I am using time series analysis for predicting the future cost of materials, where I encounter some negative predicted values. I checked the input and everything seems okay. What to do in this case?
1) Should I replace the negative value with 0?
2) Should I take the averages of the previous years and keep adding them?
Or is there any alternative to avoid the negative values other than taking the absolute.
Thank you Janusz Milek for your fruitful suggestions.
As time series analysis requires larger data, therefore, I developed my own forecasting model to predict the data for lesser observations (in my case only 7 years cost data was taken, i.e. 7 annual observations only). Now, I encounter an issue, i.e. for some dataset, (Out of 250 only for 10 variables) I am getting a negative forecasted value. It is because I am incorporating the impact of the inflation rate as well. So, with time the prices are getting decreased, ultimately bringing a negative value. The negative value appears because the price is very less, e.g. Brick per piece is equal to RM 0.31. In this case what to do with the negative forecasted costs?
It seems that your model (equation) is severely misspecified. One should never add an absolute value like the price of a brick and a percentage change (inflation), even if they are multiplied by different coefficients. With a principally wrong specification, logs would not help - apart from the problem, that for a deflation (i.e. negative price change) a log does not exist. If you think that the price of inputs (pi) depends (partly) on the consumer price (pc), you can estimate ln pi = c + a*ln pc + ..., where a shows you the elasticity. One problem is, that an estimate on the base of absolute values (even if one takes their logs) is likely not very "sharp" (what concerns significance etc.). Therefore, I prefer the form %pi = a*%pc + ...%, with %pi and %pc as percentage changes of pi and pc, respectively. One must, of course, pay attention to a reasonable transformation of the other explaining variables, but with 7 observations one is rather restricted what concerns the number of explaining variables (7 is low even for one).
Thank you Anton Rainer for a detailed answer. In my case, I am taking the inflation rate in percentage, not the percentage change over the year. The issue arises only with those variables which are having a very low cost.
What is the inflation rate in percentage, what is the difference to a percentage change?
Anton Rainer e.g. Inflation in 2013 is 4%, in 2014 5%, so the percentage change will be 25%. In my study I am considering the inflation rate directly, not the percentage change.
Janusz Milek The impact of inflation is such that whatever the correlation appears, it has been considered.
Thank you for your suggestion, I will give it a try.
Inflation is already a percentage change, in general, a change of a consumer price index. In my approach I proposed yesterday, pc is, of course, not inflation, but the consumer price index. If e.g. pi is the brick price, than ln pi = c + a*ln pc or %pi = a*Inflation can never lead to a negative brick price. Nevertheless, I doubt, that consumer prices have a strong influence on brick prices, other than via wage cost, because wage increases may take into account inflation. In this case, there will likely be a time lag. I think, for bricks the energy cost could be more important than wages.
It seems that you and Janusz need some reading and exercise on indices (indicators) and transformation of variables.
Anton Rainer Thank you for a detailed fruitful suggestion. I will incorporate this in my analysis.