Relative Strength Index (RSI), Measured Against a Random Entry

What it does

RSI compares the size of recent gains to the size of recent losses over a lookback window, then expresses the result on a 0–100 scale. The standard window is 14 bars. Readings under 30 are conventionally called oversold and over 70 overbought, though the indicator measures the ratio of up-moves to down-moves, not distance from any fair value — a market can stay under 30 for weeks while it keeps falling.

The signals tested here are the crossings, not the states: the bar on which RSI(14) rises back through 30 (leaving oversold) and the bar on which it falls back through 70 (leaving overbought). Entering on the exit rather than on the extreme is the more common practitioner reading, and it avoids the trap of buying into an ongoing slide.

How this was measured

Every figure below comes from our own tick-derived archive, not from a vendor summary. The method is identical to the one behind our indicator signal study, so the numbers here are directly comparable to every other signal we publish.

• Data. Dukascopy bid-side OHLC for 93 instruments — major and cross FX pairs, gold, silver, stock indices and crypto — from 2010 to August 2026.
• The trade. Enter at the open of the next bar. Stop at 1.5 × ATR(14), targets at 1:1 and 1:2, walking forward up to 40 bars.
• Ties go against you. A bar holding both stop and target is scored a loss every time.
• Real costs. Per-instrument median spreads from our own bid/ask measurement.
• The control. For every signal, a trade with the same direction and the same stop distance at a random bar — five draws each. The gap between the two columns is all the signal is worth.

The result

Breakeven is a 50% win rate at a 1:1 target and 33.3% at 1:2. "Matched random" is a trade in the same direction with the same stop distance, opened at a random bar.

Two things stand out. The first is that the oversold exit is genuinely the strongest daily result in our whole indicator study — not merely positive, but significant, and larger than anything MACD, Bollinger Bands or the moving-average crosses produced. The second is that it is asymmetric: the overbought exit does not mirror it. Selling into recovering momentum is a materially worse trade than buying into it, which is what you would expect in a sample period where indices, gold and crypto trended upward.

The 4-hour rows are the cautionary half. The sample is several times larger, which drags the statistics toward significance, while the edges themselves shrink. That combination — certain, and worth almost nothing — is the single most common shape in indicator research.

How to use it

What survived the test is narrow, so treat these as the conditions under which RSI earned its result rather than as a strategy:

• Daily bars, long side, wide target. The +3.0-point result is specifically RSI(14) leaving oversold, on daily charts, aiming at 1:2. Every other combination is weaker.
• Take the exit, not the extreme. The signal is the cross back through 30, not the reading below it. An oversold market that stays oversold is a market still falling.
• Give it room. The edge is larger at 1:2 than at 1:1, meaning the signal is better at finding moves that run than at scalping small reversions.
• Use it as a filter. Three points of edge cannot carry a strategy. It can slightly improve one that already works — which is a meaningfully different claim from the one most RSI tutorials make.

What does not work

• The overbought short as a mirror trade. It is not the same signal reversed; the edge is smaller and the win rate at 1:2 sits far under breakeven.
• RSI on faster bars. On 4-hour data no RSI signal clears breakeven once the spread is paid.
• Reading a number under 30 as a buy. The state is not the signal. In a strong downtrend RSI can hold under 30 for a long time, and every bar of it is a loss for anyone who bought the reading.
• Quoting RSI win rates without a control. A 52% win rate on a long signal looks like an edge until you see that a random long in the same period won 51.8%.

Limitations

• Default parameters only. Tuning them on this same data would be curve-fitting.
• One stop rule (1.5 × ATR) and two fixed targets. Trailing stops, time exits and scaling are untested.
• Signals are taken in isolation, with no trend filter or confluence — which is how the isolated value of each is measured, but not how most people trade them.
• Bid data with modelled ask. Spread is a measured median, not a replayed tick-level ask feed.
• Spread only — no commission or swap. Both would push results further down.
• A statistically significant edge of one point is still only one point. Large samples make small effects certain; they do not make them large.