The least squares estimation method can be severely affected by a small number of outliers as can other ordinary estimation methods for regression models, thus providing poor out-of-sample forecasts. This paper suggests a robust regression approach, based on the S-estimation method, for constructing forecasting models that are less sensitive to data contamination by outliers. A robust linear autoregressive (RAR) and a robust neural network (RNN) model are estimated to study the predictability of two exchange rates at the 1-, 3- and 6-month horizons. We compare the predictive ability of the robust models to those of the random walk (RW), standard linear autoregressive (AR) and neural network (NN) models in terms of forecast accuracy and sign predictability measures. We find that robust models tend to improve the forecasting accuracy of the AR and of the NN at all time horizons. Robust models are also shown to have significant market timing ability at all forecast horizons.
|Number of pages||14|
|Journal||International Journal of Forecasting|
|State||Published - Jan 2007|
Bibliographical noteFunding Information:
The authors thank Paolo Colla, Eric Ringifo, Sharon Rubin, Shinichi Sakata, the associate editor, three anonymous referees and participants at the econometrics group in CORE, Université catholique de Louvain for their comments, which improved this paper. Preminger gratefully acknowledges research support from the Ernst Foundation.
- Exchange rates
- Neural networks
- Robust regression approach