Which of above is best for forecasting?
What is the time period that these methods can forecast accurately?
Hi,
I have a lot of experience with ARIMA, and some experience with Holt-Winters methods and GARCH as well.
It would be better to know what variables you are interested in.
Generally speaking, I would go first with ARIMA models. If you need to forecast a variable with a seasonal patterns, then you move to SARIMA models, (Seasonal ARIMA).
You can always add a GARCH structure to your model. GARCH structures model the conditional volatility of your variable. That is the value added of GARCH models. Well, they also need some forecast for the variable per se. I mean, if you want to forecast stock returns, then a GARCH model provides forecasts for the returns and its volatility. Sometimes it help to the return forecast to have a better model of the volatility, but that is not always the case.
Holt winters methods do not model the variables per se, they give you a procedure to forecast a given variable, but not a model.
I have used short and long term forecasts with ARIMA models. They are quite competitive with many others methods: multivariate methods, nonlinear ones, and so on and so forth.
So, I would go with an ARIMA model first.
Hope it help!
in forecasting,
if seasonal data, the ARIMA better than Holt-Winter.
if financial data, the GRACH better than ARIMA.
I agree with all the answers given. Just three points.
1. There is a model underlying Holt-Winters (HW) which is close - not identical - to the simple SARIMA(0,1,1)(0,1,1)s model (s denoting the seasonal period), see e.g. the Broze-Mélard paper on ResearchGate. That simple model can fit a large proportion of series, hence the success of HW. Of course, everything depends on the data.
2. I speak here of the additive HW method. For the multiplicative version (for strictly positive observations) which is hybrid (mixing additions and multiplications), it is not true but you can use additive HW on logarithms which corresponds to the so-called airline model (SARIMA(0,1,1)(0,1,1)s on logarithms).
3. R forecast can handle the three models and also, I think, something equivalent to Holt-Winters with GARCH.
Thank you for suggestions.
My observation is on Stock market Index with the applcation of multiple ES are getting not fitting with theory. As the Holt-Winters (HW) method have more MSE than the Holt DSE. Do it indicate that in a data series with high frequency that degree of smoothing generate higher error.
Can anyone suggest .
I do not have experience with high frequency stock market data but I wonder what "seasonal period" you can use with Holt-Winters method in that context. Also, combination of the exponential smoothing method with a model (GARCH, ...) for the volatility can help, I suppose.
Not surprising that it does not work well. Several authors have discovered weak (and very different from one market to another) weekly seasonality on daily stock, not on high frequency data, as far as I know. I have never heard of a monthly seasonality. Moreover the number of days in a month can vary.
- Badges
- Science topic