forecast

[Former Member]

Hi All,

i am working in the forecasting of the television product

i found the history with huge variation in data from year to year

and its peak sale was relevant to deevali dates.

we use period 15days and while consider the festival

in yr2005 it under the 15th period and 2006 comes in 14th period

so sales peak is deviate to yr to yr

and while firat yr its sales is moderate and next yr its going up

and the next yr its in down trend

so how we can forecast for this type of product

which method is suitable

pls reply me as adf

[srinivas_krishnamoorthy]

Since the Diwali sales is not on one day and is spread around the dates(probably couple of weeks even before the actual event), I would see the sales as seasonal in nature. One thing that I would do is to have the seasonal model have some flexibility around seasonal length (+ or - 1 month) Take a look at

In such a retail oriented sales, I would hesitate using workdays as a parameter. Sales happen more on weekends or holidays. However if there is an opportunity to rationalize the monthly sales based on calendar days in a month, one should explore that. I have found it makes sense to do a stat analysis on forecast accuracy on ex-post forecast to judge whether this adjustment is required.

[Former Member]

if you have an one time event like a festival and if its not consistent in gregorian calendars then you should try to clean up the history

clean history is an important part of getting a good forecast

its common practice to attribute a part of ur historical sales to events and remove it from the history(your sales guys would be able to say how much was due to the event) and after the forecast you can add a similar amount to the new event period to get the new forecast

you would be able to see the KFs - history adjustment and corrected forecast used in many places

similarly you can see forecast adjustment and corrected forecast and final forecast used in many places

many businesses use workdays based on a calendar as well for cleaning history since weeks with 3 workin days are give a differnt weekly forecast picture than weeks with 5 working day

the other way out is to use outlier correction.//