Feature engineering in time series analysis involves transforming raw time-stamped data into meaningful inputs for machine learning models. Since time series data is sequential and time-dependent, the goal is to create features that capture temporal patterns like trends, seasonality, and autocorrelation. This process helps models understand relationships between past and future observations. For example, predicting daily sales might require features that reflect weekly cycles or holiday effects. Unlike tabular data, time series features often rely on sliding windows, lagged values, or aggregated statistics over time intervals to encode temporal context effectively.
Common techniques include creating lag features (using past values as predictors), rolling statistics (e.g., moving averages), and datetime-based features (e.g., hour of day). Lag features, such as sales from the previous 7 days, directly model how past behavior influences future outcomes. Rolling windows calculate metrics like mean or standard deviation over a fixed period (e.g., a 30-day average) to smooth noise and highlight trends. Datetime features break down timestamps into components like month, day of week, or holidays to account for recurring patterns. Additionally, decomposition methods (e.g., separating a series into trend, seasonality, and residuals) or Fourier transforms can isolate periodic signals. For instance, decomposing hourly temperature data might reveal daily and yearly cycles, which can then be used as separate features.
Key considerations include handling non-stationarity (where statistical properties change over time) and avoiding data leakage (using future data to create features). Differencing—subtracting the previous value from the current one—is a common fix for non-stationarity. For leakage prevention, features like rolling averages must be computed using only historical data up to the prediction point. Domain knowledge also plays a role: a retail model might include features for promotions or local events. Finally, validation must respect time order—splitting data into sequential train/test sets rather than random splits. For example, when forecasting energy demand, using a lagged feature of the previous day’s demand while ensuring no future data is included in training ensures the model generalizes to unseen time periods.