Is historical time-series demand data really the best source for quantifying the demand random variation we need for determining correct safety stock?
Another source offers compelling consideration: forecast error. After all, we’re looking for random variation – variation that’s inherent in the demand process, but with unpredictable timing.
From a practical perspective – and correct safety stock must be practical, not just theoretical – random variation is variation that is not predicted in the forecast. Certainly, better forecasting may enable us to associate some variation that is currently random with predictable extrinsic or intrinsic predictors.
In other words, we could argue that not all random demand variation is truly random.
But in a practical sense, we need to achieve our service-level targets now, so we need correct safety-stock levels now. So, for the most part, demand random variation is the difference between forecast and actual demand – forecast error.
This may appear to be the ideal source of demand random variation, but its drawbacks, in combination, really cannot be overcome:
- Granularity. Customer sensitivity to on-time vs. late is at the daily level. Forecasts are rarely at the daily level. Instead, forecasts are often weekly or monthly. Total forecast and actual demand during that time interval does not have the granularity to represent actual demand random variation within the interval.
- Number of data points. Statistical reliability – the probability that the sample represents the population – depends entirely on the number of samples. A year’s worth of daily demand data offers perhaps 250-360 data points, and provides statistically acceptable reliability. By contrast, a year’s worth of monthly forecast-error data is only 12 individual points, and statistical reliability is very low.
- Multiple forecasts. Businesses constantly update forecasts. As an example – for a given item, the forecast that drove an item’s replenishment or reorder may be different to the forecast when that reorder was received.
- Forecast bias. Most forecasts have chronic bias, especially at the inventory-item level. Bias is definitely not random, and it – or its effects – must be isolated and removed from the quantification of random variation.
- Unpredictable special causes. Natural disasters are just one example of special causes that have an enormous impact on demand, may never be predictable, and should not be included in safety-stock-related random demand. However, they do affect forecast error. In order to use forecast error as a data source for determining safety stock, they must be isolated and removed from random variation.
Some of these drawbacks of using forecast error simply cannot be overcome, and in combination, they make forecast error an unacceptable data source for correct, optimal safety stock.
As a result, the best source of random variation in demand and lead-time is daily historical time-series data. This source is granular, has many data points, contains only actual results, and is not biased.
However, historical actual data may contain not only random variation – which we need for determining correct safety-stock levels – but also other variation that may make safety stock suboptimal if not addressed.
How we begin to address that in our next post……
