I have a time series data of log returns and wish to remove the serial correlation so that I can model using either ARIMA and Garch model. The question is how do I approach the problem, there is two alternative (or atleast what I think) using ARIMA and choosing the residuals or decomposing.
I tried fiting an ARIMA model but the box jenkins test return a pvalue lower than 0.05, which still suggest serial correlation.
residuals_kcb=residuals(auto.arima(kcb_log$Log_returns))
Box.test(residuals_kcb,lag=10,type="Ljung-Box")