Leveraging Machine Learning in Supply Chain Management
Utilizing Advanced Methods in Demand Forecasting
Author: Clayton Nesslein
Manufacturers, distributors and retailers have a strong incentive to improve demand forecasting techniques. Inventory costs, along with shipping, logistical and increased labor costs have incentivized all entities within a supply chain to use the latest technologies to improve upon traditional planning techniques. Using history as a predictor of future demand has many flaws, and with the rise of big data and real-time information using point of sale, social media, and connected devices, forecasters have more tools available to improve accuracy and thus reduce costs.
The competitive nature of today’s marketplace has increased complexity for all stakeholders in a supply chain. The number of channels available for consumers is far greater due to the rise of e-commerce, consumers have a larger variety of products to choose from, and consumers demand a level of customization and customer service along with their products that was not experienced in prior decades. To improve responsiveness across multiple channels and effectively manage larger product lines with long tails, companies today use robust software, algorithms, machine learning, and neural networks to interpret complex patterns in consumer purchasing to improve the accuracy of their forecasts, which are the backbone for demand planning.
The benefits to a company of effectively forecasting demand can improve return on assets (ROA) to be a magnitude better than competitors, revenue growth can outpace competition to improve market share, and a leaner inventory management process can lead to improved inventory turnover. Moving towards a big data and analytics approach to demand planning is seen by managers today as one of the most important and high value investments a firm can make (Figure 1).
Figure 1
Foundations of Demand Forecasting
Forecasting demand is a foundational element for a successful business. The benefits are far-reaching, and are a key component in gaining the advantage over competitors, improving market share, and creating happy customers. There are 2 key and more specific outcomes that demand forecasting itself can achieve: 1) the expected demand and 2) the level of uncertainty given a companies’ current forecasting.
The more broadly defined results for a company that the forecast can achieve are far greater, including 1) reduced cycle time 2) improved on time fulfillment 3) reduced inventory 4) improved inventory turnover 3) reduced transportation costs 4) reduced inventory obsolescence and write-offs and 5) reduced labor costs in many facets of a business.
These local improvements to results are a great tactical consideration and should convince management to invest sufficiently in their demand planning department, but these individual components with improved results lead to an ultimate goal for the company: “Profitable customer fulfillment. The objective is to ensure that your supply network-including inventory, manufacturing, logistics and sourcing functions can profitably respond to the predicted demand. Building a responsive supply network requires strategic initiatives like network design and supplier selection, tactical efforts like defining a robust S&OP process as well as operational initiatives in areas like order fulfillment and available to promise” (Tohamy). All of the above considerations justify intense focus and investment on an organization’s demand forecasting systems.
Traditional Demand Forecasting Methods
Traditional Demand Forecasting methods generally fall into three categories. 1) Qualitative Forecasting 2) Extrapolative Forecasting 3) Causal Forecasting. The common link is that they all use historical data to predict future trends, and it is highly likely that the frequency between forecast updates is probably one month or greater. The primary argument of this article is that companies are moving towards Advanced Methods where real time data is used to make the most accurate decisions possible, thus reducing any latency between data collection and subsequent planning. The section on Advanced Methods also touches on advanced pattern recognition, whereas traditional methods are constrained to simple patterns related to seasonality, sales, and other macroeconomic trends.
Qualitative Forecasting is the least technical method, and uses the opinion of experts to predict demand. Panels of experts, or even just managers and personnel, can be convened in order to generate forecasts, but its shortcomings are clear. Any number of biases can arise, from the injection of anecdotal evidence into the forecast, or omissions of key data points that can lead to erroneous and incomplete results.
Extrapolative Forecasting is a quantitative method that uses time series techniques to forecast demand based on past sales. It is an extrapolation of what is going to happen in the future, based on historical trends. Two simple methods that are commonly employed include weighted average and exponential smoothing. For there to be a reasonable expectation of accuracy within some acceptable range, there needs to be consistency in the environment, meaning that conditions in the past are expected to prevail in the future. The introduction of new variables within the marketplace, such as new competitors or new consumer dynamics, will severely discount the value and results of any extrapolative forecast (MBA Knowledge Base).
Causal forecasting, aka Multiple Linear Regression (MLR), is a statistical method using regression analysis to predict demand. Many external factors and their links to expected purchasing patterns can be analyzed. Ordinary least squares (OLS) method obtains a linear approximation of the data set that is as close to the observed responses as possible. Independent variables can then be classified by level of correlation they have with the dependent variable. Internal and external factors can be added to such a model, including but not limited to advertising spending and promotions, new product launches, strategic partnerships, macroeconomic conditions, competitor pricing, and technological advances (Monahan).
The downfall of extrapolative and causal forecasting is that:1) historical trends do not predict future demand, and 2) relationships are not always linear. This is a key concept that spans any field or industry that uses forecasts. There will always be a question of how significant the past is in estimating the future. Humans are unpredictable in their purchasing patterns, and it is not likely that past trends will repeat themselves exactly, as external factors and forces will play a role in creating an everchanging and dynamic environment. Without advanced methods, the accuracy level of forecasts is capped at a rate that will inevitably produce wasted inventory and inefficient systems.
Advanced Methods
The common theme of all advanced methods is that they use real-time data, big data, and artificial neural networks. Recent advances in technology have allowed companies to leverage new technologies in communication and advanced computer processing techniques to analyze and respond in real-time to changes in their supply chain.
Advanced Method: Machine Learning
Recent developments in technology have led to applied Machine Learning (ML) techniques, which include neural networks, recurrent neural networks, and support vector machines. All have been applied in supply chain management and demand forecasting to reduce forecasting error and more effectively combat the bullwhip effect. Mean Squares Error (MSE) and Mean Average Error (MAE) are two primary metrics used to measure the accuracy of ML models against traditional methods such as Multiple Linear Regression (MLR). We can compare the methods to understand how the deficiencies of MLR can be overcome using more advanced techniques. “Trend-based forecasting is based on a simple regression model that takes time as an independent variable and tries to forecast demand as a function of time. The MLR tries to predict the change in a demand using a number of past changes in demand observations as independent variables (R. Carbonneau et al, 1143).” To reinforce the point about the shortcomings of linear modeling, MLR is used as a causal method to determine the relationship between a forecast variable and causal variable. This means that past data can establish a causal relationship between the dependent and independent variable. The problem lies in the fact that causality and correlation are very different, and a correlation can be established erroneously due to certain temporary anomalies and phenomena in a given data set.
We expect to see a high level of non-linearity in supply chains, as complex and irrational behavior is often exhibited. Advanced methods incorporate non-linear models and pattern recognition, and thus can serve as a better forecasting tool than linear models. The underlying structure of artificial neural networks is what allows for the advanced pattern recognition techniques and an improved sense of “learning”. As shown in a Figure 2 (Mathworks), the network’s MSE declines rapidly as more data is fed into the model and it subsequently learns. Learning, in this case, can be defined as reduced error beyond that of a linear model.
Figure 2
Researchers have examined and compared MLR techniques alongside those of neural networks to establish how effective Machine Learning can be. In the research article “Application of machine learning techniques for supply chain demand forecasting”, R. Carbonneau et al. analyzed the bullwhip effect within the Canadian Foundry industry and applied different forecasting techniques. Foundries were intentionally used as the subject, since they are far upstream and many layers exist between steel manufacturing and the final customer. Many forecasting techniques were applied to same data set, including Naive Forecast, Average, Moving Average, Trend, MLR, Neural Networks, Recurrent Neural Networks, and Support Vector Machines. The goal was to compare Mean Average Error (MAE) of the techniques to see which would be the most efficient. The results show that Recurrent Neural Networks and Support Vector Machines (both are types of Machine Learning) outperform MLR and produce the lowest MSE, and are thus the most effective methods of demand forecasting.
Advanced Method: Demand Sensing
Demand Sensing is the ability for companies to use real time data to predict when and where consumers will buy a product. Massive amounts of data are synthesized from a variety of sources including point-of-sale and social media in real-time. Historically, many supply chain decisions were made at the national and regional level, but Demand Sensing is able to focus on SKU and location level information. The ultimate goal is to drive planning at a higher level of granularity, using a variety of internal as well as external data sources to construct a more responsive demand model (Figure 3).
Figure 3
The underlying mathematics and pattern recognition utilized by Demand Sensing software is much more sophisticated than that used in traditional techniques. Rules-based formulas such as moving average and exponential smoothing are replaced by software that sifts through large quantities of real time data to determine specific influence factors that affect demand. The algorithms utilized can determine what data points are predictive and what is just static that needs to be ignored.
Predicting sales at an aggregate level can be replaced by a model that is focused on predicting consumer behavior. Demand Sensing tools are used to “get a better notion of actual buying patterns” (SupplyChainBrain - Kellogg). These buying patterns can be used to influence production and delivery decisions in the near term, leading to higher levels of efficiency and a leaner inventory.
Folinas & Robi state two primary goals of Demand Sensing. The first, On-Shelf Availability, aims to improve the availability of products in-store. Efficient and accurate replenishment will close any gaps between actual demand and supply. Demand not met due to inadequate shelf supply is a sale that is lost forever. The second goal is to improve the working capital position of the business. Inventory is money trapped in the system. Inaccurate forecasts lead to inventory write-offs and mismatched on-shelf availability, thus leading to inventory write-offs and a reduction in profit margins.
Humphrey & Laino outline three key concepts that are central to Demand Sensing:
1. Latency Reduction: Traditional methods in demand forecasting generally set a cadence on data collection and subsequent planning in which the frequency can be one month or greater. Demand Sensing seeks to move away from these long cycle times, towards a “Sense and Respond” model. “Sense and Respond” reduces the latency between sensing demand and execution response (Bodenstab, Jeff). The goal is to more closely align production and warehousing with downstream pull signals such as point-of sale, or even social media activity that is deemed to influence product demand. Companies can then react in real-time to demand spikes and stock-outs.
2. Downstream Data Integration (DDI): DDI includes downstream supply chain information in a demand forecasting model. It attempts to move away from the traditional model of forecasting past sales to predict future sales, and instead uses a variety of real-time sources as “predictors” within the model. Demand Sensing software can differentiate between demand signals that include demand versus those that do not. Utilizing the most current daily sales and current stock position data reduces uncertainty and increases responsiveness to expected demand. DDI can therefore enhance accuracy of deployment, buffer stocks, and shipping plans.
3. Measuring the impact of demand shaping actions (DSA): Demand shaping occurs any time a company runs a promotion, advertises, or launches a new product in order to influence demand. Capturing that information along with determining the impact of such events in a near term model is a key component of Demand Sensing.
Now that the foundations of Demand Sensing have been covered, real world examples of their applications are discussed.
Examples
Case Study: Nike
Situation - Nike is a global provider of footwear and athletic apparel. The company manages a global supply chain and manufacturing network consisting of 529 factories in 41 countries. Accurate demand planning for such a behemoth of interrelated suppliers is paramount. Production lead times can be weeks or months, but none of that matters if Nike doesn’t truly know what its customers want, and where. Profit Margins are diminished when products are overstocked, and the company is forced to offer price reductions to spur demand.
Resolution - Nike invested heavily in digital demand sensing, consumer data and analytics to improve speed and flexibility in the supply chain. At the Nike concept store in Los Angeles, CA, all its on-shelf inventory is specifically tailored to local trends, using digital sales data as the predictor. Product assortments update frequently based on what is trending in nearby zip codes. This initiative along with similar endeavors has cut lead times in half and drives double digit growth in key market segments (Cosgrove).
Case Study: Kellogg Company
Situation - Kellogg Company, with 2018 revenue of $13.5 billion, is the world’s largest cereal maker. Famous brands in the portfolio include Rice Krispies, Froot Loops and Apple Jacks. They are also a top competitor in the snack aisle, with brands such as Eggo, Cheez-it and Pop-Tarts. The cereal aisle used to be predictable and boring, with products benefiting from long life cycles. Not much changed, which made demand planning at Kellogg simple. In the 21st century where brands rise and fall at a rapid pace, the company has had to adapt more sophisticated planning techniques to capture market share and capitalize on the most recent trends.
Resolution - Kellogg has shifted away from traditional demand planning, which typically looks at seasonal patterns and average sales. They have been able to aggregate all data that is relevant to selling a product, including historical trends, POS data, and inventory data, along with other transient short-term signals to improve the demand forecast, resulting in better customer service and better cost controls (SupplyChainBrain - Kellogg).
Case Study: Granarolo
Situation - Granarolo is the leading producer and distributor of milk and yogurt in Italy. The company has eight manufacturing plants, along with 6 DC’s and 35 transit points. The product line consists of 1200 SKU’s, 200 of which are milk. The inherent nature of dairy as a perishable good brings to the forefront the need for accurate demand planning and inventory management. To complicate matters, Granarolo’s sales model is centered around promotions, which causes spikes in sales that distort interpretations of what true demand might be, and would make it difficult to analyze predictive signals using traditional methods.
Resolution - Granarolo implemented Demand Sensing systems that focused on Trade Promotion Forecasting (TPF), “which uses machine learning technology to translate historical data into reliable estimates of future promotions. Based on past promotions, TPF automatically generates proposals consistent with promotional peaks. The system proposes dynamic safety stock levels that take each product class’ forecast accuracy and store replenishment frequency into account, so Granarolo can maintain high service levels in the face of changing demand” (Granarolo Case Study).
Case Study: Lennox
Situation - Lennox is a major provider of HVAC and Refrigeration systems in North America. The residential heating and cooling business, which represents a large portion of the company’s total revenue, was faced with a myriad of demand planning and inventory issues. Heating and air conditioning units are a highly seasonal business with high capital demands. Undershooting forecast targets can lead to lost sales that will never be recovered, and overshooting targets leads to additional capital tied up in slow moving inventory. The supply chain consisted of 80 locations and was growing quickly. Lennox was managing 450,000 SKU’s including many that were part of a long tail and were slow moving. External forces such as weather and macroeconomic trends played a major role in sales performance, therefore using traditional methods such as past sales data to predict future demand was not a good indicator.
Resolution - Lennox implemented Demand Sensing systems that model and plan pre-builds in the highly seasonal environment. They created latency reduction systems with weekly reorder points, safety stock and order quantity updates. The demand sensing itself worked on a weekly cadence to focus on exception management. S&OP benefits from a highly responsive monthly inventory planning frequency with the most up-to-date information. Perhaps the most exciting aspect of their new systems includes machine learning to reliably model the highly variable seasonal demand patterns (Lennox Case Study).
Case Study: NHS Blood and Transplant
Situation - England’s National Health Service (NHS) was responsible for managing the blood supply for hospitals throughout the country. Donations were stored at 5 manufacturing sites and 15 stock holding units. The stakes were high, as stock-outs of certain blood types were unacceptable and could cause patient deaths. Blood storage, transportation, and effective inventory management poses many challenges. There was major supply and demand variability, as there are many blood types that need to be stocked, while demand can spike from major incidents such as fires or terrorist attacks. Also, blood is perishable, so the system needs to be responsive and move rapidly. Historically, NHS relied on the manual collection of data that was crunched by databases and even within spreadsheets. The organization knew that it needed to adapt to meet the needs to patients, and to minimize waste within the system.
Resolution - NHS implemented Demand Sensing systems which can rapidly respond to the dynamic inventory environment. Their new systems forecast demand at a high level of granularity, all the way down to the blood product, type, and on a hospital-by-hospital basis. Their new level of visibility allows workers to rapidly respond to shortages of specific blood types by reaching out to those very donors who may be able to help replenish stocks. NHS now has an automated “pull” system that analyzes stock levels across the country every 30 minutes, which is the translated into signals for distribution, manufacturing, collection, and supply (NHS Blood and Transplant Case Study).
Conclusion and Managerial Implications
Traditional methods of demand forecasting lack the level of accuracy and responsiveness that is required of organizations in the 21st century. The standard method of using monthly data on sales, production and inventory needs to be augmented with external and unstructured sources such as websites, beacons, apps, social media, and connected devices. Managers who can effectively utilize and implement advanced techniques will be able to gain an edge on competition while reducing inventory, lead times, and labor costs. Advanced methods in demand forecasting will also lead to reduced bullwhip effect, while creating more streamlined, agile, and responsive supply chains.
References
Figure 1: Ellis, Simon. “Sales & Marketing Report 2017: What to Do with Data”. https://consumergoods.com/sales-marketing-report-2017-what-do-data. 6/2/2017.
Bodenstab, Jeff & de Kok, Stefan. “What’s Wrong with Demand Forecasting?”. http://www.supplychain247.com/article/whats_wrong_with_demand_forecasting. 12/28/2015.
Logility. “Pratical Tips to Improve Demand Planning”. http://www.supplychain247.com/paper/pratical_tips_to_improve_demand_planning. 2/12/2018.
Tohamy, Noha. “Attaining The Next Maturity Level In Demand Management”. https://www.supplychainbrain.com/articles/4760-attaining-the-next-maturity-level-in-demand-management. 2/19/2009.
MBA Knowledge Base. https://www.mbaknol.com/research-methodology/extrapolative-forecasting/.
Monahan, Ciarán. “Supply Chain Models – Multiple Regression Models”. http://ciaranmonahan.com/supply-chain-models-multiple-regression-models/. 6/21/2016.
Figure 2: Berttram, Phillip & Schneider, Judith. “The Magic of Predicting Demand from Data”. https://www.strategy-business.com/article/The-Magic-of-Predicting-Demand-from-Data?gko=94906. 1/15/2018.
SupplyChainBrain - Kellogg. “In a Complex Retail World, Kellogg Company Turns to Demand Sensing”. https://www.supplychainbrain.com/articles/13955-in-a-complex-retail-world-kellogg-company-turns-to-demand-sensing. 8/6/2012.
Folinas, D. & Rabi, S. J Database Mark Cust Strategy Manag (2012) 19: 245. https://doi.org/10.1057/dbm.2012.22
Humphrey, Evan & Laiño, Federico. “Improving Forecast Accuracy through Demand Sensing”. https://www.supplychain247.com/article/mit_improving_forecast_accuracy_through_demand_sensing/nulogy. 11/19/2018.
Bodenstab, Jeff. “Reduced Latency – Gartner on “Sense and Respond” Supply Chain Planning”. https://www.toolsgroup.com/blog/gartners-sense-and-respond-supply-chain-planning-model/. 5/5/2015.
Cosgrove, Emma. “Digital demand sensing' is Nike's next frontier”. https://www.supplychaindive.com/news/Nike-demand-sensing-next-frontier/538484/. 9/30/2018.
Granarolo Case Study. ToolsGroup. www.toolsgroup.com
Lennox Case Study. ToolsGroup. www.toolsgroup.com
NHS Blood and Transplant Case Study. ToolsGroup. www.toolsgroup.com
Mathworks. www.mathworks.com
R. Carbonneau et al. “Application of machine learning techniques for supply chain demand forecasting”. European Journal of Operational Research 184 (2008) 1140-1164.
