What Is a Moving Average? Why Being Late Is the Whole Point

There is a smooth line curling through the candles on almost every price chart you will ever be shown — in a broker's app, behind a news presenter, over a stranger's shoulder on a train. Almost everybody has one switched on. Stop ten of them and ask what it is actually doing, and you will get ten answers, of which perhaps four will be confident and rather fewer will be right. Which is a shame, because it is one of the very few things in this field that is exactly as simple as it looks.

It is an average. That is the entire trick.

Take the last ten closing prices. Add them up, divide by ten, and put a dot at that height above today's bar. Tomorrow, drop the oldest of the ten, add the new close, do the sum again, put down another dot. Join the dots and you have drawn a 10-day moving average. The window slides forward one day at a time — that is the moving part — and the average is just the average, which is the rest of it.

Nothing is being predicted, modelled or inferred. There is no signal processing behind the curtain and no theory of markets buried underneath. It is the sort of sum you could do on paper with the closing prices out of a newspaper, which for a long time is exactly how it was done. What you get out is not a forecast but a summary: one number standing in for the last N days, so that a chart's worth of dithering collapses into a single readable direction.

Which leaves exactly one decision, and it is the one doing all the work. How many days?

The lag is the point

Here is the thing everybody says about moving averages, usually with a small sneer: they are lagging indicators. Quite true. It is also a job description, not an accusation. An average of the last hundred days cannot possibly know anything about day one hundred and one — it is made entirely of things that have already finished happening. A line that responded to today's price the instant today's price moved would not be a moving average. It would be price, in a different colour.

And you cannot pull the two properties apart, because they are one property seen from two sides. Smoothing is lag. Every bar you add to the lookback makes the line calmer and makes it later, by the same amount, for the same reason: the newest close is now one voice in a bigger crowd. There is a single dial on this instrument and both labels are printed on it.

So when you type a number into that box you are not adjusting a setting. You are answering a question about yourself: how big does a move have to be before you are willing to call it real? A 10-day average believes essentially everything — one energetic Tuesday and it turns. A 200-day believes almost nothing; it takes a genuine change in the weather to bend it, and by the time it has bent, the weather changed a while ago. Neither is correct. They are answers to different questions, held by people with different amounts of time.

A moving average is the friend who tells you about the party on Monday. Thoroughly well informed, completely reliable, no use to you at all on Saturday night.

The dial below is that whole argument in one control. Drag it and watch two things move together: the line calms down, and the crossings it makes with the 200-day get fewer and later. You do not get to have one without the other, and there is no setting at which it quietly stops describing the past and starts predicting the future.

The read-outs there are the same figures as the chart above, computed the same way. Across that window the 10-day line travels $149. The 50-day travels $63. The 200-day travels $13, on identical closes — the short line covers roughly twelve times the ground the long one does, and every extra step it takes is a chance for it to have been about nothing. Turn on the 200-day and the other half of the trade-off shows up: down at the restless end the crossings come thick and most of them are cancelled within weeks, and as you climb they thin out and arrive later — right up until the last stretch, where you have simply dialled the fast line onto the slow one and the two tangle for want of any daylight between them.

SMA vs EMA: a smaller argument than it looks

There are two common recipes for the same idea, and they differ rather less than the arguing about them would suggest. A simple moving average, or SMA, is the one described above: every close in the window counts for exactly as much as every other. An exponential moving average, or EMA, leans on recent closes more heavily, with each older bar's influence decaying smoothly as it recedes. Nothing ever drops out of an EMA's window entirely — old prices just fade out.

So the EMA is slightly quicker and slightly twitchier, which — you will have spotted the shape of this by now — is the lookback trade-off again, in miniature. It is a far smaller decision than choosing N. If you have no strong feeling either way, the widely quoted 50-day and 200-day are conventionally simple ones, so picking simple keeps you looking at the same line as everybody else.

The golden cross and the death cross

Two lookbacks get more attention than all the others put together: the 50-day and the 200-day. Draw both on the same chart and sooner or later the 50 will cross through the 200. Upwards, that is a golden cross. Downwards, a death cross. Each gets a name, a colour and its own segment on the financial news — which tends to air at roughly the moment the move is half over.

Because look at what has actually occurred. The average of the last fifty closes has risen above the average of the last two hundred. That is the whole event. Translated out of chart language: the recent past has been better than the longer past. It is genuine information — it takes a real, sustained move to drag one average through the other, which is why the signal is not noise — but it is information about the past, delivered in arrears, and by construction it cannot possibly arrive until the move that caused it has already happened.

How far in arrears? Our chart is synthetic, so treat this as an illustration of the arithmetic rather than evidence about any real market — but the arithmetic is not negotiable. The golden cross on it fired 52 bars after the actual low, by which point price was already 13.1% off the bottom. The death cross fired 63 bars after the actual high, 23.7% below it. Both calls were right. Neither was early. A cross is not a forecast; it is a receipt.

The whipsaw, or when the cross keeps changing its mind

Trends are where a moving average earns its keep, and markets spend long stretches not trending at all. When price is going sideways, the two averages settle onto roughly the same value and stay tangled there — and a rule that fires whenever they cross begins firing constantly, in both directions, about nothing.

In the sideways stretch of our chart the 50 and the 200 cross five times, and three of those crossings are cancelled by the opposite crossing within twenty bars. That is a whipsaw, and it is the exact mirror of the lag problem: instead of turning up too late to a real move, the signal turns up promptly and repeatedly for a move that never existed. Anyone trading it faithfully would have bought the top of a range, sold the bottom, and then done it again, paying costs each time for the privilege.

This is why "the crossover works" and "the crossover doesn't work" are both perfectly defensible positions, held by people describing two different kinds of market. In a trend the rule is late but right. In a range it is prompt and wrong, over and over. And nothing tells you in advance which one you are standing in — the same structural inconvenience, once again, sitting underneath everything else on this page.

Does any of this actually work?

This is one of the few corners of technical analysis with a proper academic paper trail behind it, and the story that trail tells is considerably more interesting than either side of the usual argument.

The famous starting point is a 1992 paper in the Journal of Finance by Brock, Lakonishok and LeBaron. They tested two families of simple rules — moving-average crossovers and trading-range breaks — on the Dow Jones from 1897 to 1986, using bootstrap methods to check the results against four different models of how prices are supposed to behave. Their findings, in their own words, provide strong support for the technical strategies: buy signals consistently generated higher returns than sell signals, and returns following sell signals were negative. Two qualifiers travel with that headline and usually get left behind. It is a result about the Dow between 1897 and 1986. And it is before transaction costs — the authors' own closing caveat is that transaction costs should be carefully considered before such strategies can be implemented, which matters rather a lot when the rules trade as often as these do.

Then comes the audit, and this is the part almost everyone repeats backwards. In 1999 Sullivan, Timmermann and White went after the obvious objection: test enough rules on one dataset and some of them will look brilliant by sheer luck. So they expanded the original 26 rules to 7,846 and ran the lot over a full century of daily Dow data, 1897 to 1996, applying White's Reality Check bootstrap to price in the data-snooping. The 1992 result survived. It appears, they write, to be robust to data-snooping, and the finding is valid in all four of the original subperiods.

What did not survive was time. The best rule as of the end of 1986, carried forward into the following decade, simply stopped working: over 1987 to 1996 it produced a mean return of 2.8% with a nominal p-value of 0.322, which is statistics for "that could comfortably be nothing." The authors offer several readings of the failure, among them the plain one — the edge was real, then markets became cheaper and more liquid to trade in, and it was competed away.

So the crossover is not a myth that dissolved under inspection. It is something that worked, passed the inspection, and then expired. Which is arguably the worse outcome, if you were hoping to use it.

The third paper asks a different question, and it is the one that has held up. Mebane Faber's A Quantitative Approach to Tactical Asset Allocation, published in 2007 and updated in 2013, does not use an average as a buy-and-sell trigger at all. It uses it as a regime switch, checked once a month: hold the market while price is above its 10-month simple moving average, sit in cash while it is below. That works out at fewer than one round-trip trade a year, and the results exclude taxes, commissions and slippage.

In the 2013 update, covering 1901 to 2012, the average annual return was 11.26% for buy-and-hold against 11.22% for the timing model — a wash, and marginally the wrong way. Compounded, the timing model came out ahead, 10.18% against 9.32%, and Faber puts that gap down to lower volatility rather than better calls: high volatility drags compound returns down, and sitting out the worst stretches leaves less of it to do the dragging. The drawdown figures are where it stops being subtle. The model improved the drawdown in all but two decades, and through the 1929–32 collapse it would have turned a drawdown of 83.66% into one of 42.24%.

Every number on this page, ours included, comes out of a backtest, and the SEC's own investor bulletin is blunt about what one of those is worth: back-tested performance is hypothetical and does not reflect actual performance, and past performance cannot predict how a strategy will do in future. Hold all of the above at that arm's length. Held there, the honest conclusion comes in two halves. As a buy/sell trigger, the crossover's edge is real but perishable: it survived a data-snooping audit on ninety years of Dow data, then stopped working in the very next decade. As a regime filter — am I above or below the long-term average — it has earned its keep a different way: not by being right more often, but by sitting out the worst drawdowns.

How to actually use one

1. 1Pick the lookback by answering a question about yourself, not by trying numbers until one looks clever on the chart in front of you. How long do you intend to hold something, and how much movement are you prepared to sit through without changing your mind? That is your number, and it does not change when the chart does.
2. 2Prefer the filter to the trigger. "Price is above its long-term average" is a slow, cheap, deeply unexciting question that keeps you on the right side of long moves. "The 50 has just crossed the 200" is a specific, dramatic instruction that arrives late in a trend and arrives wrong in a range.
3. 3Check whether the last few crossings stuck. If the previous two were undone within a few weeks, you are in the sort of market that manufactures whipsaws, and the next one will probably be undone as well. The chart tells you this for nothing, before you have risked anything on it.
4. 4Budget for the lag before you need it. You will not get in at the low or out at the high — on our chart the two signals were 52 and 63 bars behind them. Any plan that only works if you catch the turn is not a plan that uses a moving average, whatever else it may be.
5. 5Use one, then stop looking. Stacking a 10, a 20, a 50, a 100 and a 200 on the same chart does not give you five opinions. It gives you one opinion drawn five times, plus a standing guarantee that some pair of them agrees with whatever you already wanted to do.
6. 6Remember which way the information runs. A moving average describes what has already happened. Every time you catch yourself reading one as a statement about what happens next, you have quietly swapped a summary for a prophecy, and it never agreed to that.