Blog-hero.jpg

The Innovator's Solution

     

Five Ways Machine Learning Can Improve Demand Forecasting

Posted by Jeff Bodenstab on Jan 24, 2017 4:40:00 PM

Modeling Demand can Improve Demand Forecasting.jpg

Machine learning is being incorporated into solutions in every walk of life - home thermostats, health monitoring systems, equipment maintenance, marketing software, etc. In the latest generation of products, machine learning is adding intelligence pretty much everywhere you look.

Data is driving this trend. More data is available than ever before, but tools are needed to take advantage of it. Machine learning that allows the computer to “learn” from data even without rules-based programming nicely filling this need for improved analysis.

When it comes to demand forecasting, machine learning can be especially helpful in complex scenarios, allowing planners to do a much better job of forecasting difficult situations. It leverages the knowledge, experience, and skills of planners and other experts in a highly efficient and effective way across a broad range of data. Here are five areas where we have seen machine learning deployed specifically to improve demand forecasting, filling a clear and present need and delivering results and benefits.

  1. Trade promotions and media events
  2. New product introduction (NPI)
  3. Social listening (social media)
  4. Extreme or complex seasonality
  5. Weather data

1. Trade Promotions and Media Events

Promotions, advertising, and other “demand shaping” are expensive, yet determining their impact, or “lift,” is challenging. A large number of variables with complex interactions are buried in huge amounts of data with a high degree of noise. Even with substantial expertise and fairly consistent baseline demand, it’s usually not possible to understand correlations among variables.

But capturing this behavior is critical to producing an accurate forecast. To solve this problem, machine learning accounts for a multitude of attributes, ranging from product and market to social activity. This technique recognizes the shared characteristics of promotional events and identifies their effect on normal sales. Multi-dimensional modelling that handles both qualitative and quantitative variables is particularly well suited to describe and predict the non-linear demand driven by promotional activity.

For instance, at Danone, more than 30% of items are sold on promotion—accounting for nearly 70% of forecast error. The global foods company wanted to predict promotional lift to baseline demand to get timely production and balanced inventory deployment for channel and store supplies. Using machine learning, Danone lowered forecast error 20% and lost sales by 30%. It increased service level to 98.6%, and realized a 30% reduction in product obsolescence. Danone also cut demand planner workload in half, allowing planners to focus on more value-added activities.

2. New Product Introduction (NPI)

It's tough to forecast demand for a product without a sales history. But you can use machine learning to predict the performance of a product launch. Models can include early indicators such as web analytics, product attributes or even social media data, thereby predicting the performance and launch profile early on.

For example, Luxottica, a global leader in eyewear, adds 2000 new styles to its collection annually. It uses machine learning to cluster the behaviors of past launches, select the most probable performance for the new product, then “learn” common demand behaviors in the first launch period through detailed demand profiles. Luxottica improved global WMAPE (Weighted Mean Absolute Percent Error) by 10% and reduced the forecast baseline on new launches by about 30%.

3. Social Listening  

Traditional demand planning relies mostly on transactional data, creating latency between customer needs and supplier reactions. But social listening (now used by marketing departments to assess how their brand is perceived and their marketing campaigns are being received) can be used by the supply chain team to correlate social sentiment with demand signals.

These enhanced forecasting models can be especially helpful for providing early indicators of how the market perceives a promotional offer and upcoming demand. Or for new product introductions, where you can measure pre-launch—and during the first days of launch—sentiment, activity, reach, and location for an early inkling of demand.

To this end, we are leveraging a system called “Groover” that listens to social channels and gauges consumer sentiment to enhance supply chain planning. It monitors and stores live tweets on specific brands. This is big data—e.g., 45,000 tweets archived in just minutes. Natural language processing interprets the social communication and its impact and reach.

4. Extreme or Complex Seasonality  

Demand planning software can factor in seasonality—like the fact that more ice cream is sold in summer than in winter. But sometimes seasonality can become so extreme or complex that it is not as well suited to normal regression analysis-based techniques. For instance, Lennox Residential uses machine learning and cluster analysis to identify and track seasonality patterns and trends for its HVAC systems. The system recognizes more than 200 “micro-climates” within the United States and their seasonal timing variations. Machine learning sifts through the SKU-Locations to identify “clusters” with similar seasonality profiles. Lennox improved service levels by 16% while simultaneously increasing inventory turns by 25%.

5. Weather Data

Improving forecast demand based on weather depends on geographic areas, products, and demand lags. It’s tough to model statistically, as there are too many variables—but machine learning can crunch that data. Machine learning can let you use weather forecasting the way you evaluate causal factors like pricing and traffic—to get the best picture of demand for a particular product during a specific time series.

In an application that involved temperature and ice cream, a company used machine learning to develop a model for across several European countries. It improved demand forecasts as high much 5.8% in one country and averaged 3.8% across all five nations. The ability to incorporate weather data was most helpful in months of rising temperatures, such as the late spring, where sales were particularly sensitive to temperature spikes that prompted earlier than normal ice cream demand.


 Click below for a CSCMP article on machine learning and demand forecasting.

ToolsGroup CSCMP Machine Learning

Topics: Forecasting Demand and Analytics, Machine Learning