forecasting error measure

[Former Member]

Hi Experts,

I have small doubt about the forecasting error measure like MAD, MSE, RMSE, MAPE, MPE & ET.

Does we get different result if we choose different error measure for strategy 40 i.e. forecast with seasonal trend model?

I am getting same forrcast for all error measure for this case.

The error measures are working fine & result changing in use Autoselection strategy - 56 for forecasting.

Does it mean that, all above error measure works only for specific forecasting strategy.

Your any inputs/experience are highly appreciated in this case!

Thanks for looking into it...

Rgds/Jay

[Former Member]

Hi Jay,

Since all the error measures are giving you same results, it means that strategy 40 is a better fit for the historical data, i.e. nullifying randomness in the long run.

You might get a lower error value using the Automodel 56, but it is transient as it will vary with a different data set. Automodel 56 is giving different values because it is trying to minimize error using different measures of error.

The result suggests that strategy 40 is better for long run forecasting and automodel 56 is good on your data for short run forecasting (if the appropriate error measure that you choose in your business context is lower than that you are getting from strategy 40)

If you are able to get in touch with a statistician, I believe he will be better in explaining this scenario.

regards,

biplab

[Former Member]

Thanks Biplab for your view on question....

I hv tested various forecasting strategy for my historical data & found out that 35 (seasonal + lienar) is the best fit strategy for my history data.

Also, while using strategy 40, which according to above discussion seems to be best fit, actually not giving correct forecast. My history horizon in from Sept to Dec 2009 with 0 history for some days. While, 40 calculate forecast only till 31 Jan eventhough i hv selected fsct profile till Apr end. Also, the forecast follows the lowering trend & has only value 2 or 3 for 30 & 31 Jan...

Rgds/Jay

[Former Member]

Hi Jay,

As you have rightly stated, assuming that you are forecasting in perishables / retail / FMCG, as you are taking 4 months history in daily buckets, and assuming that your season is of 1 month or less, a seasonal+linear model would be the best choice.

But also note that planning with such short seasons would not be able to capture the YoY / fall-summer trend of your business. You might want to revisit some mix with trend models if you also forecast in the long run.

As far as forecast with strategy 40 till 31 Jan is concerned, can you check if you have defined the correct number of periods (i.e. number of days)?

regards,

biplab