Moving Averages: Smoothing, Not a Signal

A moving average predicts nothing: it retells what has already happened in a more readable form. It is arithmetic over a window — you take a stretch of the series, calculate its mean, shift the window a step forward, and repeat this across the whole history. The only thing such a construction can do is damp fluctuations that are shorter than the window. That is exactly why it is useful for reading the level and useless as an order to trade: the signal you see is assembled from prices that have already receded into the past.

What a moving average actually calculates

The mechanism is simpler than its reputation. A window of fixed length slides along the series, returning at every point the mean of the values that fall inside it. A new value comes in, the oldest drops out. Three properties follow from this, and everything else follows from them.

The average is not defined everywhere: until the window has filled up there is no value, and the smoothed series has a shorter start to its history than the original series does. The average changes for two reasons at once — not only when a new price arrives, but also when an old price leaves: a sharp jump from half a year ago will move a long average on the day it drops out of the window, even though nothing has happened in the market today. And the average has no meaning of its own in money terms: it is a characteristic of the series, not a price at which anyone is prepared to trade. The general logic of the transformation is covered in the glossary entry on series smoothing.

Why it is always late

The centre of gravity of the window lies in its middle. This means the smoothed value describes not today but the notional midpoint of the stretch — and the longer the window, the further it lags behind the price. This is not a flaw in the implementation, nor something that can be fixed by tuning: the lag is the price paid for suppressing noise, and the only way to reduce it is to shorten the window, which means letting part of the noise back into the series.

An unwelcome consequence for trading systems follows: any rule of the form "the average has turned — we buy" is triggered after the price has reversed, not before. In a strong trend this costs part of the move. In a sideways market it costs far more: the price oscillates around the average, crossovers come in succession, and each of them is a trade with costs attached.

Simple and exponential: different memory

A simple average gives every value in the window the same weight and abruptly forgets whatever has fallen over the edge. An exponential average applies declining weights: recent data weigh more, older data weigh less, but they never disappear, and the memory stretches out in a tail. The practical difference lies in how the series reacts to an outlier. A simple average will jerk twice from one shock — when it enters the window and when it leaves it. An exponential average will jerk once and then fade smoothly.

Neither of them is "more accurate". They answer different questions: the simple average answers "what was the average level over the stretch", the exponential answers "what is the level now, if the recent matters more than the distant".

The window is a hypothesis, not a setting

The length of the window determines which fluctuations you declare to be noise. A short window treats intraday jitter as noise and leaves weekly moves visible. A long window declares almost everything to be noise except a drift lasting many months. In choosing the length you have already made a substantive statement about the scale at which you are looking at the security — and from then on the chart merely confirms your choice.

The danger begins when the length is fitted to the past: by trial and error you find the window that produced attractive entries on historical data. Such curve fitting almost always carries over to the future worse than the backtest promises, because it has been fitted to a particular realisation of noise rather than to the mechanism of the market.

What corrupts the series before you average it

Smoothing does not correct errors in the input data; it smears them. The main source of distortion in the Russian market is the dividend cut-off: on the ex-dividend day the quote drops by the amount of the payout, and an unadjusted series will show a slump that did not exist in economic terms. It is convenient to check the dates for a security in the dividends section, and to take the series itself already adjusted. Splits and reverse splits break the series more crudely still — there the gap is a multiple. The instrument card SBER shows what history is available and in what form.

Where smoothing really works

The most honest use of a rolling window is not the price but financial reporting. Quarterly figures are seasonal and uneven, and comparing quarter with quarter is pointless. A rolling annual window removes the seasonality: the stretch shifts with each new publication while remaining comparable with itself. This is precisely TTM — see the glossary entry on TTM for more detail, while the latest publications are gathered in the reporting calendar.

Here the lag turns from a weakness into a strength: we do not need a reaction to the latest day at all; we need a stable level of profit or revenue on which valuation multiples are calculated.

How to work with it

Use the average as a lens, not as an oracle. Look at share prices together with the news feed: if the average has turned, the reason for the turn has almost always already been named in an event, not in the geometry of the line. Definitions of related concepts are gathered in the glossary.