ADX: what it is, and whether it works

1,048 of 5,246 out-of-sample tests beat simply buying and holding the same asset — 20.0%. On the other 4,198 it did not. That is indistinguishable from the 20.1% rate across all 382 indicators we test (one pooled rate over all 660,005 tests we have run, not a mean of the per-indicator rates) — the difference is inside the margin this many tests can resolve, so read it as ordinary, not better or worse.

One test is one variant on one asset on one timeframe. We tested 3 variants of ADX, so an asset that carries several of them contributes several tests — this is the family's record across everything we ran, not a count of assets. Those variants are not alike: individually they beat buy-and-hold between 19.4% and 20.6% of the time, and 20.0% is the pooled rate over all of them, not a property every version of this indicator has. The table below gives each one separately. "Beat" means a higher return than holding that same asset over that same window. Measured out-of-sample, on data the setup was not chosen on.

Directional movement is measured bar to bar: how much the high extends above the previous high (the up move) and how much the low extends below the previous low (the down move). Only the LARGER of the two counts — if the up move is bigger and positive it becomes +DM and -DM is zero, and vice versa; on an inside bar both are zero. Each is smoothed and divided by average true range to give the +DI and -DI lines. ADX is then a smoothed average of the absolute difference between +DI and -DI relative to their sum. The lookback is conventionally 14. Because the difference is taken in absolute terms, ADX is direction-blind by construction.

How it's read. ADX is read as a level: low values are taken to mean the market is ranging and trend-following readings should be distrusted, higher values that a trend is present and worth following. Direction, if wanted, comes from whether +DI is above or below -DI — never from ADX itself.

Where it struggles by design. It is a smoothed average of smoothed averages, so it is slow: by the time it confirms a trend is strong, a large part of the move has usually happened. It also cannot distinguish a strong uptrend from a strong downtrend, which is a frequent misreading.

Origin: Introduced by J. Welles Wilder Jr. in 1978, alongside RSI and ATR.