Technical Analysis · Chapter 5 of 12
Moving averages and trend following
The oldest indicator, the one with the best evidence behind it, and a trade-off between lag and noise that no setting can escape.
A moving average is the simplest possible smoother: the mean of the last n closes. Everything about it — its usefulness and its limitations — follows from one fact: it is a lagging function of price by construction, and the lag is the price you pay for the smoothing.
The variants, and when the difference matters
| Type | Weighting | Lag | When to prefer it |
|---|---|---|---|
| SMA — simple | Every bar equally | Highest, ≈ n/2 bars | When you want a stable, well-understood reference. Most published research uses SMA |
| EMA — exponential | Recent bars more | Lower | When responsiveness matters more than stability. Never fully forgets old data |
| WMA — weighted | Linearly declining | Between the two | Rarely worth the extra complexity over EMA |
| Hull, KAMA, ZLEMA | Various adaptive schemes | Lowest | They reduce lag by extrapolating, which means they overshoot at turns. There is no free lunch here — you have moved the error, not removed it |
The honest position on variants: the difference between SMA and EMA at the same period is usually far smaller than the difference between two periods of the same type. If your strategy's result flips when you change SMA to EMA, you have a fragile strategy rather than an interesting finding. Test that explicitly — it is a cheap and very informative robustness check.
The three ways moving averages are used
- As a trend filter. Only take long signals when price is above the 200-day average. This is the highest-value use by a wide margin. It is not a signal, it is permission, and adding a long-term trend filter improves the risk-adjusted results of a great many strategies — mostly by keeping them out of sustained declines rather than by improving the entries.
- As a dynamic support or resistance reference. Price pulling back to a rising 50-day average is a reproducible, non-arbitrary version of ‘buy the dip in an uptrend’. Its main virtue over a drawn trendline is that it is computed, so two people get the same answer.
- As a crossover signal. The classic golden cross and death cross. This is the weakest use and the most popular, which is a familiar pattern. Crossovers are late by construction and generate a great many false signals in ranges. They are not useless — long-horizon crossover systems have a real record in trend-following futures — but a 50/200 crossover on a single stock is not that.
What trend following actually is, and why it is hard for reasons that are not analytical
Time-series trend following — long when above a long-term average, flat or short when below — has a genuinely good long-run record, particularly across a diversified basket of futures. It is one of the few technical approaches with decades of live, audited track record rather than only backtests.
Its characteristic profile is low win rate, high payoff ratio. You will be wrong on most trades, with many small losses and a few very large gains. That distribution is mathematically fine and psychologically brutal.
The failure is almost never the rule. It is that people abandon it during the drawdown, which can run for years, and it is why the honest question to ask of any trend-following system is not ‘what is the return?’ but ‘what is the longest period underwater, and would I have kept going?’
Cross-sectional momentum, which is different and better evidenced
Do not confuse the two. Time-series momentum asks ‘is this instrument going up?’. Cross-sectional momentum asks ‘which instruments are going up most, relative to each other?’ The second has the stronger academic record, including in Indian equities.
- The canonical construction is 12-1: rank the universe by return over the last twelve months excluding the most recent month, buy the top decile, rebalance monthly. The skipped month matters — short-horizon reversal works against you, and including it measurably degrades the result. That detail is a good example of a finding you would never guess and would discover by testing.
- It is a relative signal, so it needs a universe and it needs the universe to be point-in-time. Ranking today's index members over the last year is survivorship bias in its purest form.
- Turnover is the practical constraint. Monthly rebalancing of a decile produces substantial turnover, and in India the cost stack — STT on both delivery legs, plus impact — consumes a meaningful share of the gross edge. Any credible test must be net of realistic costs, and reporting gross returns here is close to dishonest.
- Momentum crashes. The strategy has severe, well-documented drawdowns at sharp market reversals, because the recent winners are precisely what unwinds hardest. Anyone presenting momentum without mentioning this has not read the literature.
A test you can run this week
Take a liquid Indian index. Compare four rules over ten years, net of realistic costs, with results reported per year rather than aggregated:
- Buy and hold.
- Long only when the close is above the 200-day SMA, otherwise in cash.
- The same, but with a 1% band around the average to reduce whipsaw — and note whether the band helps, because it usually does and almost nobody adds it.
- A random control: in the market on the same number of days, chosen at random, averaged over at least five draws.
The fourth is the one that makes the exercise worth doing. If the trend filter's advantage over the random control is small, then most of what you measured was simply being invested rather than being invested at the right times. That is a completely different conclusion, and it is the conclusion most people never reach because they never build the control.
Check yourself