Technical Analysis · Chapter 6 of 12
Oscillators: RSI, MACD and the overbought fallacy
How they are built, what divergence really is, and why the most common way people use RSI is the reverse of what the indicator supports.
Oscillators normalise recent price change into a bounded or centred range so that different instruments and periods become comparable. That normalisation is genuinely useful and it is also the source of the most expensive misconception in retail technical analysis.
RSI, actually constructed
Relative Strength Index, Wilder's original, over n periods (default 14):
- Split each bar's change into gain and loss. Take a smoothed average of each — Wilder's smoothing, which is an EMA with α = 1/n, not a simple average. This detail matters: implementations that use a simple average produce visibly different values, and mismatched RSI between your backtest and your platform is a classic and maddening bug.
- RS = average gain ÷ average loss.
- RSI = 100 − (100 / (1 + RS)).
- So RSI is a normalised ratio of up-movement to total movement over a trailing window. Nothing more. It is not a measure of value, expense, or exhaustion.
“Overbought” means strong, not expensive
The most common retail use of RSI is to sell above 70 and buy below 30. In a trending market this is close to the worst possible rule, and the reason is structural rather than empirical: RSI is high precisely because the instrument has been going up. It is a momentum measure, and selling high momentum is selling the thing that has the best-evidenced continuation property in the whole discipline.
In a strong trend RSI pins above 70 and stays there for weeks. Every counter-trend entry loses. This is not a rare edge case; it is what happens in every substantial trend, which is where the money is.
Wilder himself described the levels as indicating strength. The interpretation as ‘buy’ and ‘sell’ signals was added later by other people.
The defensible uses: RSI as a filter within an already-decided direction (buy pullbacks to RSI 40 in an uptrend, not RSI 70 reversals); RSI as a cross-sectional ranking tool; or RSI thresholds as a regime classifier, where a persistently high reading tells you to use faster, trend-following rules rather than mean-reverting ones. That last use is genuinely good and rarely taught.
MACD, which is two moving averages wearing a disguise
MACD line = 12-period EMA − 26-period EMA. Signal line = 9-period EMA of the MACD line. Histogram = the difference between them.
That is it. It is a moving-average crossover system with a smoothing on top, plotted separately. Understanding this removes most of the mystique and tells you exactly what to expect: it is a trend indicator and it inherits every trend indicator's properties — late at turns, excellent in sustained moves, poor in ranges. The histogram crossing zero is simply the MACD line crossing its own signal line, which is a smoothed version of an already-lagging quantity.
Divergence: the one genuinely interesting idea here
Divergence is when price makes a new extreme and the oscillator does not. Price makes a higher high, RSI makes a lower high. The interpretation is that the second push had less momentum behind it.
This is a real observation and it is also where hindsight does its worst work in all of technical analysis, for a reason worth stating plainly: divergences are extremely common, and most of them resolve by the oscillator catching up rather than by price reversing. On any chart you can find a dozen. The ones in textbooks are the ones that preceded reversals.
How to actually test divergence — and this is a good research exercise.
Define it mechanically: a swing high in price above the previous swing high, with the RSI value at that bar below the RSI value at the previous swing high, within a maximum bar distance. Every term there is codeable.
Count all of them, not the ones you noticed. This alone will surprise you.
Define the outcome in advance: a decline of at least x% within k bars. Fixing this before you look is what stops you from choosing the definition that makes the result work.
Build the control: the same measurement at randomly selected swing highs without divergence. This is the step that decides the question.
Report expectancy, not hit rate, net of costs.
When this is done carefully, divergence usually shows a small effect that is nowhere near as strong as its reputation, and is heavily dependent on the maximum-distance parameter. That is a useful, publishable, honest result — and being the person who runs the test rather than repeating the claim is exactly the profile a research role hires.
The family trap, restated because it matters
RSI, Stochastics, CCI, Williams %R and ROC are all normalised rate-of-change measures. They are computed differently and they are measuring the same thing. Requiring three of them to agree before entering is not confirmation — it is a single opinion with a delay added, because the slowest of the three now gates your entry.
Genuine confirmation comes from a different family: a momentum reading confirmed by volume, or by volatility regime, or by structure. That is an independent measurement. Three oscillators are not.
Check yourself