ADX is the odd one out among common indicators because it deliberately discards direction. Wilder built it in 1978 to answer a prior question: is this security trending at all? Applying a trend-following rule to a stock that is chopping sideways is the fastest way to bleed on costs, and ADX is the filter that was meant to prevent it. The call is @ADX(high, low, close, 14).
How it is computed
Wilder first derives two directional movement series from consecutive bars:
+DM = current_high - prev_high when that exceeds the downward move, else 0
-DM = prev_low - current_low when that exceeds the upward move, else 0
Each is smoothed and divided by ATR to give +DI and -DI. Then:
DX = 100 * abs(+DI - -DI) / (+DI + -DI)
ADX = smoothed average of DX over N bars
The absolute value in DX is where direction is thrown away. A strong uptrend and a strong downtrend of equal force produce the same ADX. The output is unitless and compares to a threshold:
@ADX(high, low, close, 14) > 25
Typical readings
| ADX | Conventional description |
|---|---|
| Below 20 | no trend, range-bound |
| 20 to 25 | ambiguous |
| 25 to 50 | trending |
| Above 50 | strong trend |
Wilder’s 14-period window and the 20 and 25 lines are his own conventions from commodity futures. They were not calibrated on NSE equities, and treating them as thresholds with statistical meaning is a mistake.
How to read the output
The honest failure mode is baked into the definition: ADX measures trend strength without direction. An ADX of 40 on RELIANCE tells you the stock is moving decisively and nothing about which way. It is a gate, not a signal. Pairing it with a directional rule such as close > @SMA(close, 200) in a separate pipeline step is how it is meant to be used.
Risks and caveats
ADX is doubly smoothed, once into +DI and -DI and again into the final average. It is therefore slow. By the time it clears 25, a large part of the move has usually happened.
It also drifts down during a healthy trend that pauses. A stock in a genuine multi-month uptrend that consolidates for three weeks will see ADX fall below the threshold, so a rule gated on ADX exits perfectly good trends during ordinary consolidation and re-enters higher.
Both the window and the threshold are free parameters over one history. Sweeping ADX from 15 to 40 across a dozen windows and keeping the best pair is overfitting, and the effect is easy to miss because ADX is used as a filter rather than the headline signal.