Exponential moving average (EMA)

Definition

The EMA is a recursively weighted average that gives recent prices exponentially more weight than old ones — the standard lower-lag alternative to the SMA, and the building block of MACD (indicator-macd). Same job as the SMA (trend smoothing), different lag/noise trade-off.

How it works / structure

• Formula: EMA_t = α·P_t + (1−α)·EMA_{t−1}, with α = 2/(N+1) for the “N-period” convention.
• Parameters (engine-executable): N (or α directly), price input, initialization (seed with SMA(N) — early values differ by seed, a replay-reproducibility detail the engine pins), and derived signals (price-vs-EMA, EMA slope, fast/slow EMA crossovers — 12/26 being MACD’s pair).
• Properties vs SMA: responds faster to new prices (less lag), never fully forgets old ones (infinite memory, geometrically decaying), no drop-off artifact — a shock decays smoothly instead of exiting abruptly.
• Equivalence caveat: EMA(N) and SMA(N) are not comparable at equal N; matching effective lag requires different windows — parameter conversions matter when porting rules.

When it applies

Anywhere the SMA applies with a preference for responsiveness: faster trend-state flips, crossover systems on shorter horizons, MACD construction. The choice between SMA and EMA is a lag-vs-noise parameter decision the engine treats as tunable, not a doctrine.

Risk profile & failure modes

• Faster also means noisier: the EMA’s responsiveness converts to more whipsaw signals in ranges — it moves the trade-off, it does not escape it.
• Seed sensitivity in short replays: early-window EMA values depend on initialization; short backtests inherit seed bias.
• False precision folklore: claims that specific EMA windows (8/21, 9/13) carry special power are uncited lore; window choice is a fitted parameter with all of lens-quantitative’s overfitting caveats.

Evidence & limits

The evidence base is the same MA-rule literature as the SMA (Brock-Lakonishok-LeBaron 1992; Sullivan et al 1999’s snooping correction) — EMAs were among the tested family; no study establishes EMA superiority over SMA as a class. Trend-state content is real (time-series momentum); crossover timing edges are weak-to-unproven after costs.

Falsifiable-thesis examples

Illustrations only, not signals:

• “X, with its 21-day EMA above its 50-day EMA, will outperform its sector over the next month” — falsified by the pair’s returns.
• “Replacing SMA with lag-matched EMA in system S improves replay Sharpe this year” — falsified by the paired replay.

Cross-references

• Arithmetic sibling: indicator-sma; principal consumer: indicator-macd
• Effect proxied: strategy-momentum
• Platform binding: trend_state
• Method caveats: lens-technical, lens-quantitative