The Best Approach to Forecasting

An article recently recirculated on Linkedin described the secret to forecasting as: “the best way to improve your forecasting is to reduce your forecasting.”  The article also provided a picture of some sort of martial arts contest and stated “it’s perhaps the humility of Japanese culture that allowed companies like Toyota to focus on responding to volatility, rather preoccupying themselves with [sic] more hopeless task of predicting and controlling it.”  The first statement is a good guideline, less forecasting is usually better, but that doesn’t help when forecasting is required. On the second statement, you don’t have to look to the Japanese culture to understand how to forecast better. Don’t get me wrong, my wife and I lived in Japan the first couple of years of our marriage and loved it. The best way to use forecasting is to understand the science behind forecasting and then apply the concepts appropriately to your business.

Science is the process of making observations about nature and then testing those observations to see if they can provide predictive laws of nature. For more on this topic, see Chapter 3 in Factory Physics for Managers. There are some useful laws of forecasting:

  1. The forecast is always wrong
  2. The forecast will always change
  3. The further out in time the forecast, the less accurate the forecast
  4. The greater the level of detail forecast, the less accurate the forecast
  5. If the product lead time promised to the customer is shorter than the time required to procure and produce the product, demand MUST be forecast or products will always be late.

Law 5 is a problem. It names a common business condition that requires forecasting. If I get my supply of suits from China and that takes six weeks, I must order at least six weeks ahead of time if I want a customer to be able to walk into my store, buy a suit and walk out with the suit. No way around it.

Law 4 is a killer also. Forecasting the number of size 44 navy blue suit jackets with pants that are 36 waist and 34 length gives forecasters the cold sweats. Typically, companies will allow their forecasters to report forecast accuracy in families. Forecast accuracy was 95% for volume of navy blue suits sold. “Great.” Forecast accuracy was 40% for navy blue suits with 44 jacket size and 36 x 34 pants. “Aww, that’s terrible.” Do companies sell suits that are only characterized by their color? No, so if the demand is not forecast by the color AND size, it’s much less useful for production. Here’s the thing, companies must forecast and then translate demand to the item level. Otherwise, production doesn’t know what to make.

The Linkedin article’s observations about responding to variability (volatility) rather than chasing it are dead on. Even then, you must forecast the level of variability to which you will respond. Will you invest resources to meet demand of 100 items/week +/- 10 or 500 items/week +/- 100? The “demand-driven” movement has appropriately identified the chaos caused by long cumulative lead times in MRP systems and ways to put inventory in place to buffer those lead times. Even then the amount of buffer stock must be forecast per Law 5. So in that regard, “demand-driven” is nothing new. All businesses are driven by demand. For more detail on inventory optimization driven by demand, as is all inventory optimization, click here or here. When forecasting for inventory, here’s the best way to use forecasting:

  1. Make a strategic decision about the average and variance you are willing to tolerate in demand and replenishment time for all items to be stocked. Replenishment time average and variance are strategic levers also—that is often overlooked.
  2. Measure forecast error at replenishment time for each item. Use Mean Squared Error to measure forecast error.
  3. Use this data to determine appropriate policies (when to order and order quantities) for all items. See pp. 115-137 in Factory Physics for Managers for more detail on the math.
  4. Monitor inventory position (on hand + on order – backorders) and make sure it stays between the appropriate control limits.

This approach ensures that as long as your assumptions about the range of variability hold up, you will get the results you expect in terms of required inventory investment and customer service level. This is fundamentally different than just forecasting an average demand for each item. However, operations science dictates that more variability requires more buffers. If you decide to buffer high variability with inventory, you will have a large investment in inventory.There are many different ways to successfully buffer high variability but that’s another discussion and why we wrote Factory Physics for Managers.

Here’s a little-known secret: if an item’s forecast error is greater than its historical variance of demand, you are better off using historical demand as the item’s future forecast. Understanding forecast error is very important:

  1. If forecast error is greater than historical variance of demand, forecasting is actually making things worse—increasing variability. This can happen no matter what fancy algorithm, e.g. Runge-Kutta or Box-Jenkins, you use for forecasting. Naturally, known future events such as sales promotions should be used to modify the forecast when using historical demand as your forecast.
  2. For any level of forecast accuracy, there is an optimal tradeoff of inventory vs. service. If the demand for suits has to be forecast and if 40% is the best forecast accuracy that can be attained for a suit with size 44 jacket and 36 x 34 pants, you at least want to make sure that you aren’t carrying twice the inventory required for that level of accuracy.

If you can shorten your lead times or if your customers will wait (make-to-order), those are good ways to reduce the variability associated with forecasting. Usually companies can not get completely away from forecasting. It pays to understand how best to use forecasting when you have to use it. – ESP

1 Nick Green. “The Secret To Improving Forecast Accuracy.” Linkedin, 15, April 2015, http://bit.ly/2qoOJhb. Accessed May 16, 2017.

    Comments are closed