Hello,
i am trying to forecast using eviews the stock price returns of a specific company. I am running automatic arima forecasting so as eviews evaluates and proposes the best model fit regarding to Akaike or Schwarz criterion. My problem is that when Akaike gives good results , the proposed model for Schwarz criterion is (0,0). Doesn't this imply non-stationary data and that ARIMA should be considered as a better model fit? On the other hand, running an ADF test gives results of stationary data. So, i am confused on how to use the SIC for my forecasts.
If all var are stationary, then ARIMA is not the appropriate choice, go for arma or arch/garch families as per requirements. AIC and SIC selection depends on the discretion of researcher, albeit sic performs well if sample is relatively small in T
As far as I know stock price returns are generally not autocorrelated. It is therefore not surprising that SIC (also called BIC or SBIC) leads to an ARMA(0, 0) model. AIC puts a smaller penalty to useless parameters. Not autocorrelated random variables does not mean independent variables. Squared returns are often autocorrelated, justifying the use of GARCH or other nonlinear models, as Muhammad Shakeel recommends. See a good book on GARCH models. Even if the forecasts are the same as a linear model, the forecast intervals will be more realistic.
In my experience, BIC may yield ARMA(0,0) when AIC doesn't. Your model choice depends on what goal you pursue. I discuss this in
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3362361
- Badges
- Science topic
- Similar topics
- Biological Science
- Kinesiology
- Running