I want to forecast demand for the cement industry. I have data for 10 years- Monthly data. Is that desirable for forecasting?
$\begingroup$
$\endgroup$
2
-
1$\begingroup$ Why don't you try building a model for the first 9 years of data, and use the 10th year to validate your model? $\endgroup$– Max FlanderMay 28, 2015 at 5:02
-
$\begingroup$ The 10th year may not provide particularly good validation, but it'd be better than nothing - with small data sets there are always limitations. $\endgroup$– SilverfishMay 28, 2015 at 11:16
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
3
You can forecast a time series based on a single historical observation, by simply extrapolating this observation into the indefinite future (the so-called naive no-change forecast). Or you could even forecast time series with no historical observations whatsoever, in your case e.g., by estimating the current size of the industry and assuming a certain growth rate.
Of course, forecasts based on very few data points will likely not be very good. Usually more historical data translates into better forecasts. So your question really isn't whether 10 years of data is enough for forecasting, but whether this is enough for your forecasts to be good enough. And that, in turn, depends heavily on what you want to do with your forecast.
Then again, 10 years is quite a lot of data. The question really is whether any more data will improve accuracy. I'd imagine that the cement industry may have changed a lot over the last decade, so 11 year old data may not be comparable to more recent data. Your forecast accuracy will likely be influenced much more by how well you can forecast the future economic growth and industrial policy in China.
answered May 28, 2015 at 6:23
-
$\begingroup$ As per your explanation, minimum data like that of two years can be used to forecast. Initially the company provided me with their monthly sales and price data of two years but not sure of how accurate the forecast will be, I thought of doing it for the entire cement industry as the data I am getting from various sources is for 10-20 years. $\endgroup$– ChaMay 28, 2015 at 9:07
-
$\begingroup$ The dependent variable being production of cement The independent variables selected are Production of( Coal, steel, electricity, Oil),GDP, Inflation rate, Interest rate, Import export. My question is: Should I go for the two year data which the country has provided me with or should I take the data for the entire industry. ( I can change my aim from specific to general). Thanks! $\endgroup$– ChaMay 28, 2015 at 9:07
-
$\begingroup$ If you do causal forecasting, you will need to forecast your causal variables, too. Keep that in mind. In your specific case, I'd recommend building two models, one including prices and two years of history, the other one excluding prices but using ten years of history. Forecast out using both models. Then average the two forecasts within each future time bucket. Averaging forecasts very often improves accuracy. $\endgroup$ May 28, 2015 at 9:49
$\begingroup$
$\endgroup$
It is good that you have 10 years of data, but you should consider what many ignore and check your data for changes in variance. You can use Tsay's test to do this and then use WLS to estimate, if necessary. See more here Outliers, Level Shifts, and Variance Changes in Time Series www.unc.edu/~jbhill/tsay.pdf
You should also check your model to see if the model is changing using the Chow test for structural breaks as patterns change over time within the data due to policy, competition or unknown factors. See more here http://en.wikipedia.org/wiki/Chow_test
