PEMODELAN RUNTUN WAKTU FINANSIAL DENGAN VOLATILITAS GARCH MENGGUNAKAN ADAPTIVE NEURO FUZZY INFERENCE SYSTEM (ANFIS); MODELING FINANCIAL TIME SERIES WITH GARCH VOLATILITY USING ADAPTIVE NEURO FUZZY INFERENCE SYSTEM (ANFIS)

tarno, Subanar

2015 | Tesis | FMIPA

The purpose of this research is to develop the modeling procedure of Adaptive Neuro Fuzzy Inference System (ANFIS) for forecasting financial time series data. The focus of the development is selecting the optimal ANFIS model by using the inference of Lagrange Multiplier (LM) test. To date, there is no research about model selection in ANFIS using statistical inference based on LM test. The inference procedures include selecting the input variables, determining the number of membership functions (clusters number) and determining the number of rules, especially the application of ANFIS model for forecasting financial time series data. The results of theoretical analysis related to the development of ANFIS model include deriving estimator properties and its asymptotic distribution, which is consequence parameter estimation using recursive least square (RLS) method, asymptotic distribution of LM statistic. Distribution of the LM statistic is asymptotic chi-square. Modeling procedure with the basic of LM-test resulting new algorithm for selecting the input variable, determining the number of membership function and determining the number of the appropriate rules based on LM statistic. Procedure of selecting input variable is started by identifying the variable of lag input based on lag plot to the data and partial autocorrelation function (PACF) plot. Lag plot is used to test the linearity of lag variable to the data. Lag plot and ACF plot is used to identify autoregressive (AR) input. Lag input variable is selected by the inference of LM test based on LM statistic value of each input variable or the combination. The number of membership function (cluster number) depend on the heterogeneity of the data, which mean if the heterogeneity level of the data is higher, so the number of membership function is higher too. More number of clusters, more number of rules can be formed. Maximum number of rules that can be formed is mp , where m is the clusters number and p is the number of input. Verification of the developed procedure accuracy is validated using the results of simulation study. In the simulation study, four types of data are generated representing AR stationary process, AR non-stationary process, seasonal process and nonlinear process. Developed procedure of ANFIS modeling is applied to those four types of data and appears to be work out. The developed procedure is also applied for constructing the Indonesian inflation data, return of single asset data and portfolio assets return. That model is applied for forecasting the inflation and for predicting volatility of the LQ-45 stock return. Based on the result of empirical study, the model can work out and constructing the model with good accuracy of prediction when was applied to predict Indonesian inflation or to predict volatility of the LQ-45 stock return.

Kata Kunci : Financial time series, ARIMA, GARCH, ANFIS, LM test.