Individual time series and composite forecasting of the Chinese stock index

Abstract

We explore the short-run forecasting problem at horizons of 1, 5, 10, 15, and 20 days for three forecasting periods within one year for the Chinese stock index from April 16, 2010, the launch date of the index futures, to May 19, 2014 with daily closing prices. We study forecast performance of 51 individual time series models that are different variations of autoregressive models, (Bayesian) vector autoregressive models, and (Bayesian) vector error correction models, and 41 composite models based on different trimming strategies of these individual models. The composite models, including the previous best forecast, equal-weighted average, inverse mean squared error, bias adjusted mean, shrinkage, and odds matrix approaches, utilize the idea of model boosting to diversify against possible mis-specifications, breaks, and structural changes in individual models, and aim at more robust performance. Across all forecasting horizons and forecasting periods investigated, we arrive at a shrinkage composite model with the shrinkage parameter of 0.25 that is optimal based on the mean squared error. This result is robust against the choice of futures series used in individual models and the pre-processing of structural breaks in data. We also discuss empirical findings at a more granular level, including comparisons of individual models and those of composite forecasts. Our results should fulfill different forecasting users’ information needs for decision making and policy analysis. The empirical framework also has potential of being adapted to similar time series forecasting problems in different fields.