As technologies like RFID become more common, we will have the ability to monitor the pulse of every part at every location. And MCA fits very well with the autonomic concept
because as more and more information which reduces the uncertainty about future requirements becomes available, down to the level of an individual part or location or even down to a serial number level, we will be able to take that information and use it to optimize decision making for either the initial deployments or the redeployments of stock.
Our approach to forecasting begins with predicting parameters that describe and specify the underlying probability distribution of demand, the underlying random factors. In a low-volume, low-demand environment, there are a variety of probability distribution families that can be used. We then use those parameters, combined with historical time series, to come up with a projection for future periods.
What that generates is a sequence of probability distributions that you are projecting forward into time. To measure the accuracy of that we can compare it with past data - take a historical time series, divide it up, make our estimates, make projections, compare the observed distribution with the predicted and then use fairly classical non-parametric statistical goodness-of-fit tests to see if in fact we are fitting the correct distribution. This enables us to verify that we have in fact identified the right class of distribution family, the right shape and that we have estimated the parameters – the mean and the variance, and whatever else is involved – effectively.
It’s important to keep in mind, however, that the real question is not whether the forecast is accurate, but if the decision is correct. The forecast is a means to an end, a means to select the strategic targets stocking levels for deployment; it drives the tactical redeployment decisions. At the end of the day, the question is: have I achieved the performance, the availability, the fill rates and the predicted delay? How do they compare to what actually is experienced? So the accuracy is not in terms of the forecast and the numbers but in terms of the quality of the decision. And specifically have I achieved the planned for level of risk? If I stock to a 99% fill rate, am I achieving 99%? If not, I either have the wrong forecast or I have used that forecast incorrectly and picked the wrong level.
Managing the service supply chain, however, is a problem of risk management. In the service world, we don’t know those requirements in time, and in fact, we’ll never know those requirements exactly, and therefore we have to manage the risk, we have to keep track of the odds. It’s as if you went to Las Vegas and someone asked you to forecast what the next roll of the dice would be and based on that, make a decision.
You know that you don’t know for sure what the next roll of the dice will be, you know that it’s a matter of calculating the odds and finding the optimal way to manage the risk. And ERP systems are just not designed for that. They’re not equipped to predict when a part might fail – especially one that might have an MTBF of years or decades – and to figure out where the optimal location is to position replacements.