Time series and regression methods for univariate environmental forecasting: An empirical evaluation

Area of Science:

• Environmental Science
• Data Science
• Statistical Modeling

Background:

• Accurate environmental forecasting is crucial for human well-being.
• The comparative performance of time series versus regression methods for univariate environmental forecasting remains unclear.
• Evaluating numerous methods across diverse environmental variables and frequencies is needed.

Purpose of the Study:

• To determine the superior method for univariate time series forecasting in environmental applications.
• To compare the performance of conventional time series models against various regression techniques.
• To assess forecasting accuracy across multiple frequencies (hourly, daily, monthly) and horizons (1-12 steps).

Main Methods:

• Conducted a large-scale comparative evaluation of 68 environmental variables.
• Included six statistical time series methods (e.g., ARIMA, Theta) and fourteen regression methods (e.g., Huber, Random Forest, Gradient Boosting Machines).
• Evaluated forecast performance for one to twelve steps into the future across hourly, daily, and monthly frequencies.

Main Results:

• While strong time series methods (ARIMA, Theta) showed high accuracy, several regression methods (Huber, Extra Trees, Random Forest, Light Gradient Boosting Machines, Gradient Boosting Machines, Ridge, Bayesian Ridge) demonstrated superior performance.
• Regression methods yielded more promising results across all forecasting horizons.
• Performance varied by frequency, with specific methods being more suitable for different data granularities.

Conclusions:

• Certain regression techniques offer enhanced accuracy for univariate environmental time series forecasting compared to traditional methods.
• The optimal forecasting method depends on the specific application, data frequency, and desired trade-off between computational cost and performance.
• Further research should consider method suitability for different environmental variable types and forecasting needs.