The Research Of Financial Time Series Forecasting Models

This paper predicts the future trend of Shanghai Composite Index by applying four major forecasting methods, thirteen models in total. We select the daily closing price of Shanghai Composite Index from January4,2007to December31,2013as training samples, which are used to build a model, the daily closing price from January2,2014to March13,2014as testing samples, which are used to verify the model predictions.At the beginning, we predict the daily closing price from January2,2014to March13,2014with Double Exponential Smoothing Method and Holter-Winter No Seasonal Model respectively, and then we compare the results of the two models, calculate RMSE and MAE values under the two models, after that, we draw the conclusion that Holter-Winter No Seasonal Model has more accurate predictions than Double Exponential Smoothing Method.Then, we utilize ARIMA model to fit the training samples, by the application of EACF code and AIC information principal, the step number of the ARIMA model is selected, and we make predictions on the daily closing price from January2,2014to March13,2014with the selected ARIMA model. After that, we denoise the original data of training samples with the wavelets theories, and make predictions by applying the selected ARIMA model combined with denoised data afterwards. Then the predictions made from these two methods are contrasted, we find the model using denoised data generates better predicting results.Thereafter, considering the volatility clustering of financial time series data, we utilize ARCH model, GARCH (1,1) model, GARCH (1,1)-M model, TARCH (1,1) model, EGARCH (1,1) model, PARCH (1,1) model, Component ARCH (1,1) model, Asymmetric Component ARCH (1,1) model to fit the training samples, we make the conclusion that Component ARCH (1,1) model fits the data best and GARCH (1,1)-M model has the greatest predicting ability.Finally, we compare the prediction results of GARCH(1,1) under Normal Distribution, Student’s t Distribution and General Error Distribution(GED) hypothesis, we find that although t-GARCH(1,1) model and GED-GARCH(1,1) model fit the data better than N-GARCH(1,1) model, the predictions from the former two models are not accurate as N-GARCH(1,1) model.