如何在R语言stl函数中生成季节性成分？含趋势成分相关疑问

Hey there! Let's break down your questions about R's stl() function and those behind-the-scenes utilities step by step:

First off, let's clear up a common misconception: the trend component in stl() is NOT derived from differencing. Instead, stl() uses LOESS (locally estimated scatterplot smoothing) to decompose a time series into three parts: seasonal, trend, and remainder. Differencing is a technique to remove trend/stationarize a series, which is separate from how stl() extracts trends.

To generate (extract) the seasonal component, it's straightforward once you run the decomposition:

• The stl() function returns a list object where the $time.series element is a matrix containing all three components. You can directly pull the seasonal column.

Here's a quick example using the built-in AirPassengers dataset:

# Load data and run STL decomposition
data(AirPassengers)
stl_result <- stl(AirPassengers, s.window = "periodic")

# Extract seasonal component
seasonal_component <- stl_result$time.series[, "seasonal"]

# Extract trend component (for reference)
trend_component <- stl_result$time.series[, "trend"]

The s.window = "periodic" argument tells stl() to use a fixed seasonal pattern across all cycles (great for time series with stable seasonality, like monthly airline passenger numbers).

Let's walk through that snippet line by line:

which.cycle <- cycle(x)
z$seasonal <- tapply(z$seasonal, which.cycle, mean)[which.cycle]

• which.cycle <- cycle(x): The cycle() function returns the position of each observation within its cycle. For monthly data, this would be 1-12 (each January gets 1, each February gets 2, etc.).
• z$seasonal <- tapply(z$seasonal, which.cycle, mean)[which.cycle]: This line smooths the seasonal component by averaging values across the same cycle positions.

In plain terms: If you have monthly data, it calculates the average seasonal value for all Januarys, all Februarys, ..., all Decembers. Then it replaces every January's seasonal value with that overall January average, every February's with the February average, and so on. The result is a more consistent, fixed seasonal pattern (instead of the slightly variable seasonal estimates from the initial decomposition).

The functions fdrfourier, backgroundData, ar1analysis, and fourierscore are internal helper functions used by stl()—they're not exported for direct use by users (you'll get an error if you try calling them directly unless you use stats:::function_name to access private package functions). Here's a quick breakdown of their roles:

• fdrfourier: Assists in fitting Fourier series to the seasonal component. It helps identify the right number of Fourier terms to capture the seasonal pattern's frequency characteristics, laying the groundwork for the LOESS smoothing step.
• backgroundData: This is either an internal data object or function that stores/handles temporary parameters and intermediate calculations during decomposition (like smoothing window sizes or intermediate statistics). It's purely for the algorithm's internal bookkeeping.
• ar1analysis: Runs AR(1) analysis on the decomposition's residual series. It checks for autocorrelation in the residuals—if residuals have no significant autocorrelation, it means the trend and seasonal components have captured most of the signal in the original series.
• fourierscore: Calculates a "score" for the Fourier fit of the seasonal component. This score helps the stl() algorithm choose the optimal Fourier terms to ensure the seasonal component is accurately estimated.

In short, these are all behind-the-scenes tools that stl() uses automatically—you don't need to interact with them directly for standard decomposition tasks.

内容的提问来源于stack exchange，提问作者h-y-jp