Comparison of data-rich and small-scale data time series models generating probabilistic forecasts: An application to U.S. natural gas gross withdrawals

- Incorporating factors improves forecasting ability, but including too many factors tends to exacerbate forecasts performance.
- This contradicts claims in literature that including additional factors does not alter outcomes.
- Factors may add information about seasonality for forecasting natural gas withdrawals.

Time series models derived from using data-rich and small-scale data techniques are estimated to examine: 1) if data-rich methods forecast natural withdrawals better than typical small-scale data, time series methods; and 2) how the number of unobservable factors included in a data-rich model influences the model's probabilistic forecasting performance. Data rich technique employed is the factor-augmented vector autoregressive (FAVAR) approach using 179 data series; whereas the small-scale technique uses five data series. Conclusions drawn are ambiguous. Exploiting estimated factors improves the forecasting ability, but including too many factors tends to exacerbate probabilistic forecasts performance. Factors, however, may add information about seasonality for forecasting natural gas withdrawals. Results of this study indicate the necessity to examine several measures and to take into account the measure(s) that best meets the purpose of the forecasts.