8

I have a an ARIMA model which gives a pretty good forecast when compared to actuals. However it occasionally dips to negative values, while the quantity being predicted can never be negative.

Is there a way of forcing the ARIMA model to be positive?

Skander H.
  • 11,888
  • 2
  • 41
  • 97
  • 2
    If zeros don't occur, one possibility is to consider modelling the log-actuals. (If zeros do occur, there's a couple of possibilities, depending on how zeros arise/what you're modelling) – Glen_b Dec 16 '16 at 23:01
  • 1
    the simplest way is to constrain the forecasts to be > 0 by simply replacing/changing all negative numbers to zero . Logs kike any transformation/drug http://stats.stackexchange.com/questions/18844/when-and-why-should-you-take-the-log-of-a-distribution-of-numbers/18852#18852 should only be taken under special circumstances because of resultant side-effects. – IrishStat Dec 16 '16 at 23:20
  • I can see that if the series must always be positive then because of say measurement noise observations could go negative. Changing negative values to 0 as suggested by Glen_b and IrishStat satisfies the constraint that the series must be non-negative. I am curious as to why the series can never be negative. Can it sometimes be zero. – Michael R. Chernick Dec 16 '16 at 23:40
  • 4
    See http://robjhyndman.com/hyndsight/forecasting-within-limits/ – Rob Hyndman Dec 17 '16 at 09:47

1 Answers1

-1

ARIMA model while forecasting the future values uses differencing method intenally to stationarize the data,The negative values predicted are differenced values if u convert them to actual values that will be your End Result!

user225773
  • 1
  • 1
    I don't see how this answers the question. How do you know the negative values predicted are differenced values? Even if they were, what would prevent the undifferenced values from being negative? Furthermore, ARIMA does not use "differencing method internally to stationarize the data." You can specify how much differencing you want, but if you specify "$0$" for the differencing parameter, there will be no differencing. – jbowman Nov 04 '18 at 17:05