Statistical and Machine Learning forecasting methods: Concerns and ways forward.

Area of Science:

• Time Series Analysis
• Computational Statistics
• Machine Learning

Background:

• Machine learning (ML) methods are increasingly proposed as alternatives to statistical methods for time series forecasting.
• There is limited empirical evidence comparing the accuracy and computational demands of ML versus statistical forecasting approaches.
• The M3 Competition provided a valuable dataset for evaluating forecasting methods.

Purpose of the Study:

• To empirically evaluate the performance of popular ML methods against traditional statistical methods for time series forecasting.
• To compare accuracy and computational requirements across multiple forecasting horizons.
• To provide insights into the relative strengths and weaknesses of ML and statistical forecasting.

Main Methods:

• Utilized a large subset of 1045 monthly time series from the M3 Competition dataset.
• Compared the post-sample accuracy of popular ML methods with eight traditional statistical methods.
• Assessed computational requirements alongside accuracy measures.

Main Results:

• ML methods were outperformed by statistical methods in terms of accuracy across all examined forecasting horizons.
• ML methods demonstrated considerably greater computational requirements compared to statistical methods.
• The study found statistical methods to be more accurate and computationally efficient for time series forecasting.

Conclusions:

• Traditional statistical methods remain superior to current ML methods for time series forecasting in terms of accuracy and computational efficiency.
• The paper discusses reasons for ML model underperformance and suggests future research directions.
• Highlights the importance of large-scale, open competitions for objective forecasting method evaluation.