RSI (relative strength index)

Definition

RSI (Wilder 1978) is a bounded 0-100 oscillator measuring the ratio of average up-moves to average down-moves over a window (14 by convention — the platform’s rsi_14). High readings mean recent gains dominated; low readings mean losses dominated. The folklore labels — “overbought” above 70, “oversold” below 30 — are conventions Wilder proposed, not measured thresholds.

How it works / structure

• Formula: RS = (average gain over N) / (average loss over N), smoothed by Wilder’s method (an EMA variant with α = 1/N); RSI = 100 − 100/(1+RS).
• Parameters (engine-executable): window N, smoothing convention (Wilder vs standard EMA — values differ; the engine pins Wilder), signal definition — threshold cross (30/70 or fitted), threshold-and-return (re-cross back through the band), divergence (mechanically pinned), or midline (50) regime.
• Two contradictory readings: mean-reversion practice buys low RSI (strategy-mean-reversion); momentum practice reads RSI > 70 as strength confirmation (strong trends SUSTAIN high RSI — the “overbought can stay overbought” reality). Which reading applies is a regime question the indicator itself cannot answer.

When it applies

Short-horizon reversion triggers in range regimes (with the event exclusions of strategy-mean-reversion); pullback qualification within uptrends (strategy-swing-trading); normalized cross-instrument momentum comparison (bounded scale travels across instruments, unlike MACD).

Risk profile & failure modes

• Trend blindness: in a sustained trend RSI pins high (or low) for weeks; fading it is the falling-knife failure with an indicator’s endorsement.
• Threshold folklore: 30/70 have no demonstrated universal edge; fitted thresholds inherit overfitting risk — either way the number needs replay support, not tradition.
• Window sensitivity: RSI(2) and RSI(14) are different instruments (very short RSIs are reversion-tuned and friction-intense).
• Divergence subjectivity: same discipline as MACD — pinned mechanical definition or unfalsifiable.

Evidence & limits

Wilder (1978) is the construction source. Oscillator families show mixed, cost-fragile results in the surveyed literature (Park-Irwin 2007); the short-horizon reversal effect RSI proxies is documented (strategy-mean-reversion citations) but simple public RSI rules capture it inconsistently after friction. RSI level lore (“30 means bounce”) is folklore, labeled as such.

Falsifiable-thesis examples

Illustrations only, not signals:

• “X, with RSI(14) < 25 and no news event, will close above today’s close within 10 sessions” — falsified at the mark.
• “RSI(2) < 10 entries on liquid ETFs, exited at RSI(2) > 60, are net profitable after friction this year in replay” — falsified by replay P&L.

Cross-references

• Sibling oscillators: indicator-stochastics, indicator-roc
• The two effects it straddles: strategy-mean-reversion, strategy-momentum
• Platform binding: rsi_14
• Method caveats: lens-technical