What is RSI divergence?
RSI divergence is a disagreement between price and the relative strength index at consecutive swing points: regular bearish divergence is a higher high in price with a lower high in RSI, regular bullish divergence is a lower low in price with a higher low in RSI, and the hidden forms are the reverse pairings read as continuation signals.
The reasoning is that RSI measures the pace of recent gains against recent losses, so a new price extreme reached with less momentum than the previous one suggests the move is tiring. The signal is therefore a claim about momentum, expressed as a comparison between two swing points on each of two series.
That last phrase is where the difficulty lives. A comparison between two swing points requires knowing which points are swings, on both price and the oscillator, and both series have to be read as of the same bar. A chart drawn after the fact chooses the points that made the divergence visible; a rule has to choose them before knowing whether a divergence will appear.
The hidden forms add a second set of comparisons with the opposite interpretation, and a test should treat them as separate signals rather than pooling everything labelled divergence.
Why is divergence hard to count?
Divergence is hard to count because it needs a swing rule on two series, a lookback that decides how far apart the two swings may be, and a confirmation that says at which bar the divergence became known, and each of those is chosen by eye in most presentations of the signal.
The swing rule on price is the familiar problem. A peak with two lower bars on each side is a different set of points from a peak with six. The swing rule on the oscillator is the same problem again and is almost never stated, because on a chart the oscillator's peaks are simply read off wherever they look highest.
The lookback bounds the pair. Two peaks a hundred bars apart can be connected by a line, but a signal that compares momentum across a hundred bars is a different signal from one that compares across ten. A fixed maximum separation is needed, and a minimum too, since adjacent bars are not a meaningful pair.
Confirmation is the hindsight leak. A price peak is only a peak once the bars after it have closed, and the same applies to the oscillator's peak. The divergence is therefore knowable only some bars after the second peak, and a test that enters at the peak itself has used bars it did not yet have.
How do you define it mechanically?
Define it mechanically by fixing a swing rule and applying it to both price and RSI, requiring the two most recent confirmed swings on each series to fall within a stated separation, comparing their heights, and marking the divergence on the bar at which the second swing was confirmed.
Applying one swing rule to both series is a simplification that removes a parameter and makes the definition easy to state. It is not the only choice; the oscillator is smoother than price and a tighter rule on it may be defensible. Whatever is chosen has to be chosen before the results are seen.
The comparison itself needs a tolerance. A higher high in price by a fraction of a pip against a lower high in RSI by a hundredth of a point is technically divergence and practically noise. Requiring the price difference to exceed a volatility-scaled minimum and the oscillator difference to exceed a fixed minimum removes the noise cases at the cost of some real ones, and the threshold is a parameter to report.
Marking the event at the confirmation bar rather than the peak bar makes the signal later and less pretty. It is also the only version a trader could have acted on, and a test that grades the earlier version is grading something that never existed.
What baseline must divergence beat?
Divergence must beat the reversal rate at oscillator extremes that occur without divergence, because the RSI being at an extreme is already a mild reversal signal on its own, and the divergence is only worth its extra conditions if it improves on that.
This is the control that most divergence claims omit. RSI above seventy or below thirty, with no divergence at all, is followed by a pullback some fraction of the time, simply because extremes are followed by regression more often than not. If divergence at those extremes is followed by a pullback at about the same rate, the divergence added nothing; the extreme was the signal.
The second baseline is the random bar. Over the same horizon, how often does a reversal of the stated size occur starting from any bar at all? A signal whose reversal rate is not materially above that number is not a signal.
Both comparisons need the same horizon and the same definition of a reversal, and the honest report shows all three rates with their counts.
How does QuantParadox test RSI divergence?
QuantParadox has no divergence primitive; a divergence rule compares the RSI value at the two most recent confirmed price swings with the price at those swings, using the platform's swing rule and its as-of-the-bar confirmation, which makes the definition explicit and later than a drawn line.
The absence of a single divergence flag is deliberate. A flag would encode one of the many possible definitions silently, and a trader who meant a different one would be grading a rule they did not write. Composing the rule from the swing points and the oscillator values forces every choice, the swing rule, the separation, the tolerance, to be stated.
The limitation is that the platform's swing rule is applied to price, and the oscillator is read at those bars rather than at its own independently detected peaks. That is one of the defensible definitions and not the only one, and a trader who wants the oscillator's own swings expresses that as an additional condition.
The baselines are straightforward to construct from the same primitives, the extreme-without-divergence group as a comparison rule and the random-bar rate as a control, and the out-of-sample split applies to the whole comparison.