The M3-Competition found that simple models outperform more complex ones for time series forecasting. As part of these competitions, several claims were made that statistical models exceeded machine learning (ML) techniques, such as recurrent neural networks (RNN), in prediction performance. These findings may over-generalize the capabilities of statistical models since the analysis measured the total forecasting accuracy across a wide range of industries and fields and with different interval lengths. This investigation aimed to assess how statistical and ML methods compared when individuating series by category and time interval. Utilizing the M3 data and building individual models using Facebook© Prophet and R packages: tswge, forecast, and nnfor, there were significant differences in model performance. The statistical models performed better for monthly – industry, macro, and micro combinations (Wilcoxon signed-rank adjusted p-value < 0.0001) for short-term forecast horizons (h=5). However, the multilayer perceptron (MLP) surpassed the statistical models in quarterly – industry data (p-value < 0.001) for the same forecast length. The statistical models also outperformed ML methods for long-term forecasts in the same category by interval combinations (p-value < 0.01). Thus, identifying which model may have increased performance in specific category, interval and horizon combinations provides direct value for time series analysis.

Creative Commons License

Creative Commons Attribution-Noncommercial 4.0 License
This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License