Improve forecasting accuracy of short-term highway traffic flows by applying robust statistics to combination of forecasts

Abstract

The real highway traffic flows are time series sampled from typical complex systems. Combination of forecasts is necessary for the accurate, fast and reliable forecasts of them. The real traffic flows are complex stochastic processes, with time-varying probability distribution functions, and many outliers. The last two factors reduce the authenticity of point estimations of variances and correlation coefficients from the forecasting error series of the all individual methods, and directly reduce the accuracy of theoretical best combination weights. Using estimators to the complex time series by robust statistics, can improve the authenticity of point estimations of variances and correlation coefficients, then can improve the combination of forecasts accuracy of short-term highway traffic flows. Numerical test results show the improvements.