Empirical study of robust combination of forecasts for short-term highway traffic flow forecast

In order to improve forecast accuracy and reliability of expressway traffic flows, the variance reciprocal weighting methods in linear combination of forecasts are compared numerically with the simple average. Ten individual methods for combination include the autoregression, exponential smoothing models, moving average models, and cybernetics method. Six variance estimators for the variance reciprocal weighting methods are the standard deviation, mean absolute deviation, median absolute deviation from median, fourth-spread, biweight estimator and Andrews wave M-estimator of scale. The empirical results show that the variance reciprocal weighting methods are usually better than the simple average, and they can be further improved by robust scale estimators.