An Interview with Torsten Oldhues

A conversation with Torsten Oldhues, Managing Director of Operations at HAVI Germany, and Dr. Jens König, Principal Consultant at io.

To be more specific: How do forecasting methods help in managing and operating a logistics center?

JK: For example, you can forecast incoming orders and, accordingly, the picking workload. The same applies to incoming goods and the sales figures for individual items. The forecasted workload can also be used to determine staffing levels. However, the real benefit doesn’t come from the forecast itself, but from how you implement the insights.

And how do they help with planning the supply chain for restaurants in the chain restaurant industry?

TO: Exactly like that. Essentially, you use past data to predict the future. It’s easy to forecast sales for all products whose sales are stable and not volatile.

How do forecasting algorithms work? Can you explain that simply?

JK: Basically, you’re trying to find a mathematical function that best fits a given time series. Extending that function into the future is the forecast. 

Here’s an example: An item appears in five orders every day, with ten units per order. 

When represented as a function, this is a horizontal line. Any person or algorithm can conclude from this that five orders of ten units each will also be placed in the coming days. In real life, however, there are fluctuations rather than straight lines. We try to categorize these fluctuations into periods and cycles and predict them. Classic fluctuations, such as the summer slump, are easily identified. There are various methods for modeling such cycles. “Learning” methods determine the most suitable model with the best parameters for a given time series.

The goal, then, is to represent the given time series as accurately as possible using a mathematical function. Once a good function is found, the forecast is calculated using it. Modern forecasting is essentially mathematics and statistics combined with the automated search for a good solution.

So there isn’t just one forecasting method?

JK: No, there are various methods for modeling fluctuations, outliers, or peaks—or for the way algorithms find the best solution. You have to determine the right method for the problem at hand. Depending on the dynamics of the underlying time series, the best method may also change over time. It is therefore sensible and important to check from time to time whether you are still using the optimal model or whether another one might work better.
 

TO: That’s right. We, too, continuously maintain our planning system. We use JDA’s Advanced Planning Software and rely exclusively on the MLR method. Its implementation across the entire market took about one to two years. Incidentally, the forecast system discussed in the preceding article is used in this form only by our client McDonald’s—and even then, not by every single franchisee. Our clients are primarily from the chain restaurant sector. Due to the varying complexity of the system and our customers’ differing requirements for consistent availability of all products, it is too time-consuming and too costly for most of our customers. For many, demand planning based purely on historical data is usually sufficient.

When do forecasting methods and computational models reach their limits?

JK: It’s difficult to forecast rare extreme events, such as the effects of the COVID-19 lockdowns. With some clients, we saw a complete collapse in business. Others, on the other hand, experienced a sharp spike due to the sudden, dramatic increase in online retail. No forecasting model will detect such effects based solely on time series data. To identify them, you have to look for leading indicators or incorporate other external events—which makes developing and maintaining forecasting models a complex task.

TO: When it comes to chain restaurants, I like to mention the tour bus example presented in the article. I did say that looking at statistics on tour buses alone wouldn’t yield any results. Because even if you somehow managed to track all bus trips and had valid traffic forecasts and predictions about bus drivers’ break patterns, you still wouldn’t know, for example, which rest stop the bus driver will choose for a break.

Are there differences between various industries?

JK: The forecasting methods are completely indifferent to industry and content. It’s rather the characteristics of the underlying situation that determine whether something is easy or difficult to forecast. Our customers, for example, often look for extreme cases: days with unexpected peaks when everyone at the logistics center was struggling. In our experience, these events are difficult or impossible to forecast because they occur sporadically.

TO: We’ve had the same experience and have therefore established an internal S&OP process. The individual departments coordinate in advance and work together to prepare for more critical phases of the year that are difficult to forecast.

Where are demand forecasts easy, and where are they more difficult or even impossible?

JK: For a forecast to work well, there must be predictable patterns in the data series. This means that, first of all, there must be a sufficiently large statistical sample of—for example—items with consumption data for the approach to work in principle. Furthermore, it’s helpful if there are certain regularly recognizable patterns or trends. If specific external events or influencing factors—such as holidays or the weather—come into play, such information can be integrated into the forecasting model to improve the results if necessary.

Furthermore, data is required for the analysis and development of the forecasting model. However, this data is not available for a product that is new to the market. In such cases, substitute products or product groups are used that are assumed to behave similarly. However, it takes some time to determine how accurate this assumption was.

It is also important to always be clear about the required forecast granularity: should it be hours, days, weeks, or months? Often, it is sufficient to forecast the behavior of product or customer groups to arrive at the necessary result.

The path to a good forecasting model is therefore an iterative process. You start by analyzing the initial data and training the model, then implement it—on a trial basis if necessary—observe the model’s behavior and the quality of the forecasts it produces, and refine it, for example by adding additional factors or adjusting parameters. During this time, ideally, a sense of trust in the model develops, and users become familiar with its behavior. If the results are sufficiently good and everyone involved understands how and when the model reacts, it can be put to good use. Current forecasting methods already deliver astonishing results, as HAVI impressively demonstrates.

Is there still room for improvement?

TO: There’s always room for improvement. Of course, we don’t rest on our laurels either. For example, we’re currently working on implementing an AI-based forecasting system.