Are you reading the right signals from your customers?

An established “crawl-walk-run” approach to digital demand planning will improve forecasts, transparency, and efficiency.

Ken Spigarelli

Ken Spigarelli

Advisory Managing Director/Supply Chain Consumer and Retail Lead, KPMG US

+1 734-231-1311

Davis McNeil

Davis McNeil

Director Advisory, C&O Commercial, KPMG US

Eric Salazar

Eric Salazar

Director Advisory, C&O Commercial, KPMG US

+1 214-840-8192

In a prior KPMG blog, A digital path to Consumer & Retail business model transformation, we discussed the impact of rapid changes in today’s retail sector, including customer expectations, increased demand volatility, supply chain disruptions, and the emergence of new competitors such as platform companies and direct-to-consumer businesses.

A key strategy to address these changes is harnessing data and machine learning (ML) to enhance the demand-planning process. When properly designed and executed, digital demand planning provides greater insight into customer preferences, increased engagement across the purchase lifecycle, and improved product delivery models.

However, layering advanced analytical techniques into demand planning does not mean that retail organizations must redesign their existing processes all at once. In fact, the most effective way to develop a digital demand planning solution is with a gradual “crawl-walk-run” approach.

From crawling to running

The following example describes how to leverage ML into demand planning in five, carefully structured steps based on a crawl-walk-run approach.

Prior to making changes to the planning process, organizations should review underlying data, such as historical demand or sales, and ensure this data is accurate and complete. If the data feeding into the model is missing periods or segments, those gaps must be understood and addressed. Also, though COVID-19 has drastically changed trends for many businesses, that data should not be disregarded.

Segmentation: An easy place to start adding ML is with segmentation. Not all customers, product groups, channels, and geographies are the same. Segmenting, or creating the right clusters for the correct combinations of similar segments, is critical to developing an actionable set of insights. ML will provide greater insights than using product hierarchies or manually defined groups. This can be especially useful when addressing questions around what customers buy, how they buy it, and how customers expect to be served.

Leverage ML forecast models: Strategically review products and associated models to determine which would be best suited for switching to ML. Typically, an ensemble technique of identifying a winning model per segment works best versus a one-size-fits-all approach, as what is best for one company or product might not be the right choice for another. While many people associate ML with signals, the first step with ML is without signals. As confidence and proficiency with ML grows, expand the use of ML to other products or segments. By leveraging ML, you enhance your forecasting beyond traditional methods, which may not account for volatile demands that many are experiencing.

Maximize internal signals: Not all drivers are created equal—the key is finding the ones that are appropriate for specific needs. Organizations should look first at internal signals. These might include demand signals such as point of sale (POS) data, shipment data, order backlog, and proximity to true demand decreases from POS to order backlog; discounts and promotions; and customer relationship management (CRM) data.

Signal optimization becomes critical at this stage. Review the importance of each signal; only use a handful of meaningful ones to truly improve the forecast. Once you integrate internal data, you will be closer to an accurate, timely, and actionable model that accounts for diverse inputs from key stakeholders and functions across your value chain.

Evaluate and leverage external signals: Tens of thousands of external signals are now available, such as those involving commodity prices, demographics, weather forecasting, wages/payroll, construction permits, and job postings. These signals need to be carefully considered, systematically evaluated, and selected prior to model fine-tuning and testing.

Experience is key to a timely review and selection of signals. The goal is to find the right signal and not get distracted by the noise. Avoid unneeded complexity; if a signal does not improve the model, do not add it. Incorporating external signals will provide a confident leading indicator to better understand long-term timeframes.

Fine-tune the final model: Once the segments, algorithms, and signals have been selected, organizations need to continue refining the model. Incorporate structured feedback loops that enhance prediction accuracy over time. This approach involves retraining models with updated demand results, data, and signals, generating an improved predictive model structure for greater forecasting accuracy. The model is not a build-and-forget objective; reassess the model on a standard cadence.

Signals are a powerful tool to improve the accuracy of a forecast, but they won’t fix a model that isn’t built right. Ensure that the foundation is built correctly before adding additional layers. Once the foundation and structure are in place, fine-tune with appropriate signals.

Demand planning in action


A global consumer products company enjoyed strong growth during the pandemic. However, impacts on supply chains led the client to need a more accurate forecast. Manufacturing the wrong products and not allocating inventory levels across the right geographies and region resulted in lost sales.


KPMG presented the client with a blueprint for improved forecasting with ML and the incorporation of carefully selected signals. The KPMG team convinced leadership that adding signals to the existing model would not produce the desired results. Instead, a new demand-planning initiative was proposed based on the KPMG Cognitive Planning Solution.

  • Select the correct base algorithms to help fix the existing model.
  • Segment products, regions, and customers down to the correct levels.
  • Incorporate internal signals, such as order backlog data and planned price adjustments.
  • Incorporate external signals, such as total construction spending and employment, hours, and earnings to increase model accuracy.


Increased forecast accuracy, supplementing decision-making processes through a statistically tested and data-driven approach

Enhanced transparency and insight, providing additional understanding of internal/external drivers of customer demand and their impact on business results

Improved speed and efficiency, helping to accelerate cycle times for demand planning

Improved accuracy with ML and external signals

When using ML, we have typically observed an improvement of 10 to 20 percent in error rate over simple average models. The actual improvement varies based on the quality of historical data, as well as industry, function, and level of forecasting.

With the addition of external signals to a model (multivariate model) used in conjunction with relevant internal signals like pipeline data, e have seen improvements reaching 10 percent in error rate compared to a model based on internal data only (univariate mode). Some industries such as consumer goods/retail, or certain functions such as supply chain, are influenced more by external signals than others, so results may vary.

Advanced predictive techniques


  • Facebook prophet
  • Exponential smoothing
  • Bayesian Structural Time Series (BSTS)


  • Multiple linear regression
  • Bayesian ridge
  • Lasso
  • Elastic Net

Machine Learning

  • Random Forest regression
  • Catboost
  • LightGBM
  • Xgboost
  • Neural Network

A predictive-analytics build process applies and evaluates multiple established and leading techniques to determine the combinations that work best for specific requirements. In general, multiple techniques may be needed within a single organization or forecast.

How KPMG can help

The KPMG Consumer & Retail practice understands that supply chain leaders must think strategically about their supply chains to meet rapid changes in consumer demand, challenging capacity constraints, and rising input costs to the supply chain. Our multidisciplinary approach and deep, practical industry knowledge, skills, and capabilities help our clients meet challenges and respond to opportunities in a rapidly evolving business world.

With over 50,000 external signals available through our KPMG Signals Repository, we are primed to help your organization determine the right approach forward. Our proprietary active platform continuously harvests a broad variety of data from public and private sources, including commodity prices, demographics, and weather forecasting.

KPMG Ignition

KPMG Ignition services are designed to help our clients proactively plan for disruption, explore fresh insights, support new business models, and develop breakthrough solutions. Backed by our network of wide range of advanced capabilities, consumer and retail organizations can:

  • Discover insights on signals of change impacting their business.
  • Envision strategies to innovate, compete, and win.
  • Design bold scenarios with one of the leading art technologies that help decision makers visualize and interact with their data.
  • Deliver solutions that address complex business issues and align to changing business models.