Academic journal article The Journal of Business Forecasting

Beyond Traditional Time-Series: Using Demand Sensing to Improve Forecasts in Volatile Times

Academic journal article The Journal of Business Forecasting

Beyond Traditional Time-Series: Using Demand Sensing to Improve Forecasts in Volatile Times

Article excerpt

EXECUTIVE SUMMARY | Demand Sensing is a term that we are hearing more about recently. This article explains a way to think about Demand Sensing as many large, multi-national consumer packaged goods companies use it. It requires different thinking with respect to the sources and granularity of data, frequency of forecast updates, and how forecast accuracy is measured. While large companies have taken the lead on this, any organization can get started and realize benefits just by looking at data that are already available in their Enterprise Resource Planning (ERP) system. Companies that have implemented Demand Sensing have shown impressive improvements across many different Key Performance Indicators (KPIs) across the supply chain and entire business system.

With volatility now the new norm, what happened two years ago has become increasingly disconnected to what will happen next week, next month, or next year. So it is no surprise that Demand Sensing - defined by Gartner, Inc. as "the translation of demand information, with minimal latency, to detect who is buying the product, what attributes are selling, and what impact demand-shaping programs are having" - is top of mind for many manufacturers as a way to build a more agile and responsive supply chain. More than one-third of the total North American consumer packaged goods trade is already using Demand Sensing to manage demand; but if it is not yet at your company, you may be wondering what Demand Sensing is and how leading manufacturers use it to create demand-driven networks. What's different today is that Demand Sensing technology provides new capabilities to significantly improve forecast accuracy, and enable a fundamental change in the way companies view and measure forecast error by looking at error over exact lead time instead of approximating lagged error. Traditional demand planning systems work in buckets of time (e.g., a week) so error can only be measured by lag. However, the error necessary to understand supply chain performance is error over lead time. Demand Sensing enables companies to achieve better business outcomes by measuring error over exact lead time rather than using approximations when setting safety stock, for example.


The story of Demand Sensing starts with an appreciation of the capabilities and limitations of conventional forecasting systems. Popular demand planning solutions such as SAP-APO, JDA-Manugistics, and Oracle-Demantra typically use statistical forecast engines to create seasonal models based on traditional time-series analysis of prior sales data. Looking solely at sales history, these solutions are limited to producing an estimate of average sales for this time of year, with perhaps some adjustments to account for changes in promotional activity.

As the cost of computing dropped, statistical models grew in sophistication. Leading statistical engines now offer a myriad of model options including Fourier transform; single, double, and triple exponential smoothing (Holt-Winters); autoregressive moving average (Box-Jenkins); and intermittent and causal models, along with numerous configurable parameters. Despite this apparent sophistication, the forecast accuracy of time-series models has proven to be limited. Simply put, even with a full datasetfor every item, there is only so much information contained in prior sales data. To make matters worse, in practice, only half the consumer packaged goods items sold in North America have the required two years of historical data necessary for proper statistical analysis that accurately accounts for seasonality. The net effect is chronically poor forecast errors, with weekly mean absolute percentage error (MAPE) for the North American consumer packaged goods industry at 50% at the manufacturer distribution center level.

Overcoming these inherent limitations requires more data and new pattern recognition algorithms to extract meaningful information. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.