Neural network forecasting of short, noisy time series

Abstract

The use of neural networks to forecast short, noisy time series was investigated by comparing neural networks used as function approximators for individual time series with neural networks used to optimally combine traditional forecasting methods based on training across a group of series. Neural network forecast performance on 384 economic and demographic time series was compared to two traditional alternatives: linear regression and the even average of six exponential smoothing methods. On these kinds of partially stochastic series, the use of neural networks as function approximators was found to generally give significantly less accurate forecasts than linear regression. The use of pre-processing to remove seasonal variation improved the forecast accuracy of neural networks, but an even average of exponential smoothing methods was still more accurate. In contrast, the use of neural networks to combine traditional forecasts gave small but significant performance improvements over conventional forecasting methods. Neural network combining of forecasts limits overfitting and, furthermore, can extract general rules of forecasting based on examples across many different time series, making it a potential alternative to rule-based systems.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 126 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.