What it measures

The relative strength index measures how one-sided recent price changes have been. It takes the average size of up-moves and the average size of down-moves over a window and turns their ratio into a number between 0 and 100. Near 100 the window was almost all gains; near 0 almost all losses; 50 means balance. It is the most used oscillator in the strategies the desk screened.

The calculation

RS = average gain / average loss, both Wilder-smoothed over length bars RSI = 100 − 100 / (1 + RS)

The averages use Wilder's smoothing (an RMA), which is slower than it looks: a 14-period RSI responds roughly like a 27-period EMA would. Gains and losses are measured close to close.

Inputs and defaults

  • Length: 14, Wilder's original. Short lengths (2 to 5) make a twitchy oscillator used in mean-reversion rules; long lengths (20 and up) make a slow trend gauge.
  • Thresholds: 70 and 30 by convention; 80 and 20 for stricter rules; 50 as a trend line.
  • Source: close.

How it is read

Three different ways, which is the source of most confusion. As a reversion signal: above 70 is "overbought", sell; below 30 is "oversold", buy. As a trend filter: above 50 means buyers have the upper hand, trade long only. As divergence: price makes a new high and RSI does not, so the move is tiring. The first and the second contradict each other, and a strategy has to pick one.

Where it fails

In a trend the RSI goes to 80 and stays there while price keeps rising; every "overbought" sale is a short against a move that is not finished. The reversion reading only works in a range, and the indicator has no way of telling you which regime you are in. The thresholds are the whole strategy and are the first thing anyone tunes, so an RSI rule that looks sharp on one window is often the thresholds fitted to that window.

How the desk reads it

The desk reads an RSI rule first for which of the three readings it codes, reversion, trend filter or divergence, because they contradict each other, and then on how far the thresholds can move before the result changes. Across the screen, rules that sold overbought on their own tended to fail in the trending year.

In the desk's screen

Of the 614 screened strategies in the library, 189 (30.8%) use this indicator somewhere in their logic.

71 of those were gross-profitable on enough cells to reach the after-fee replay.

The count is read from the compiled code of each strategy in the library; a strategy that computes the indicator by hand under another name is not counted. The desk's own rules are not part of the count.