We measured RSI: its edge flips when the exit moves

Every tuning guide answers the same question: does this entry filter work? The answer usually arrives as a single number — a win rate, an expectancy, a line on an equity curve.

We measured our own RSI entry filters across seven months of hourly candles. The number changes sign depending on where you take profit. Not the filter. The exit.

What we measured

The instrument is read-only and needs nothing private — public candles, no keys, no database.

• •Window: 22 February – 18 September 2026 (UTC), hourly bars, ~5,000 per pair.
• •Market: ten perpetual pairs on Bybit — BTC, ETH, SOL, XRP, DOGE, ADA, LINK, AVAX, LTC, DOT.
• •Rules: the RSI entry presets our own product ships, at the exact thresholds a user receives. A long enters at RSI ≤ threshold, a short at the mirrored value.
• •Measure of edge: a race the size of the take-profit. From the close of the signal bar, did price travel X % in our favour before it travelled X % against? A bar that touches both edges is scored a loss — we do not know the order of events inside a bar, and you must not round in favour of your own conclusion.
• •Baseline: entering on *every* bar, same race. The edge is the difference in win rate, in percentage points.
• •37,059 signals judged.

Then we ran it three times, changing exactly one thing: the size of the race — 1 %, 2 %, 3 %.

Finding 1: the same filter, three exits, three answers

Nothing about the filter changed between those columns. The same bars, the same thresholds, the same 2,209 short signals in every column. Only the distance to the exit moved — and the RSI ≥ 70 short travels from clearly negative to apparently positive.

There is a reason, and it is not noise. An entry filter is a claim about where price sits in its local range. RSI ≥ 70 says "extended upward". Over a 1 % race an extended market usually keeps extending, and those signals lose. Over a 3 % race the horizon is long enough for reversion to arrive. The rule did not get better; the question changed.

Which is why "does RSI work" has no answer. It is not a property of the indicator.

Finding 2: the average passes, the stability test does not

Our instrument refuses to call an edge on the average alone, and this measurement is why.

2,115 signals are not 2,115 independent observations. They are computed on overlapping bars: an RSI below 30 usually stays below 30 for several hours, so a single drawdown on a single pair contributes dozens of near-identical "signals". Averages over clumped data look far more decisive than they are.

So the verdict checks the sign of the advantage in every cell of *pair × half-of-window* — twenty cells — and requires it to hold in 70 % of them.

Across the four rule/side combinations at three exit sizes — twelve verdicts — exactly one clears that bar: RSI ≤ 30 long at a 1 % take-profit (+3.91 pp, 15 of 20 cells). The same rule at a 2 % exit has a nearly identical average (+3.55 pp) and fails at 11 of 20. An average anyone would publish, and a coin flip on the sign.

Finding 3: an average does not say which instrument paid for it

Take the RSI ≥ 70 short at a 3 % exit. Average +1.83 pp — it reads like a working filter. Per pair:

ETH +13.8   DOGE +11.0   AVAX +6.7   LINK +3.6   ADA +0.5
LTC  −1.7   BTC   −2.1   SOL  −2.5   XRP  −2.6   DOT −10.3

Positive on five pairs of ten. Two instruments carry the other eight.

Now the one candidate that did earn a verdict — RSI ≤ 30 long at a 1 % exit:

DOT +11.3   LINK +5.9   XRP +5.5   ETH +5.3   ADA +5.2
AVAX +3.4   SOL  +2.8   DOGE +2.7  LTC +0.7   BTC −3.1

Positive on nine of ten. Similar-looking summary statistics; entirely different claims about the world.

Finding 4: mirrored thresholds are not a mirrored rule

RSI ≤ 30 long fires on 4.3 % of bars. RSI ≥ 70 short fires on 4.5 %. The two sides are almost exactly as selective — by construction, they are the same strictness reflected.

They do not behave the same way. The long side is positive on average at all three exits (+3.91 / +3.55 / +3.01) and broad (9, 8 and 6 pairs of ten). The short side swings from −1.90 to +1.83 and never gets past 6 of 10.

This is not a claim that longs are better than shorts. It is a claim that a symmetric setting is not a symmetric rule — and a product that mirrors the number and calls the job done is shipping two different behaviours under one label.

We have already paid for that once. One of our presets shipped RSI 45 for longs and 70 for shorts for two months, because the long side had been "fixed" in response to a complaint. A live short system then went 52 hours without placing a single order, while its paper twin on a byte-identical configuration traded four cycles.

If you are tuning a system

• •Never judge an entry filter without naming the exit. The pair is the unit of measurement. A filter is not good or bad; it is good or bad *at a horizon*.
• •Signal count is not sample size when bars overlap. Check whether the sign survives being cut by instrument and by time.
• •Ask where the average lives. "+1.8 points across ten pairs" and "+1.8 points, two pairs carrying eight" are different findings printed identically.

How to check this yourself

You need candles and about forty lines. Pick your rule, walk the bars, and for every bar where the rule passes, run the race forward to the first touch of +X % or −X %. Do the same for every bar without the rule. The difference in win rate is your edge. Then cut it: by instrument, and by the first and second halves of your window. If the sign does not survive both cuts, you measured a market regime, not a rule.

Limits of this measurement

State them plainly, because a measurement without its boundaries is a claim.

• •One venue's candles, ten pairs, seven months — that is one market regime, not "the market". February–September 2026 is what we looked at.
• •The measure is a race, not a P&L. No fees, no slippage, no funding, no position sizing. That matters here: a round trip costs roughly 0.1–0.2 % on these venues, which is a large share of a 1 % race. A filter with an edge on this measure can still lose money after costs.
• •Signals overlap. The stability test is a guard against that, not a cure.
• •Ties inside a bar are scored as losses. This biases every number downward — equally for the filter and for the baseline.
• •We measured RSI only. Our product ships 29 entry filters; this says nothing about the other 28.
• •This is not investment advice, and none of the above is a recommendation to trade any of it.

What we changed

Two things in the instrument, both of them gaps this article exposed. It used to name its window as "5,000 bars" — which names neither a start nor an end, so the same run tomorrow is a different measurement and nobody can check yours against ours. It now reads the boundaries off the candles themselves and prints dates.

And it now prints the per-pair breakdown, because the summary line could not distinguish "+3 points across ten pairs" from "+3 points collected on one". Those were being printed identically, and only one of them is an edge.