Assessing predictability of environmental time series with statistical and machine learning models.

Area of Science:

• Environmental statistics
• Machine learning applications
• Scientific modeling

Background:

• Machine learning methods are increasingly popular across scientific disciplines, including environmental statistics.
• Techniques like neural networks and decision trees are now common for environmental process forecasting.
• This trend challenges traditional statistical modeling and necessitates evaluating the role of established methodologies.

Purpose of the Study:

• To investigate the comparative performance of statistical and machine learning models in environmental statistics.
• To assess forecasting skills, uncertainty quantification, and computational efficiency of different modeling approaches.
• To inform the discussion on whether classical statistical methods should be retained or adapted for machine learning contexts.

Main Methods:

• Two time series case studies were conducted.
• Selected models from both statistical and machine learning literature were employed.
• Comparative analysis focused on forecasting accuracy, uncertainty quantification, and computational time.

Main Results:

• Machine learning models generally demonstrated superior forecasting skills.
• Statistical models provided more robust uncertainty quantification.
• Computational time varied significantly between different model types.

Conclusions:

• Neither statistical nor machine learning approaches are universally superior for environmental statistics.
• A hybrid approach, leveraging the strengths of both, may offer the most effective solution.
• Further research is needed to integrate model-based statistical principles with machine learning frameworks.