What is it about the Kalman filter that makes it attractive from a retailer’s point of view ?
As mentioned in passing two weeks ago the original idea of the Kalman filter when applied to missiles and spacecraft was that as new information became available on the position, velocity and acceleration of the vehicle, the Kalman filter only needed to process the latest data. It was no longer necessary to have to reprocess all of the flight telemetry data recorded since launch to work out the position of the vehicle.
The application of this approach to retail point-of-sale data means that rather than having to process 52 weeks of sales data each week, only the most recent data needs to be processed. A single initial pass through 52 weeks of data is enough to calculate the filter’s coefficients. The filter can then be updated by addition of only the most recent week’s sales data. If you are having to estimate millions of SKU locations this is going to be pretty important. As new data becomes available, updating of the filter can occur at a greater speed than calculating a moving average model. All the more so with the advent of 64-bit servers.
This all well and good, but why go to all this extra trouble to calculate a stock model using a relatively complex method? The main driver of retail inventory levels is forecast accuracy. The stock you order today is the stock you have to live with tomorrow! From real-life retail experience, a Kalman filter estimate is often more accurate three weeks out than a moving average from only one week out. This translates to a 35% improvement in stock turns over a moving average model. If you get 3.5 turns from a moving average, you will get 4.7 turns from a Kalman filter.
There are probably better things to spend money on than unnecessary inventory.