Machine Learning for Supply Chain Efficiencies

A few weeks ago I attended the ToolsGroup Supply Chain Transformation conference in Boston. One of the more interesting sessions I attended was “Predictive Commerce and Machine Learning” featuring ToolsGroup CEO Yossi Shamir. While Yossi obviously touted the benefits of the ToolsGroup product suite, the session focused more on the big concept of why supply chain needs automation now. In particular, he looked at how machine learning can help create better demand forecasts.

Traditionally, companies want to capture demand as close to the customer as possible, using inventory models to determine the appropriate product mix across the entire supply chain. Yossi outlined two basic principles behind the use of statistical models within the supply chain to accurately forecast demand. First, companies must use statistical models that include uncertainty. Uncertainties inherently exist within the supply chain. If statistical modeling is not taking these uncertainties into account, whether they are market outliers, supply chain disruption, etc., then the statistical model is only showing part of the big picture. Second, companies must look at deviations using the same statistical model. This ensures that uncertainties are also taken into account when looking for exceptions or deviations.

The big question from Yossi was “why machine learning”? First and foremost, creating statistical models is no easy task. To get really deep into these models, companies require a team of Ph.D.’s to analyze, interpret, and create fluctuating statistical models. Secondly, these models can only go so far. Yossi outlined eight factors that contribute to advanced modeling, which can reduce volatility in the supply chain. These factors can use machine learning to analyze data from many sources to take a forecast from the daily baseline to a more advanced and efficient form of demand shaping. As companies use each of these data sources, their demand insight increases.

• Statistical forecast: this is the baseline companies use to create initial demand forecasts. Companies are capturing demand and using multi-echelon inventory optimization to determine the appropriate product mix.
• Trend, seasonality, calendars, and daily patterns: supply and demand fluctuate for items over the course of the year. Understanding that there is a seasonality effect is key, as using summer demand, for example, to forecast winter supply, can be a devastating mistake.
• Trade promotion: trade promotions are a key relationship between retailers and suppliers. The trade promotion is an important tool for increasing demand of certain products and categories.
• Media event effect: media events can create a lot of excitement and hype for a product, which will increase demand, or they can have the opposite effect. Machine learning can decipher and analyze these trends to better understand the impact on the supply chain.
• New product introduction: new products generally create increased demand for a brand. However, the demand for the new product typically results in reduced demand for older versions. This information must be applied to the timing of a new product release as well as to the supply chain implications.
• Web indicators / sentiment: social media monitoring and data mining can reveal a lot about how consumers feel towards brands and products. Machine learning can identify trends in advance which might impact future demand of a product.
• Special events: these events can have a significant impact on the future of supply and demand. Understanding how events impact customer demand allows a company to better prepare for unexpected deviations from their traditional forecasts.
• Market intelligence: this is the everyday information that is available to a company about its markets. The information that is analyzed is used to determine future product, marketing, and supply chain strategies. Machine learning can synthesize this information into more readily available and digestible formats.

So the question remains, why machine learning? The factors outlined above all take data streams that impact customer demand. By applying all of these streams of data, a company can greatly improve its demand insight. But how does it work? According to Yossi, you need to have a baseline first. After that, companies need to figure out multi-variant correlations within the demand forecast. A learning machine can create rules based on behavior, and can create more reliable forecasts. Machines are taking data from statistical forecasts, seasonal patterns, trade promotions, media events, new product introductions, web indicators and social media monitoring, special events, and market intelligence to create these forecasts. All of these data sources, when combined with machine learning and analysis, are going to be a key component to the future of supply chain transformation. The big issue is just how long it will take for companies to connect these data streams to make improved demand forecasts.