ADX (average directional index)

Definition

ADX (Wilder 1978) measures trend STRENGTH without direction: it compares upward and downward directional movement (+DI/−DI), then smooths their normalized difference into a 0-100 index. High ADX = strong directional movement either way; low ADX = range. Its platform role is regime arbitration — the switch that decides whether trend rules or reversion rules apply (trend_state composites use it).

How it works / structure

• Construction: +DM = today’s high minus yesterday’s (when positive and dominant), −DM the mirror; each smoothed and divided by ATR to give +DI/−DI; DX = 100 × |+DI − −DI| / (+DI + −DI); ADX = Wilder-smoothed DX (14 convention).
• Parameters (engine-executable): window, threshold convention (ADX > 25 “trending”, < 20 “ranging” — Wilder’s suggestions, treated as tunable), and the signal forms — ADX level (regime filter), ADX slope (strengthening/fading), DI crossover (a directional signal, distinct from the ADX level itself).
• Key property: ADX rises in strong DOWNTRENDS too — reading it as bullishness is the classic misuse; direction lives in the DI pair, strength in ADX.

When it applies

Regime filtering: gate trend entries on ADX rising/high (strategy-futures-trend-following MA variants), gate reversion entries on ADX low (strategy-mean-reversion regime filter), suppress breakout systems in ADX collapse. As a direction signal (DI crossovers) it is weaker and noisier — platform treats that use as secondary.

Risk profile & failure modes

• Double-smoothed lag: ADX is a smoothing of a smoothing — it confirms trends late and announces their death later; as a gate it always costs the move’s first leg.
• Threshold folklore: 20/25 boundaries are Wilder’s suggestions from 1970s commodity charts, not measured constants; fitted thresholds inherit overfitting risk.
• Whipsaw at the gate: ADX oscillating around the threshold flips the regime switch repeatedly — hysteresis (separate on/off levels) is the standard fix and one more parameter.
• Range-compression blindness: low ADX says “no trend lately,” not “breakout imminent” — pairing with compression gauges (indicator-bollinger-bands bandwidth) is a separate hypothesis.

Evidence & limits

Wilder (1978) is the construction source. Direct academic tests of ADX rules are sparse; it enters the surveyed technical literature (Park-Irwin 2007) within directional-system families with the usual mixed, cost-fragile results. The regime-gating use — its platform role — is testable per strategy: does the gate improve the gated system’s replay? That paired test is the evidence that matters here.

Falsifiable-thesis examples

Illustrations only, not signals:

• “Gating system S’s entries on ADX(14) > 25 improves its replay Sharpe vs ungated this year” — falsified by the paired replay.
• “X’s ADX rising through 25 this week will be followed by higher 20-day realized trend persistence (|ROC(20)| above its median)” — falsified by the measurement.

Cross-references

• Its denominator: indicator-atr; its arbitration clients: strategy-futures-trend-following, strategy-mean-reversion, strategy-breakout
• Regime framework: regime-volatility, trend_state
• Method caveats: lens-technical