지수 가중 이동 평균

A simple moving average treats all n observations in the window equally — a price from 30 days ago contributes the same as yesterday's price. But for predicting near-term values, recent observations are more informative than old ones. An Exponentially Weighted Moving Average (EWMA) solves this by assigning exponentially decreasing weights to older observations: yesterday counts the most, the day before counts slightly less, and so on back to the beginning of the series.

EWMA computes each value as a weighted combination of the current observation and the previous EWMA: EWM_t = alpha * x_t + (1 - alpha) * EWM_{t-1}. The alpha (smoothing factor) determines how quickly weights decay — a large alpha (close to 1) responds quickly to recent changes (less smoothing), a small alpha (close to 0) is slow to respond (more smoothing). The weights form a geometric series: the most recent value has weight alpha, the previous has alpha*(1-alpha), and so on.

Pandas implements EWMA through Series.ewm(). The most intuitive parameter is span, which is analogous to a simple moving average window: alpha = 2 / (span + 1). A span of 10 approximately corresponds to a 10-period SMA but with exponentially decaying weights instead of equal weights. Use ewm(span=n).mean() to get the EWMA, and ewm(span=n).std() for exponentially weighted volatility.

The halflife parameter defines how many periods it takes for the weight to decrease by half. If halflife=5, a value from 5 periods ago has half the weight of the current value. The relationship is: alpha = 1 - exp(log(0.5) / halflife). Halflife is often more intuitive than span or alpha because it directly expresses the 'memory' of the average in business terms — for example, 'prices from 2 weeks ago should have half the influence of today's price'.

You can also set the smoothing factor directly with ewm(alpha=). Alpha must be between 0 and 1 (exclusive). A high alpha like 0.8 gives very little smoothing — the EWMA tracks the raw signal closely. A low alpha like 0.1 gives heavy smoothing — the EWMA changes slowly and lags behind sudden shifts. Choosing the right alpha depends on the trade-off between responsiveness (detecting real trend changes quickly) and stability (ignoring random noise).

ewm(span=).std() computes an exponentially weighted standard deviation. This is the basis of the EWMA volatility model used in financial risk management (also known as the RiskMetrics model with decay factor lambda=0.94). Exponentially weighted volatility responds faster to volatility spikes than a simple rolling std because it assigns more weight to recent high-variance observations.

The MACD (Moving Average Convergence Divergence) is a classic financial indicator built from two EWMAs. It is the difference between a fast (short-span) EWMA and a slow (long-span) EWMA of the price. When the MACD crosses zero from below, it signals potential upward momentum (a buy signal); crossing from above signals downward momentum. MACD is a perfect example of applying two EWMAs to derive a derived indicator.

Pandas 1.1+ supports specifying halflife, span, and com as time offsets (e.g. halflife='7D') when working with a DatetimeIndex. This enables proper EWMA on non-uniform time series — for example, if you have hourly data with gaps on weekends, a halflife of '7D' means the weight truly halves every 7 calendar days regardless of the number of rows in that period.

Summary of when to use each:

• Rolling(n).mean(): all n recent observations equally weighted — good for simple trailing metrics and when all recent periods matter equally.
• ewm(span=n).mean(): exponentially weighted — best when you want smoothing that responds more to recent changes, commonly used in finance and signal processing.
• expanding().mean(): all historical data with equal weight — best for YTD or all-time statistics where forgetting old data is not acceptable.

Plotting EWMA and SMA together on a time series chart makes the difference in their responsiveness visible. The EWMA reacts more quickly to sudden level changes because it has not 'forgotten' all previous observations — the short-span EWMA will turn faster than an SMA with the same nominal window. This makes EWMA preferable in forecasting applications where capturing trend direction quickly is more valuable than maximum noise reduction.

Beyond finance, EWMA is widely used in product analytics and operations: customer lifetime value smoothing (recent behaviour predicts future spend better than old behaviour), server latency monitoring (an exponentially smoothed latency metric sounds false alarms faster than a rolling mean), and inventory demand forecasting (recent sales weeks should outweigh data from six months ago). The universal principle is: wherever recency matters, EWMA outperforms simple averaging.

Test your understanding of exponentially weighted moving averages from this lesson.

In this lesson you learned: ewm(span=) applies exponentially decaying weights so recent observations matter more, alpha controls responsiveness (high = fast, low = smooth), and the halflife parameter expresses memory in business-meaningful terms. EWMA outperforms SMA when detecting trend reversals quickly is important. Next up we explore rank and percentile calculations within groups using groupby().rank() and pd.qcut.