Based on the idea,mean-reversion in spread,of statistical arbitrage,this paper selects a portfolio of stock index futures with long-run equilibrium relationship via cointegration tests,and then builds three time series models,namely cointegration regression,ARX and ARX-GARCH models to estimate daily portfolio proportion,trading signals and practical spreads in-sample.By using rolling window and above models,this paper estimates parameters out-of-sample.This paper computes the optimal solution of linear programming from a view that minimizes the sum of weight differences between underlying cash as a benchmark and compares these proportion estimates from these three models.To comply with trading rules,this paper integerizes proportion estimates using rational fraction approximation and backrests trading models in-sample and out-of-sample under the hypothesis of 4 ticks of slippage and 0.27% fee rate per lot traded.From a comprehensive view of sensitive analysis and performance evaluations,this paper compares and analyzes long and short performances,of three models,in-sample and out-of-sample,focus on annualized rate of return,win ratio,average win to loss and Sharpe Ratio.The cointegration regression exhibits the top performing model in-sample and out-of-sample,for its highest annualized rate of return and win-ratio and most stable return curve.While ARX model is abandoned for its excessive volatility and drawdown.Additionally,the high return of ARX model in-sample is due to a few trades which may not happen in the future.Though,ARX-GARCH model is less preferred than the other two models in-sample,it exhibits the highest return out-of-sample and tolerable volatility.ARX-GARCH model can be developed by introducing stop-loss strategy. |