Moving Average

Moving Average in Time Series: SMA, EMA, WMA and MA(q)

Time-series data often moves up and down from one period to the next. A moving average helps us smooth some of those short-term fluctuations so that the broader pattern is easier to see.

For example, suppose daily website visits are 10, 20, 30, 40, 50. A 3-day moving average first uses 10, 20 and 30, then moves one step and uses 20, 30 and 40, then 30, 40 and 50.

What Is a Simple Moving Average?

A Simple Moving Average (SMA) calculates the arithmetic mean of observations inside a fixed-size rolling window. With a 3-period trailing SMA, every value in the most recent three observations receives equal weight.

Why does a moving average smooth data?

A single unusual observation affects only part of the window instead of becoming the entire displayed value. Averaging therefore reduces some short-term variation, although it can also make the smoothed series react more slowly to genuine changes.

Choosing the Moving-Average Window

The window size controls how much smoothing you get. There is no universally best window.

Domain structure can also matter. For daily data, a 7-day window may be useful when weekly behavior is relevant; for monthly data, 12 periods may be meaningful when yearly seasonality is important. These are examples, not automatic rules.

SMA, WMA, EMA and Cumulative Average

1. Simple Moving Average (SMA)

Every observation inside the fixed window receives equal weight. SMA is easy to understand and is useful for descriptive smoothing, but larger windows can respond slowly to new changes.

2. Weighted Moving Average (WMA)

A WMA assigns chosen weights to observations. The weights do not have to decrease linearly; the weighting scheme depends on how you define it.

3. Exponential Moving Average (EMA)

EMA gives more weight to recent observations while older observations retain progressively smaller influence. A common recursive form is:

For span = 3:

α = 2 / (3 + 1) = 0.5

Start with EMA₁ = 10. For the next value 20:

EMA₂ = 0.5×20 + 0.5×10 = 15

4. Cumulative / Expanding Average

A cumulative average uses all observations seen so far, so its window keeps expanding rather than staying fixed. It is useful for quantities such as a running lifetime average, but it behaves differently from a fixed-window SMA.

Smoothing Is Not Automatically Forecasting

A trailing 3-period SMA calculated at time t usually includes the observation at time t. That makes it a useful smoother, but it is not automatically a prediction that existed before time t was observed.

Moving-Average Model MA(q): A Different Concept

In ARMA and ARIMA, the MA(q) part does not mean averaging the latest observations. Instead, the current process value is represented using the current random shock and a finite number of past shocks/errors.

Worked MA(2) forecasting example

Suppose the fitted model has c = 50, θ₁ = 0.8, θ₂ = 0.3, with previous estimated innovations 4 and −2. For a one-step forecast, the future innovation has expectation zero, so:

The key point is that this MA(2) calculation uses past innovations/errors. It is structurally different from taking a 2-period or 3-period arithmetic moving average of observed values.

Going Deeper: MA(q), Stationarity, White Noise and ACF

Stationarity

Classical MA(q) is a stationary time-series model. In practical weak-stationarity terms, we look for stable mean and variance and an autocovariance structure that depends on lag rather than calendar time. This is different from saying that the series must stay inside fixed numerical bounds.

White noise / innovations

White noise is commonly modeled with mean zero, constant finite variance and no serial correlation. After fitting a time-series model, residual diagnostics check whether meaningful autocorrelation remains; residuals behaving approximately like white noise are a good sign that the model has captured much of the predictable linear structure.

ACF and the order q

For an ideal pure MA(q) process, the autocorrelation function has a characteristic cutoff after lag q. In real datasets, an ACF plot can therefore provide clues about candidate MA orders, but it should not be treated as an automatic rule that always reveals the “optimal q.” Mixed ARMA/ARIMA structure, finite samples and noise make identification less clean.

Python: SMA, EMA and a Time-Safe One-Step Moving Average

import pandas as pd

# Sequential time-series observations
df = pd.DataFrame({
    "Value": [10, 20, 30, 40, 50, 60, 70]
})

# 3-period trailing Simple Moving Average (smoother)
df["SMA_3"] = df["Value"].rolling(window=3).mean()

# Recursive EMA matching: EMA_t = alpha*x_t + (1-alpha)*EMA_(t-1)
df["EMA_3"] = df["Value"].ewm(span=3, adjust=False).mean()

# One-step-ahead 3-period moving-average forecast.
# shift(1) ensures the value at time t is not used to predict time t.
df["MA_3_Forecast"] = (
    df["Value"]
    .shift(1)
    .rolling(window=3)
    .mean()
)

print(df.round(2).to_string(index=False))

next_forecast = df["Value"].tail(3).mean()
print("Next-period 3-value forecast:", round(next_forecast, 2))

 Value  SMA_3  EMA_3  MA_3_Forecast
    10    NaN  10.00            NaN
    20    NaN  15.00            NaN
    30   20.0  22.50            NaN
    40   30.0  31.25           20.0
    50   40.0  40.62           30.0
    60   50.0  50.31           40.0
    70   60.0  60.16           50.0
Next-period 3-value forecast: 60.0

Common Mistakes