Skip to main content
Apr 26, 2017 1:29 AM UTC
Question

Forecast Accuracy Reporting

How do you people currently report Forecast Error/ Accuracy?  How do you bucket it ?

What parts do you report it on - only the planned parts or on all the parts ? 

What metric do you use: Forecast error %, MAD, RMSE, MAPE, Tracking signal or Bias ?

We have about 5 million part numbers in our system but we plan only for about 5000 pairs in total maybe , so doesn't make sense to report forecast accuracy on those parts, so I am trying to see what other companies / industries are doing. 

The data we can use is ipcs_forecasted_data.forecastamount and ipcs_demand_history.historyamount  will give you the recorded forecast & the actual demand history for a month when you join by partid, locid & historybegdate and forecastslicedate


Looking at the Demand Planning Cycle the first step is forecasting - pretty much with SVG under time phased planning - your forecast drives the parts that are being planned


Thanks

Regards

Natarajan


    3 replies

    May 15, 2017 2:09 AM UTC

    Hi Nate and others,

    A couple of thoughts re forecasting accuracy. (Apologies if I get all "traditional and consultancy" on you - won't get technical)  Simply put, we use a forecasting quality measure. If time pressured, stay on the improvement path for your volume A class lines.

    Here at Qantas we have started measuring our forecasts in two discrete ways.

    1 - We have large (separately codified, and discrete forcasting stream) maintenance activities running over several weeks where we are able to use an initial cut of our probablistic bills of materials to rate our A B and C class items within the event, and extend slightly wider tolerances for our C class lines. We also use the "n-1" rule to determine significance of the forecast in calculating the measure for a line item for pass/fail. When the event is finished we get to a % of items inside acceptable tolerance  per event. A pass is achieved when a part falls within tolerance or is within a single unit, and a fail otherwise.We tthen count the items in tolerance/total items forecast+unexpectedly used to arrive at the percentage. By the way we take extracts and pass them through our data warehouse and out to Tableau at lag 2 months. At this leadtime improvements in accuracy can still yield business benefits.

    2 -  The other method is still a quality pass/fail % in monthly buckets looking at part and/or pair level for the months actuals just completed. We still pass ourselves at n-1. It's effort versus reward

    When we take a version upgrade we are also going to change our forecasting consumption logic as this is causing us some overbuying. You might want to understand how your order plan copes with a part that has a large forecast, no usage month to date, and how the residual balance is reflecting in order planning / purchasing. Similarly and more impactful is what do you do when a large usage occurs that is simply a pull forward of future months. You could easily double stock without a demand control process. (We are not yet at real-time ourselves there)

    Btw taking info into our analystics tool does not mean it can't be in created in Intellicus. We had a lot of great graphs at our go-live but we have just scoped things a little tighter since.

    Cheers,

    Alastair

    May 15, 2017 10:57 PM UTC

    Hi Alastair,

    Thats nice thanks for exchanging your views and knowledge, this explanation is still at line level, part or pair level, what if the management asks for Forecast Accuracy as a percentage ? do you roll up all the sucesses / (successes + failures) at the site /region / organization level ?

    Thanks

    Regards

    Nate

    May 17, 2017 11:58 PM UTC

    We roll up and that is the % we use - it's a forecasting quality measurement.