Related calculators

Where this goes next:

More in Math, or browse all calculators.