IMPROVE YOUR OUTLOOK

by Ryan Gunderson (8 April 2021)

Don’t you just love backorders? Of course you don’t. Nobody does. Even worse is when you’ve stocked out of the one product your customer wants, but you’re swimming in products that your customer doesn’t want. If you’ve lived it, you know what I’m talking about. So painful and frustrating for everyone!

What if I told you I helped a manufacturer increase their sales by 18% while reducing inventory by 12%? And that AI (artificial intelligence) helped me do it? I wouldn’t blame you for thinking that it was science-fiction.

But it’s not science-fiction. It’s process improvement, built on math, statistics, algorithms, machine learning, lean, and six sigma. This post summarizes five steps I utilize to increase sales and reduce inventory. I hope it will help you to consider whether the same is possible for your business.

1. Define the problem

It is impossible to perfectly predict future demand for your products. As a result, the quantity you buy or manufacture is based on an imperfect forecast. Safety stock inventory can buffer against some degree of demand uncertainty. But poor forecasts lead to missed revenue, excess inventory, or both.

Missed revenue is a pretty straightforward problem. If you don’t have a product available to sell, you miss out on the sale. If you’re lucky, the customer will buy a substitute product from you today. Or they will buy their desired product from you once you have it back in stock. In other cases, you will lose the sale -- and maybe even the customer -- forever.

If you’re wondering why excess inventory is problematic, there are two parts to it. First, if you hadn't purchased inventory that is now collecting dust or mold, you could have bought something else. Second, the more excess inventory you have, the more you end up throwing away. It’s obvious why perishable items, such as spoiled food, are written off as losses rather than sold. But many items without an expiration date can still lose value over time. Designs go out of style. Technology becomes obsolete. Excess inventory may hurt your profitability as much as missed revenue.

2. Measure your forecast accuracy

If you don’t currently have a process to regularly and formally forecast demand, you’re not alone. The silver lining is that it makes your baseline measurement especially straightforward. 

Here are a few formulas to help you measure the effectiveness of your forecasting process:

Absolute Error = |Forecasted Quantity - Actual Quantity|

Forecast Accuracy = 1 - (Absolute Error / Actual Quantity)

Bias = (Forecasted Quantity - Actual Quantity) / Actual Quantity

Let's assume a forecast of 0 units and actual demand of 100 units. Absolute error is 100 units. Forecast accuracy is 0%. Bias is -100%, meaning that you under-forecasted the demand.

Now let’s assume a forecast of 200 units and actual demand of 100 units. Absolute error is still 100 units. Forecast accuracy is still 0%. Bias is now +100%, meaning that you over-forecasted the demand this time.

For our final practice calculation, let’s assume a forecast of 130 units. Actual demand is, you guessed it, 100 units. Absolute error in this example is 30 units. Forecast accuracy is 70%. Bias is +30%.

Now that we’re all on the same page about calculations, I'll jump back to the manufacturer. Their baseline forecast accuracy was 70%, with a bias of -9%. Across more than 10,000 line items, the average forecast was off by 30%. Some line items were over-forecasted. Some were under-forecasted. Their overall process tended to under-forecast demand, by 9% on average.

If the forecast quantity was only off by 9% in total, why wasn’t forecast accuracy 91%? To fulfill customer demand, you need the right quantity of the right product in the right location at the right time and in the right condition (e.g., unexpired and undamaged). Having an extra loaf of bread on the shelf in Alaska is not helpful, if the customer who needs it is in Florida.

3. Determine the root cause of forecast error

The manufacturer had several different plants. The smallest plant generated tens of millions of dollars in annual revenue. The largest generated hundreds of millions of dollars each year. Each plant was responsible for its own planning. Some plants produced the same quantity and product mix as they shipped last period. Others supplied the average from the past 13 weeks. Yet others started with historical quantities but allowed sales, marketing, and/or finance to override the figures through a consensus process.

Were these forecasting approaches reasonable? Sure. They may even be similar to the approach you use. But if you’re still reading, I’m guessing that the results haven’t been as stellar as you’d hope. Here’s why those common approaches sometimes yield unacceptable results:

Simple historical averages aren’t great at handling variability in demand. Use them with growing demand, and you’ll have backorders. Use them with declining demand, and you'll have excess inventory. Use them with cyclical or seasonal demand, and you’ll bounce between backorders and excess inventory. Use human intuition, and the simple average is replaced by a simple guess.

Ready for a better approach?

4. Implement improvements that address the root cause

Proposing computer algorithms as the solution will always be controversial to some people. “Garbage in, garbage out,” as they say. I agree that AI-enhanced demand planning does not magically fix everything. That said, automated data-cleaning algorithms can flag potential data issues more quickly and easily than I am capable of doing.

Another common criticism is that computers can’t predict market disruptions. How would a computer know about upcoming product launches or product discontinuations? My bias is that good market intelligence is superior in such circumstances.

Speaking of bias, Daniel Kahneman won a Nobel Prize in Economics for his work on the topic. Kahneman identified numerous mental shortcuts that lead to quick, confident, error-prone judgments. Algorithms replace emotions with rules. Robust algorithms can even identify and offset sources of bias. The result is that algorithms often beat human intuition at forecasting.

Even when you know that most people aren't superhuman forecasters, it’s hard to accept that about yourself. I try to ease forecast participants into that reality. My discussions during a forecasting transition look something like this:

• The current forecast is represented by the green line, below.

• Historical demand is represented by the light blue bars.

• As you can see, actuals have been lower than the forecast for each of the past nine weeks.

• The orange, navy, and purple lines represent the latest statistical forecasts. Does one of those statistical forecast lines seem reasonable to you?