# Top-Down or Bottom-Up Approach for demand forecasting

I have 5000 SKUs which all of them are highly positive autocorrelated, to get the item level forecast for all5000 SKUs (disaggregate forecast) which approach can provide more accurate forecasts, BU or TD, SES is forecasting method? and why?

BU approach: we do the forecast for all 500 SKUs directly TD approach: first we sum up the demand for all SKUs, then we do the forecast for aggregate demand, finally according to the weight of each SKU, we will obtain the disaggregate forecast for each SKU!(f(i,t)=pi*Ft)

-

This can probably not be answered based entirely on theory. It will depend on your time series, the length of history, even your error measure.

I would recommend that you fit both the bottom-up and the top-down procedure to your data with a holdout sample (say, the last three observations), forecast into the holdout sample and see which approach works best.

That said, grouping 5000 items in a top-down approach looks dubious to me. The fastest moving items will completely dominate the slower ones. Perhaps you could introduce a more fine-grained hierarchy?

SES (Single Exponential Smoothing) is certainly a good first candidate. You could also look at the other Exponential Smoothing methods. See, e.g., this free online textbook on forecasting.

-
Thank you, Actually I've done the forecasts and the Buttom-Up always works better, do you think it's because of the high positive autocorrelation? –  Roji Nov 20 '12 at 17:06
I used MSE as an error measure –  Roji Nov 20 '12 at 17:07
the time series data are stationary and follows ARMA(1,1) process –  Roji Nov 20 '12 at 17:09
That bottom-up works better may be because a top-down forecast with 5000 series is dominated by fast movers. MSE: do you average that over the 5000 series to determine which method works better for the entire sample? If so, the MSE will be determined by large errors (i.e., by series with high means). How much do your series differ on average? –  Stephan Kolassa Nov 20 '12 at 20:17
I didn't average errors, I calculate MSE for each SKU and for both approach , I'm trying to find a intutive explanation why BU works better for high positive autocorrelated data, this is my idea: When autocorrelation is highly positive, successive values of dt are positively correlated and so the process will tend to be smoother than the random series, With data are not so noisy, the signal can be detected at the bottom level, and so the bottom-up approach may do so well. Do you think that's make sense? –  Roji Nov 21 '12 at 10:14

This paper gives a framework for reconciling hierarchical forecasts. It also points out a flaw in top-down decomposition.

-