Oscillators, Momentum and Divergence

The short answer

An oscillator takes a window of recent prices, compares the latest ones with the rest of that window, and rescales the comparison into a bounded number — usually 0 to 100.

Everything people find surprising about oscillators follows from that sentence. The output is relative to the window and to nothing else, which is why an extreme reading says "recent closes were lopsided" rather than "price is too high", and why a strong trend can hold the reading at an extreme for as long as it runs.

How one is built

Two common constructions, written out in full.

A relative-strength style oscillator compares the size of up-closes with the size of down-closes:

average gain = mean size of up-closes over the last N periods
average loss = mean size of down-closes over the last N periods
RS          = average gain ÷ average loss
reading     = 100 − (100 ÷ (1 + RS))

A range-position style oscillator asks where the latest close sits inside the window's high-low range:

reading = 100 × (close − lowest low over N) ÷ (highest high over N − lowest low over N)

Read the second formula closely. If the close is the highest price in the window, the numerator equals the denominator and the reading is 100. Not "very overbought" — 100, by arithmetic, because the close is at the top of its own range. The first formula behaves the same way: if there were no down-closes in the window, average loss approaches zero, RS grows without bound, and the reading approaches 100.

Neither formula contains earnings, supply, positioning, or any notion of what the asset is worth. It contains a window of closes and a division.

Normalisation is the trick and the trap

Bounding the output is what makes oscillators pleasant to use. A number between 0 and 100 can be compared across instruments priced in entirely different units, and it can be plotted on one panel regardless of the underlying scale.

That convenience comes at a cost: the scale is local. A reading of 80 on one instrument and 80 on another describes two different things, because each is relative to its own window, its own volatility and its own timeframe. The shared scale invites a comparison the construction does not support.

It also means the reading is bounded while price is not. Price can keep rising after the oscillator reaches its ceiling — the indicator has simply run out of room to express what is happening. An indicator pinned at its maximum is not describing an extreme market; it is describing a market that has exceeded the indicator's vocabulary.

What "overbought" actually means

Unpack the reading and the word disappears.

"Overbought at 78" means: over the last N periods, up-closes have outweighed down-closes to this degree. That is a statement about arithmetic on a window of history.

It does not mean the asset is expensive — no valuation input exists in the formula. It does not mean buyers are exhausted — the indicator cannot see who is buying. It does not mean a reversal is due, and this is the expensive one. The oscillator is high because the trend is strong, so an extreme reading is evidence of trend strength before it is evidence of anything else. Selling into it is selling because the trend is working.

The thresholds themselves — 70 and 30, 80 and 20 — are conventions inherited from the indicators' original authors. They are not derived from market behaviour, they are not calibrated per instrument, and there is nothing that makes 70 more meaningful than 68.

A more defensible reading of the same number: this instrument has moved a lot in one direction recently, so the risk profile of joining now is different from the risk profile of joining early. That is a statement about the position you would be taking, not a prediction about the market, and it can be acted on with sizing rather than with a contrarian entry.

Divergence in a strong trend

Divergence describes price making a higher high while the oscillator makes a lower high, or the mirror image at lows. It is a real observation with a mechanical explanation.

The oscillator measures the rate of recent change, not the level of price. A trend that continues at a slightly slower pace makes new highs in price while producing a smaller reading than its earlier, faster leg. Divergence is the ordinary signature of a trend that is still advancing but no longer accelerating.

Deceleration is not reversal. A move can decelerate and then re-accelerate, or decelerate and drift higher for months. This is why divergence appears repeatedly inside long trends: each one is a true description of the momentum series, and none of them is a statement about what happens next.

Two habits keep divergence honest.

• Give it a level. If you act on divergence, name the price at which you are wrong before you act. Divergence alone supplies no invalidation, which is why it pairs badly with discretion.
• Refuse the retroactive labels. When divergence is followed by continuation, it is often renamed "hidden divergence" or declared "failed". A concept with a name for every outcome cannot be tested, and what cannot be tested cannot be relied on.

Parameters change the signal quietly

The lookback period is not a detail. It is the definition of "recent", and changing it changes what the indicator says while the market does nothing at all.

The numbers below are invented for teaching. They describe no real instrument and are not a trade idea.

Imagine an instrument that has risen steadily for thirty sessions with three small pullbacks. On a 7-period setting, each pullback pulls the reading well down and each resumption pushes it back to an extreme: several "overbought" episodes and a couple of divergences. On a 21-period setting, the same thirty sessions produce one long stretch of elevated readings and no clean divergence at all. Same prices, same chart — two different accounts of what happened.

Three rules follow.

1. Choose the period for a reason you can state, ideally one about the horizon you trade rather than about how the chart looks.
2. Expect graceful degradation. A setting that captures something real should work about as well at 12 or 16 as at 14. A sharp peak at exactly one value is the signature of overfitting to the sample, the same failure described in the moving averages lesson.
3. Fix it before testing. Tuning the period until the signals line up with known outcomes is curve-fitting with extra steps. The backtesting lesson sets out how to test so that a bad parameter can actually fail.

Any rule you build is a rule about that parameter on that timeframe, not a discovery about the market.

What to do with this

Run one oscillator at most, at a period you chose deliberately and do not change.

Read it as a description of condition rather than as an instruction: extreme means the recent window has been one-sided, divergence means the pace has eased. Then decide separately, using structure and levels, whether there is a trade — and let the entry, the invalidation and the size come from that decision rather than from the number on the lower panel.

The reading tells you what recent price has been doing. It was never able to tell you more than that, and the traders who get the most out of oscillators are the ones who stopped asking them to.