What Is a Triple Exponential Moving Average?
Triple exponential moving average applies the same lag-correction as DEMA one layer further, stacking three exponential averages to sit closer to price still. It is the most responsive of the smoothed averages and the most prone to reversing on moves that go nowhere.
TEMA takes the lag-correction behind DEMA and applies it again. It is the fastest of the smoothed averages, and the clearest demonstration of where this whole line of thinking runs out.
How it works
DEMA uses two exponential averages — one of price, one of that average — and adds the gap between them back on to cancel most of the lag.
TEMA adds a third. An average of the average of the average, combined so that the residual lag left over by DEMA is cancelled too.
Each layer moves the line closer to price. The sequence is consistent: simple average, exponential, DEMA, TEMA — progressively faster, progressively less smooth.
Where the idea runs out
The limit is price itself. Perfect lag removal produces a line identical to the closing price, which carries no information an ordinary chart does not.
So every layer has diminishing value. DEMA removes most of the lag an EMA has; TEMA removes most of what DEMA left, which is a smaller quantity to begin with.
And each layer adds arithmetic. More computation, more sensitivity to the starting values, and a line that is harder to reason about when it does something unexpected.
A worked example
Take this site’s shared series. Direction runs average 2.01 bars with a longest of 11; the median bar range is 0.493 and the ninetieth percentile is 1.101.
Set a 20-period average of each kind. The simple average turns last, the EMA sooner, DEMA sooner again, TEMA first.
On the eleven-bar run, that ordering is a gain. TEMA is in earliest and out latest, and captures the most of it.
On the two-bar runs — which is most of them — the ordering reverses into a cost. TEMA signals on nearly all of them, each costing a round trip of 0.0098, about 2% of a median bar.
What it is actually good for
Confirming a trend already identified elsewhere. As a fast line against a slow one, TEMA gives a responsive crossover without needing to be the whole system.
Short-horizon work where lag is genuinely expensive. A scalper holding for a handful of bars cannot afford an average that turns five bars late, and accepts the noise as the cost of that.
As a visual aid rather than a trigger. Plotted without acting on every cross, it shows the shape of recent movement more legibly than raw candles do.
What it is not good for is being the signal on its own, in a market that ranges. Nothing about the construction distinguishes a real turn from a two-bar wobble — it reacts to both faster, and that is the entire specification.
The original data
On this site’s shared series: median bar range 0.493, ninetieth percentile 1.101, largest bar 2.338. Direction runs average 2.01 bars with a longest of 11. A round trip costs 0.0098, about 2% of the median bar range.
The 2.01-bar average run is the number that decides this. An average tuned to react within two bars reacts to the market’s most common behaviour, which is changing its mind.
And the cost compounds: this site’s fee measurement shows 75 basis points a year costing 20.2% of a thirty-year balance. A spread paid on every false turn is the same arithmetic on a shorter clock.
The family, in order
Simple moving average — every bar weighted equally, slowest, smoothest.
Exponential — recent bars weighted more, faster, still stable.
Double exponential — lag measured and subtracted once.
Triple exponential — subtracted again, the fastest of the four.
And Hull sits alongside rather than beyond, using weighted averages and a square-root period to reach similar responsiveness by a different route. Choosing between them is choosing a point on one axis, not finding a better indicator — and the point that suits you depends on your holding period, not on which formula is cleverest.
Why three and not four
The correction is a series, and it converges. Each layer cancels the lag the previous one left, and what it leaves is smaller every time, so the gain from a fourth layer is a fraction of the gain from the third.
Meanwhile the cost does not shrink. Every layer is another exponential average computed on the output of the last, so noise introduced early is amplified through the stack rather than damped by it.
Quadruple constructions exist and nobody uses them. Not because of tradition, but because the line they produce is close enough to price that the smoothing has stopped doing its job.
Which makes TEMA roughly the end of the useful range. It is the last point on this axis where the output is still recognisably an average rather than a repainted price series, and that is the honest case for it over anything faster.
When it fails
The characteristic failure is mistaking responsiveness for accuracy. TEMA turns before the other averages, so in review it looks prescient on every move that ran — the eye goes to the trends and skips the flat stretches. What the chart does not make obvious is how many times it turned during the ranges, because those turns left no memorable mark. The indicator is genuinely earlier; it is earlier about everything, and only some of everything was worth being early about.
A second failure is using it at a short period. The construction is already fast; a short period makes it a redrawn price line.
A third is stacking it with DEMA or Hull. They respond to the same moves, so agreement is repetition rather than confirmation.
A fourth is judging it on a trending backtest, which is the one condition it is built for.
And a fifth is looking for a faster one still. The limit is price, and price with no smoothing is what you started with.
Related
Double exponential moving average covers the construction one layer back. Exponential moving average covers the building block. And moving average covers what the whole family is doing.
There is a natural end to this idea and TEMA is close to it. Keep removing lag and eventually the line is price, at which point you have an indicator that tells you what you could already see. The useful question is not how fast an average can be made, but how slow it can afford to be.
— Michael Whitman, from this video
This page is educational, not financial advice. Test every idea on your own charts before risking money.