Traditional Product Quantity (PQ) or ABC analysis fails to recognize that high volume may not be predictable and that low volume can be.  So when trying to determine which products may lend themselves readily to pull techniques we can run into trouble if we don’t understand demand variation.  Get some demand data – build a spreadsheet or table listing products in rows and demand in columns, calculate average demand, Cv (the standard deviation divided by demand average), and plot it as here.  Here are some considerations:

1. Altitude – this analysis typically starts with customer demand but can be done by market, customer, finished goods, in-process, raw materials, suppliers i.e. anywhere along the supply chain where there is consumption.
2. Time Bucket – depending on the ‘clock speed’ of the enterprise the demand data can be aggregated by month, week, day, or hour.  When gathering data start with the smallest time period possible.  It’s easy to convert daily demand to week or month but impossible to take monthly and say anything about daily demand.
3. Unit of Measure – the units of volume may be straight forward, for example ‘pieces’ or may be complicated due to different value streams or product lines having a mix of units. You may need to convert your data to a common unit.  Dollars can be a common unit.
4. Demand – sometimes not easy to get.  Shipments may not represent true customer demand, especially if on-time delivery and order fill aren’t very high.  Getting original customer demand can be difficult if the customer order entry system forces ship complete or when back orders lose original request dates (and quantities).
5. Geography – like altitude, this analysis can be done by production cell, value stream, plant or DC, business unit, or enterprise.
6. Horizon – use forecast to determine volume, especially if product life cycles are short.  Using history to predict future volume is like driving by looking in the rear view mirror – it can be done, but reaction time is a little slow, and its hard to see the ‘cliff event’ until it’s too late.  Sometime forecast is crap or isn’t available, so then use very recent history.
7. Timeline – usually want at least 25 data points to calculate a meaningful standard deviation of the demand history.
8. Scrub – other than filtering out abnormal orders consider weekend transactions, huge one-time orders, and zeros.  Excel and Access treat blanks and zeros differently.  As you take the transaction log and build a pivot table of part number vs. date you’ll have cells with no data because no transactions occur on that day for that item.  Use search/replace to replace blank with zero.  But for new products coming to life during the study period you might leave cells blank prior to the launch date.  Similarly if a product has regular demand, say Monday, Wednesday, Friday you might want to leave Tuesday and Thursday blank.
9. Plot – volume vs. Cv
10. Interpret – Cv’s less than 1.0 lend themselves to flow and pull techniques.  Cv’s less than 0.5 can often be handled with rate-based replenishment methods.  Remember a Cv of 1.0 means the demand variation is as great as the demand average.  Say a part has an average daily demand of 100 with a Cv of 1.0 the demand one day could be zero and the next 200 or more – not very predictable.  High Cv items are usually low volume, but not always.  Take a look at the three data points in the top center of the graph above.  Must be a story here – why are the highest volume parts so unpredictable?  Big export order?  One time price promotion?