Exponential Moving Average

Time Series

Easy

Problem

Theory

An exponential moving average (EMA) is a weighted average that gives more weight to recent observations. Weights decrease exponentially as observations get older.

\text{EMA}_t = \alpha y_t + (1-\alpha) \text{EMA}_{t-1}

where 0 < \alpha < 1 is the smoothing parameter.

The Formula

For a time series y_1, y_2, ..., y_T:

\text{EMA}_t = \alpha y_t + (1-\alpha) \text{EMA}_{t-1}

Recursive form: Current EMA combines current observation with previous EMA.

Expanded form:

\text{EMA}_t = \alpha \sum_{i=0}^{t-1} (1-\alpha)^i y_{t-i}

Weight on observation i periods ago: \alpha(1-\alpha)^i

Weights decay exponentially with age.

Smoothing Parameter

Alpha (\alpha):

Controls responsiveness to new data.

High \alpha (close to 1):

• More weight on recent observations
• Fast adaptation to changes
• Less smoothing, more noise

Low \alpha (close to 0):

• More weight on historical observations
• Slow adaptation, smoother curve
• More smoothing, less noise

Typical range: 0.1 \leq \alpha \leq 0.3

Alpha and Window Size Equivalence

EMA with parameter \alpha approximates SMA with window:

w \approx \frac{2}{\alpha} - 1

Examples:

• \alpha = 0.1: w \approx 19
• \alpha = 0.2: w \approx 9
• \alpha = 0.5: w \approx 3

Interpretation: \alpha = 0.2 gives similar smoothing to 9-period SMA.

Initialization

First value (\text{EMA}_1):

Common: \text{EMA}_1 = y_1

Alternative: \text{EMA}_1 = \text{average of first few observations}

Impact: Initial choice matters for first few periods, then influence fades exponentially.

After n periods: Weight on initial value is (1-\alpha)^n

Example: With \alpha=0.2, after 20 periods weight is (0.8)^{20} = 0.0115 (1.15%).

Worked Example

Data: [100, 105, 103, 108, 110]

Parameter: \alpha = 0.3

Initialization: \text{EMA}_1 = 100

Period 2:

\text{EMA}_2 = 0.3(105) + 0.7(100) = 31.5 + 70 = 101.5

Period 3:

\text{EMA}_3 = 0.3(103) + 0.7(101.5) = 30.9 + 71.05 = 101.95

Period 4:

\text{EMA}_4 = 0.3(108) + 0.7(101.95) = 32.4 + 71.365 = 103.765

Period 5:

\text{EMA}_5 = 0.3(110) + 0.7(103.765) = 33 + 72.636 = 105.636

Result: [100, 101.5, 101.95, 103.765, 105.636]

Weight Distribution

Weight on current observation: \alpha

Weight on observation 1 period ago: \alpha(1-\alpha)

Weight on observation 2 periods ago: \alpha(1-\alpha)^2

Weight on observation k periods ago: \alpha(1-\alpha)^k

Sum of all weights: \alpha \sum_{k=0}^{\infty} (1-\alpha)^k = \alpha \cdot \frac{1}{1-(1-\alpha)} = 1

Weights sum to 1 (proper average).

Comparison to Simple Moving Average

SMA weights: \frac{1}{w} for last w observations, 0 for older.

EMA weights: Exponentially decaying, never exactly zero.

SMA: Step function in weights.

EMA: Smooth decay in weights.

Advantage SMA: Clear interpretation (average of last w values).

Advantage EMA: All history considered, continuous weighting, faster computation.

Computational Efficiency

SMA: Requires storing last w values. O(w) per update.

EMA: Only requires previous EMA. O(1) per update.

Memory: EMA needs constant memory (just previous value), SMA needs O(w) memory.

Real-time systems: EMA preferable for low memory footprint and fast updates.

Multiple EMAs

Fast and slow EMAs:

• Fast: Small window (high \alpha), e.g., \alpha = 0.2 (10-period equivalent)
• Slow: Large window (low \alpha), e.g., \alpha = 0.05 (40-period equivalent)

Golden cross: Fast EMA crosses above slow EMA (bullish signal).

Death cross: Fast EMA crosses below slow EMA (bearish signal).

Application: Trading strategies, trend confirmation.

MACD Indicator

Moving Average Convergence Divergence:

\text{MACD} = \text{EMA}_{12} - \text{EMA}_{26}

Signal line:

\text{Signal} = \text{EMA}_9(\text{MACD})

Histogram:

\text{Histogram} = \text{MACD} - \text{Signal}

Interpretation:

• MACD crosses above signal: Buy signal
• MACD crosses below signal: Sell signal
• Histogram shows momentum strength

Lag Properties

Lag in EMA:

\text{Lag} \approx \frac{1-\alpha}{\alpha}

Example: \alpha = 0.2 gives lag of \frac{0.8}{0.2} = 4 periods.

Comparison to SMA lag: \frac{w-1}{2}

For SMA with w=9: lag = 4 periods (same as EMA with \alpha=0.2).

Implication: EMA and equivalent SMA have similar lag characteristics.

Forecasting with EMA

One-step ahead forecast:

\hat{y}_{t+1} = \text{EMA}_t

Multi-step ahead:

\hat{y}_{t+h} = \text{EMA}_t \text{ for all } h > 0

Flat forecast: EMA assumes no trend, predicts constant future.

Use case: Appropriate for stationary series with no trend.

Limitation: Poor for trending data. Use double exponential smoothing instead.

Connection to Simple Exponential Smoothing

EMA is equivalent to simple exponential smoothing (SES):

SES forecast:

\hat{y}_{t+1} = \alpha y_t + (1-\alpha)\hat{y}_t

Substituting \hat{y}_t = \text{EMA}_{t-1}:

\hat{y}_{t+1} = \alpha y_t + (1-\alpha)\text{EMA}_{t-1} = \text{EMA}_t

Interpretation: EMA is both a smoothing technique and a forecasting method.

Optimal Alpha Selection

Minimize forecast error:

\alpha^* = \arg\min_{\alpha} \sum_{t=1}^{T} (y_t - \text{EMA}_{t-1})^2

Grid search: Try \alpha \in \{0.01, 0.02, ..., 0.99\}, select best.

Closed-form: No analytical solution. Numerical optimization required.

Cross-validation: Split data, optimize on training set, validate on test set.

Typical result: \alpha \in [0.1, 0.3] for most applications.

Adaptive Exponential Smoothing

Problem: Fixed \alpha may not suit all periods.

Solution: Let \alpha vary over time.

Example adaptive rule:

\alpha_t = |\frac{e_t}{y_t}|

where e_t = y_t - \text{EMA}_{t-1}

Interpretation: Increase \alpha when forecast errors are large (adapt quickly).

Constraints: Bound \alpha_t to [\alpha_{\min}, \alpha_{\max}] to prevent instability.

Centered vs Trailing EMA

Trailing EMA (standard):

\text{EMA}_t = \alpha y_t + (1-\alpha)\text{EMA}_{t-1}

Uses only past values (causal).

Centered EMA:

Apply EMA forward and backward, then average.

\text{EMA}_{\text{centered},t} = \frac{\text{EMA}_{\text{forward},t} + \text{EMA}_{\text{backward},t}}{2}

Advantage: No lag, better smoothing.

Disadvantage: Requires future values (non-causal). Cannot forecast.

Use: Historical analysis, not real-time forecasting.

Variance of EMA

For white noise input with variance \sigma^2:

\text{Var}(\text{EMA}_t) = \frac{\alpha}{2-\alpha} \sigma^2

Higher \alpha: Higher variance (less smoothing).

Lower \alpha: Lower variance (more smoothing).

Variance reduction factor:

\frac{\text{Var}(\text{EMA}_t)}{\text{Var}(y_t)} = \frac{\alpha}{2-\alpha}

Example: \alpha = 0.2 gives variance reduction to \frac{0.2}{1.8} = 0.111 (11.1% of original).

Signal-to-Noise Ratio

Goal: Maximize signal retention while minimizing noise.

Trade-off: Low \alpha removes more noise but also removes signal.

Optimal choice depends on:

• Signal frequency
• Noise characteristics
• Application requirements

High SNR data: Use higher \alpha (less smoothing needed).

Low SNR data: Use lower \alpha (more smoothing needed).

EMA for Trend Detection

EMA slope:

\text{Slope}_t = \text{EMA}_t - \text{EMA}_{t-1} = \alpha(y_t - \text{EMA}_{t-1})

Positive slope: Upward trend.

Negative slope: Downward trend.

Acceleration:

\text{Acceleration}_t = \text{Slope}_t - \text{Slope}_{t-1}

Monitoring: Track slope changes for early trend reversal detection.

Bollinger Bands with EMA

Middle band: EMA of price.

Upper band: \text{EMA}_t + k \cdot \sigma_t

Lower band: \text{EMA}_t - k \cdot \sigma_t

where \sigma_t is rolling standard deviation and k=2 typically.

Interpretation:

• Price near upper band: Overbought
• Price near lower band: Oversold
• Squeeze (bands narrow): Low volatility, potential breakout

Exponential Weighting in Other Contexts

Exponentially weighted variance:

\sigma_t^2 = \lambda \sigma_{t-1}^2 + (1-\lambda) r_t^2

Used in GARCH models for volatility.

Exponentially weighted covariance:

\text{Cov}_t(x, y) = \lambda \text{Cov}_{t-1}(x, y) + (1-\lambda) x_t y_t

RiskMetrics: \lambda = 0.94 for daily data.

Principle: Recent data more relevant for current risk.

Double Exponential Smoothing Extension

Problem: EMA assumes no trend.

Solution: Add trend component (Holt's method).

\ell_t = \alpha y_t + (1-\alpha)(\ell_{t-1} + b_{t-1})

b_t = \beta(\ell_t - \ell_{t-1}) + (1-\beta)b_{t-1}

Forecast:

\hat{y}_{t+h} = \ell_t + h \cdot b_t

Linear trend instead of flat.

Triple Exponential Smoothing

Holt-Winters method:

Adds seasonal component to level and trend.

Three equations:

1. Level: \ell_t = \alpha(y_t - s_{t-m}) + (1-\alpha)(\ell_{t-1} + b_{t-1})
2. Trend: b_t = \beta(\ell_t - \ell_{t-1}) + (1-\beta)b_{t-1}
3. Seasonal: s_t = \gamma(y_t - \ell_t) + (1-\gamma)s_{t-m}

Forecast:

\hat{y}_{t+h} = \ell_t + hb_t + s_{t+h-m}

Handles trend and seasonality.

Residual Analysis

One-step forecast errors:

e_t = y_t - \text{EMA}_{t-1}

Good fit: Residuals should be white noise.

Diagnostic plots:

1. Time plot of residuals (no patterns)
2. ACF of residuals (no significant autocorrelation)
3. Histogram of residuals (approximately normal)

Ljung-Box test: Test for autocorrelation in residuals.

Heteroscedasticity: Check if residual variance is constant.

EMA in Volatility Modeling

EWMA volatility:

\sigma_t^2 = \lambda \sigma_{t-1}^2 + (1-\lambda) r_t^2

where r_t is return at time t.

RiskMetrics approach: \lambda = 0.94 for daily returns.

Interpretation: Variance forecast adapts to recent squared returns.

Application: Value at Risk (VaR) calculations, portfolio risk management.

Filter Interpretation

EMA is a discrete-time first-order IIR (infinite impulse response) filter:

Transfer function:

H(z) = \frac{\alpha}{1 - (1-\alpha)z^{-1}}

Frequency response:

Low-pass filter. Attenuates high-frequency components (noise).

Cutoff frequency: Depends on \alpha. Higher \alpha has higher cutoff (less filtering).

Outlier Robustness

Problem: Single outlier disproportionately affects EMA.

Example: Outlier y_t = 1000 when typical values are 100.

\text{EMA}_t = 0.2(1000) + 0.8(100) = 280

Large jump. Persists for several periods.

Robust alternatives:

1. Median-based smoothing
2. Trimmed mean exponential smoothing
3. Winsorization before applying EMA

Trade-off: Robustness vs sensitivity to legitimate changes.

EMA for Time Series Decomposition

Trend extraction:

\text{Trend}_t = \text{EMA}_t

Detrended series:

\text{Detrended}_t = y_t - \text{EMA}_t

Seasonal component: Apply EMA to detrended series at seasonal lag.

Residual:

\text{Residual}_t = y_t - \text{Trend}_t - \text{Seasonal}_t

Application: Decompose into components for separate analysis.

Practical Considerations

Choice of \alpha:

• Financial trading: \alpha \in [0.1, 0.2] (smooth trends)
• Web analytics: \alpha \in [0.2, 0.4] (responsive to changes)
• Sensor data: \alpha \in [0.05, 0.15] (heavy smoothing)

Initialization: Use first observation or average of first 5-10 observations.

Outliers: Consider robust preprocessing before applying EMA.

Trend: If data has strong trend, use double exponential smoothing instead.

Seasonality: If seasonal patterns exist, use Holt-Winters method.

Examples

Hints

1. Initialize the output with the first input value.
2. Append alpha times the current value plus one minus alpha times the previous EMA.

Requirements

• Initialize EMA[0] with the first value in the series
• Apply the recursive EMA formula for each subsequent value
• Return a list of the same length as the input

Constraints

• values has at least 1 element
• 0 < alpha <= 1
• Return a list with the same length as values
• Time limit: 300 ms

Starter Code

def exponential_moving_average(values: list, alpha: float) -> list:
    """
    Returns the exponential moving average at every position.
    """
    # Write code here
    pass

Test Cases