Stationarity: A Time Series's Stable Personality

THE MENTAL MODEL: A stationary process is a time series whose fundamental statistical properties don't change over time. Imagine taking a snapshot of the data today and another a year from now. If the process is stationary, the average value (mean) and the typical spread of values (variance) in both snapshots should be roughly the same. The process has a stable, predictable character, not one that is constantly evolving.

HOW IT WORKS: Formally, stationarity means the joint probability distribution of the process is invariant to shifts in time. This is a very strict definition. In practice, analysts usually work with a looser definition called "weak-sense stationarity," which requires just three conditions. First, the mean of the series is constant. Second, the variance is constant. Third, the correlation between two data points depends only on the time lag between them, not on their absolute position in the series. Most statistical procedures aim to achieve this practical form of stationarity.

WHEN TO USE IT: Stationarity is a prerequisite, not a tool you choose to use. It's a property you must check for or induce before applying many classical time series forecasting models, such as ARIMA. These models learn patterns from the past and assume those patterns will continue. If the data has a trend (e.g., a constantly increasing mean), this core assumption is violated, and the model cannot produce a reliable forecast without first transforming the data to remove the trend.

WHEN NOT TO USE IT: You don't need to enforce stationarity for all models. For instance, some machine learning models like tree-based methods or certain neural networks can often handle non-stationary data directly by learning trends and seasonality as features. However, even with these advanced models, transforming the data to be stationary can sometimes improve performance and stability. The key is to know your model's underlying assumptions.

ONE CANONICAL EXAMPLE: Consider the daily stock price of a growing company. The price series likely has an upward trend, so its mean is increasing over time. This is non-stationary. To analyze it with a classical model, you must transform it. A common technique is "differencing," where you replace each price with the difference between it and the previous day's price (the daily return). This new series of daily returns often has a constant mean (hovering near zero) and constant variance, making it stationary and suitable for analysis.