What Is RSI? How Traders Use It and How to Test It Honestly

If you've looked at a trading chart, you've probably seen RSI: the little line under the price that moves between 0 and 100.

It's also one of the first things people ask an AI to build a strategy around. This post covers what RSI measures, how traders use it, and how to test an RSI idea so you know whether it works or just looks like it does.

What RSI measures

RSI stands for Relative Strength Index. J. Welles Wilder introduced it in 1978. It looks at recent price moves and asks one question: lately, have the up moves been bigger than the down moves, or the other way around?

Here's the idea in plain steps, using the common setting of 14 candles:

1. Look at the last 14 candles.
2. Take the average size of the up moves and the average size of the down moves.
3. Divide the up average by the down average. That's "relative strength" (RS).
4. Turn it into a number from 0 to 100: RSI = 100 − 100 / (1 + RS).

If almost every recent move was up, RSI gets close to 100. If almost every move was down, it gets close to 0. If ups and downs were about even, it sits near 50.

That's all it is. RSI doesn't know the news or the order book. It's a summary of recent candles.

How traders use it

Overbought and oversold. The classic levels are 70 and 30. Above 70 is called "overbought" and below 30 "oversold". The thinking is that a price that dropped hard and fast might bounce, and one that ran up fast might cool off. This is a mean reversion idea.

Trend strength. Some traders read it the other way. In a strong uptrend, RSI can stay above 70 for a long time. To them, a high RSI means strength, not a reason to sell.

Divergence. Price makes a new low but RSI doesn't. Some read that as selling running out of steam.

Notice that the first two uses point in opposite directions. Same line, two opposite readings. That's a good reason to test, not trust.

An RSI rule, written properly

"Buy when RSI is low" is not a strategy yet. A testable rule needs every part spelled out. Here's a made-up example (not a recommendation):

• Market and timeframe: one coin, daily candles.
• Entry: when the 14-day RSI closes below 30, buy at the close of the next candle.
• Exit: when RSI closes above 50, sell at the close of the next candle.
• Costs: 0.1% fee on every buy and every sell.

Note the "next candle" part. You only know the RSI once the candle closes, so trading at that same close is a classic timing bug. See How to Avoid Look-Ahead Bias. For all the parts a rule needs, see What Is a Quant Trading Strategy?

What happens when you test it on pure noise

We ran a quick simulation to show why RSI results are easy to misread. We made 100 random price series, 365 days each. They're pure noise, so no rule can have a real edge on them. Then we ran the exact rule above, with fees and next-candle fills.

• 59 of 100 series net positive (simulated).
• 140 of 215 trades closed net positive (simulated).

On data with no pattern at all, most trades still won. A "buy the dip, sell the small bounce" rule tends to win often and lose less often but bigger.

Then we did what most people do next: tune it. For each series we tried 25 settings (entry levels 20 to 40, exit levels 50 to 70), picked the best one on the first 250 days, and then ran that setting on the next 115 days it had never seen.

• Best setting, first 250 days: 91 of 100 net positive (simulated).
• Same setting, next 115 days: 33 of 100 net positive (simulated).

The tuning didn't find anything. It found the luckiest setting for each chart. Once new data came in, the results fell apart. We dig into why in The Deflated Sharpe Ratio and Backtest Overfitting: Signs.

RSI isn't useless. But a nice-looking RSI backtest, on its own, tells you very little.

How to test an RSI idea honestly

1. Write the rule down first. Entry, exit, size, fees, timeframe. Lock it before you look at results.

2. Use next-candle fills. Signal on one close, trade on the next. Never trade at the candle that made the signal.

3. Include fees. Small costs add up fast.

4. Count every setting you tried. If you tried 25 versions and kept one, say so: "1 of 25 settings kept." The count is the honest part.

5. Look past the share of winning trades. Check how big the losers are compared to the winners. A rule that wins 7 trades out of 10 can still lose overall.

6. Test it forward. Run the locked rule on new data it has never seen, ideally in paper trading with real fees and timestamps. See Backtest vs Forward Observation.

7. Compare forward to backtest. If the backtest said 4 trades a month and forward shows 1, something is off.

Where AI fits

AI tools like ChatGPT, Claude, Grok, or Cursor can turn "buy oversold dips" into clean rules in seconds, and give you 50 versions just as fast. More versions means more lucky winners. Let the AI write the idea. Let a fixed test decide.

How Pickled Lizard tests ideas like this

Pickled Lizard is an AI trading and research platform. You bring the idea, or your own AI does, and we test it the same way we test everything:

1. Backtest on the full data set we have, with fees and next-candle timing.
2. Paper forward testing on new market data, with real fees and timestamps. The idea shows as "still collecting" while it builds a record.
3. Verified only when its forward results reach 80% of its own backtest trade rate and 80% of its backtest win rate.

Results show on the public paper trading leaderboard. Pickled Lizard isn't an exchange and doesn't hold your money. It's a non-custodial aggregator that connects to your own exchange accounts. Our servers are in Europe. The beta is free, and live trading is off during the beta.

Got an RSI idea? See if it survives new data.

Test it free at pickledlizardz.com/join

Quick recap

• RSI is a 0 to 100 summary of recent up vs down moves. Default 14 candles, classic levels 70/30.
• On pure noise, a simple RSI rule had 140 of 215 trades net positive (simulated). Winning often is not the same as working.
• Tuned settings went from 91 of 100 net positive to 33 of 100 on new data (simulated).
• Write the rule, use next-candle fills, count every try, and test forward.

New here? Start with How to Backtest a Trading Strategy Without Fooling Yourself.