RSI, measured: what a momentum oscillator actually predicts

The mechanism, without the formula

Split the last N bars into up-moves and down-moves. Add up the size of the up-moves, add up the size of the down-moves, and take the ratio. If gains dominate, the ratio is large and the index reads high; if losses dominate, it reads low. The standard period is 14 bars, and the compression into 0–100 is what makes it comparable across markets.

Two consequences follow immediately. First, RSI is a summary of what already happened. Second, the period you choose is the entire personality of the indicator: short periods produce a jumpy line that crosses thresholds constantly, long periods produce a smooth line that rarely reaches extremes.

How RSI is actually used

• As a level: above 70 'overbought', below 30 'oversold'. Simple, popular, and on its own worthless — in a strong trend the index can stay above 70 for weeks.
• As a midline cross: above 50 read as bullish momentum, below 50 as bearish.
• As a divergence: price makes a new high, RSI does not. This is the most interesting use and the hardest to test, because 'a new high' needs a precise definition.
• As a failure swing: RSI pushes above a previous peak, pulls back, and fails to exceed it.

The measurement

To find out whether the indicator carries information, treat it as a factor: take its value at each bar, shift it forward by one bar so nothing from the future leaks in, and compute the Spearman rank correlation against the next bar's return. We run this on gold hourly bars through the same engine that measures every factor in the platform.

What a small IC actually means

Rank ICs for single indicators on liquid markets are small — a few percent, not a few tenths. This disappoints people who expect a strong number, but it is the honest shape of the problem. If a single publicly known oscillator predicted the next bar strongly, it would be arbitraged away, because anyone can compute it.

A small IC is not the same as useless. It means the indicator shifts the odds slightly, and that a strategy built on it needs several things in its favour: low costs, a sensible holding period, combination with other signals, and position sizing that survives the stretches where it contributes nothing.

The mistakes that cost the most

• Treating 30 and 70 as entry signals. They are thresholds on a derived series; the market does not know they exist.
• Choosing the period after seeing which one looked best. That is fitting, and it will not survive.
• Assuming it transfers across markets. RSI on gold and RSI on a currency pair behave differently because the underlying volatility structure differs.
• Ignoring that RSI is bounded. Near 0 or 100 the indicator compresses, and small price changes produce large readings — an artefact of the scale, not a signal.

How to test it yourself

Open the MT5 Strategy Tester with your own broker's spread, add commission, and add a slippage estimate. Build the simplest possible rule from RSI — for example, hold long while RSI is above 50 and flat otherwise — because a simple rule is much easier to interpret than a complicated one. Then compare it against always being long, which is the baseline most RSI strategies quietly fail to beat.

Change the period from 14 to 10 and to 20, and watch how much the result moves. If the answer depends heavily on the exact value, you were measuring noise.