**The first forecasting blog covered an overview to inventory forecasting and the techniques available. This post continues to review the approaches to inventory forecasting.**

*Forecasting includes estimating forecast accuracy*

*Forecasting includes estimating forecast accuracy*

Forecasts are never perfect. You need to know how much actual demand may be likely to deviate from forecast, so you can plan safety stock to meet spikes. There are two ways to do this: (1) compare actual demand to the forecast for that month and (2) compare the model used to forecast future demand with previous actual history.

**Approach 1: Forecast Accuracy**

In the first situation, we compare the latest forecast for a month with the actual demand. For example, near the end of November we may have a forecast demand of 29 for December. At the end of December, we find the actual demand is 33. In that case, the absolute deviation of actual demand from forecast is 33-29 = 4. If the actual demand were 25, we would also have an absolute deviation of 4. Averaging these absolute deviations over time, we find the *Mean Absolute Deviation* or *MAD*.

Alternatively, we could sum the squares of the deviations and divide by the number of months to get the *variance*. The square root of the variance gives the *standard deviation*.

The biggest problem with the forecast accuracy approach is deciding which forecast to use. The forecast for December usage may change often as we progress from January through November. If we always compare with the latest forecast (November in that case), then the MAD or standard deviation figure may be misleading for forecast months beyond the first one. Another issue is that it may be problematic to recover the forecast data from previous predictions.

**Approach 2. Model Accuracy**

In this approach, the model used to forecast the future months is compared backwards against the historical demand numbers to see how well it fits the historical pattern. After compensating correctly for the *degrees of freedom* left in the model, either the MAD or standard deviation can be calculated similarly to the first approach.

Either approach can work to give at least a rough idea of how much demand variance to expect. From this, inventory planning can decide on a safety stock level to protect against spikes.

*Exponential smoothing can reduce computational complexity*

*Exponential smoothing can reduce computational complexity*

Forecasting can take a long time with tens of thousands of items stocked in multiple locations. The forecasting techniques described above require storing large quantities of history data and then performing calculations across these data sets. Naturally, this can absorb a lot of computer resources and force the forecasting process to occur at a down period like overnight or on a weekend.

**Exponential smoothing is a technique that reduces both the storage requirement and the computational complexity by requiring only the latest forecast and the new month of historical usage.** For example, given a new month of history, this approach forecasts the next month’s demand as a smoothing constant (between zero and one, usually a number between 0.25 and 0.33) times the new history figure plus 1 minus the smoothing factor times the old forecast.

Simple exponential smoothing can be extended to double exponential smoothing, representing both the expected demand level and the trend slope. This approach requires two smoothing factors. Often referred to as *Holt’s method*, double exponential smoothing also gives forecasts beyond the next month.

Likewise, seasonal effects can be added with a third smoothing factor, giving triple exponential smoothing. This is referred to as the *Holt-Winters* method.

Exponential smoothing can even be used to estimate the MAD of the actual demand from forecast. Simply combine the latest absolute deviation with the previous MAD, using an exponential smoothing factor, to get the new MAD estimate.

These exponential smoothing techniques vastly reduce both storage and computational burdens. However, there is no free lunch here either. Results are substantially less accurate than previously described methods. Triple exponential smoothing in particular does not do a good job of estimating seasonal effects. Because computer storage and power has become so inexpensive, most forecasting software applications have abandoned exponential smoothing in order to improve accuracy.

*Limited history causes forecasting problems*

*Limited history causes forecasting problems*

When you have a continuing product with years of history, forecasting tends to be relatively straightforward. But what about new products with limited history? How do you forecast for them? This problem is particularly pronounced with high tech products that may have a sales life as short as a few months.

- Analogy or proxy: Specify another item that has similar usage characteristics to this new one, perhaps modified by a multiplicative factor. For example, “I expect this screen display to sell 1.75 x what the older screen display sells.”
- Trend factor: Specify a starting level and trend factor. For example, “This fastener should sell 1,000 units the first month and increase by 10% per month for the first year.”
- Expert judgment: Enter a forecast from marketing or other company experts.

Want to find cash hidden in your Warehouse? Download our free ROI Calculator!