RSI divergence is the pattern traders screenshot most and test least: price pushes to a new low while RSI(14) refuses to follow and prints a higher low (bullish), or price makes a higher high while RSI makes a lower high (bearish). The reading is that momentum is quietly leaving the move before price admits it — the classic setup for fading an exhausted trend. It is taught everywhere, and almost never with rules precise enough for a machine to trade.
Here is the honest status. Detecting divergence mechanically means defining swings on two series at once — price and the oscillator — and our engine does not ship that primitive yet. We publish only numbers our own deterministic engine produced on real data, so this page carries the exact rules we will run, not borrowed statistics. The simpler cousin has already been tested: plain RSI(14) 30/70 mean reversion ran through the same 12 months of BTC and ETH data, and those results are live at /strategy/mean-reversion. The divergence run lands here the moment the detector ships.
The exact rules we’ll test
- Compute RSI(14) on the trading timeframe (planned first run: 1-hour candles, 24/7).
- Define swings identically on both series: confirmed fractals (lookback 3) on price and on the RSI line — a swing exists only after 3 later candles close, no look-ahead.
- Regular bullish divergence: price prints a lower low while RSI prints a higher low across the same two swings; bearish is the mirror — higher high in price, lower high in RSI.
- Enter at the close of the candle that confirms the second swing; stop beyond the price extreme of the divergence; target 2R.
- Risk 1% of equity per trade; 0.05% commission per side; 10× max notional leverage.
- Missing primitive: swing-to-swing divergence detection between price and an oscillator is not yet supported by the engine — this page gets its results table the moment it ships.
Backtest data: in the queue
We publish only numbers our own deterministic engine produced on real data — no borrowed or estimated stats. This setup needs a detection primitive the engine doesn’t ship yet, so its 12-month results will appear here the moment that lands. The strategies with completed research are on the strategies hub, and you can already practice this setup bar-by-bar in the free replay demo.
Why every "RSI divergence" backtest disagrees
Divergence looks binary on a marked-up chart and is anything but in code. Which swings count (fractal lookback 2, 3, 5?), how many bars apart the two swings may sit, whether the first swing must come from oversold or overbought territory, regular versus hidden divergence, confirmation on the close versus intrabar — every one of those choices rewrites the trade list. Two honest backtests of "RSI divergence" can share almost no trades, which is exactly why it resists being a one-line primitive and why quoted win rates for it are all over the map.
On top of the definition, discretionary traders stack context filters: only at a higher-timeframe level, only with the trend (hidden divergence), only after a structure break confirms, only when RSI actually visited 30 or 70 first. Each filter shrinks the sample and changes the stats — and none of them can be judged until the unfiltered baseline exists. The nearest published baseline is already live: plain RSI(14) 30/70 mean reversion on the same 12 months of data. The divergence filter has to beat that to earn its complexity.
How to backtest RSI divergence on Secuora
Until the detector ships, the honest path is manual replay plus the supported RSI baseline:
- Open the free replay demo at /backtest/demo (no sign-up) and add RSI to the chart — it is one of the built-in indicators, alongside EMA, MACD, Bollinger, VWAP, ATR and Stochastic.
- Step through the chart bar by bar and mark price swing extremes and RSI swing extremes with the drawing tools — only count a divergence once the second swing has confirmed.
- Sign up free and trade each qualifying divergence in the replay terminal with simulated stop-loss and take-profit orders — stop beyond the price extreme, 2R target — so the engine scores your sample honestly.
- Run the supported baseline in the AI backtester at /backtest/ai: RSI(14) crossing back through 30/70 is a built-in primitive, so you can see what plain RSI mean reversion earns with no divergence filter at all.
- Journal both samples (rules, confluences, screenshots) and withhold judgment until you have at least 30 replayed divergences — if they cannot beat the plain-RSI baseline, the pattern added nothing.
Frequently asked questions
What is RSI divergence?
A disagreement between price and momentum: price makes a new swing extreme while RSI fails to make one in the same direction. Regular bullish divergence is a lower low in price with a higher low in RSI; regular bearish is a higher high in price with a lower high in RSI. Traders read it as the move running out of participation before it runs out of price.
What is the difference between regular and hidden divergence?
Regular divergence signals a potential reversal — price extends to a new extreme, momentum does not. Hidden divergence is the inverse arrangement (price makes a shallower extreme while RSI makes a deeper one) and is read as trend continuation. Articles routinely mix the two under one name, which is one more reason quoted "divergence win rates" are not comparable.
What win rate does the RSI divergence strategy have?
We do not quote one yet, on purpose. Divergence detection is not a primitive in our engine, and we publish only statistics our own deterministic engine produced on real data — the 12-month run lands on this page when the detector ships. The closest tested relative, plain RSI(14) 30/70 mean reversion, already has published results at /strategy/mean-reversion.
Is RSI divergence a reliable strategy?
It is one of the most-taught and least-tested patterns in trading, precisely because "divergence" hides a stack of definitional choices that change the trade list. Any reliability claim that does not publish its exact swing rules is unverifiable. Treat it as a hypothesis to test against a plain-RSI baseline, not as a fact.