Improving Time Series Accuracy RRS feed

  • Question

  • What method or approach is recommended to measure the accuracy of the ARIMA forecast?  Once the accuracy is evaluated what parameters or approaches can be used to increase it's accuracy for a 30-day period?

    My data is sales by day with the timestamp equal to the calendar day (i.e '01/01/2012') for the period 01/01/2011 - 01/01/2013.  I forecasted the month of Jan 2013 trained on the period above, but the error between the forecasted and actual vary by as little as 2% or by as much 67% with a mean of 28%.  I have set my periodicity hint to {7,30,364} for weeks, months, and year and Auto_Detect_Periodicity to .8.  Generally speaking, our business has annual seasonality, the beginning of each month has high demand and trails to the end of the month, and a weekly pattern with high demand on weekend and slow mid-week.  The overall curve looks reasonable, but the error seems too high.

    I may be using the wrong approach to evaluate the accuracy of the model.  I thought to forecast only one day of the week like Friday's only, then Saturday's, etc.  Thanks.

    Tuesday, April 16, 2013 10:14 PM


  • Hi.

    You can try to use other variables with your algorithm.
    External variables like GDP, Consumer Confidence Level, etc.
    Or internal variables like new clients month, new items month.
    Or more random facts like special events at your town or weather.

    About measuring the accuracy i dont know either how to do it. :)

    Best Regards,
    Helder Borges

    • Marked as answer by Eileen Zhao Tuesday, April 23, 2013 5:55 AM
    Thursday, April 18, 2013 9:23 AM