Achieving the future of demand planning

Robert Byrne, CEO of Terra Technology

Despite calls for supply chain rationalisation in recent years, networks have become increasingly complex with more locations and continued SKU (Stock Keeping Unit) proliferation.

For planners, market uncertainty has compounded the problem, making forecasting the most challenging in decades. Fortunately, the amount of supply chain data available to manufacturers, both from their operations and from trading partners, has also grown exponentially. Within these masses of data, each individual data point contains small but valuable information about the current state of demand that can be unlocked to provide the most likely prediction of future demand for any planning horizon. 

Since the advent of Holt-Winters exponential smoothing in the 1950s, there has been little innovation in the fundamental techniques used to forecast demand.  During that time, many areas of the economy, such as the Internet, mobile communications and social media have undergone transformational change, sometimes more than once. 

While computer platforms used for forecasting have grown faster and more powerful, the underlying algorithms remain essentially unchanged. Most forecasting software in use today was designed for hardware capabilities of 30 years ago.  It’s time to upgrade planning software to take advantage of today’s technology. 

Developing the right tools

At a recent conference, Kellogg summed this issue up by explaining that there is a difference between simply having tools and having the right tools. They stressed that predicting demand is about making the right choices, and that includes tools, processes, and partners. To succeed, companies need the right tools. This transformation is already underway with automated pattern recognition software that uses better math and more data to sense demand, without relying on the assumption that history will reoccur.

In contrast, traditional forecasting engines use time-series methods founded on the premise that history does indeed repeat itself. This is rarely the case, even for stable products with long shipment histories. Promotions are effective at driving sales, but introduce complexity by distorting past shipment distributions. Innovation in the form of new products has no history whatsoever, which is problematic for time-series techniques since they require a minimum of two years data for the most basic analysis. 

With up to 50 percent of items for consumer products companies having less than two years of history, this can be challenging. Even small product changes like a new colour, scent or package size require labour-intensive mapping to get meaningful results.  Additionally, changing consumer preferences in response to economic pressures, new mobile technologies or weather-related disruptions like Superstorm Sandy all combine to make last year’s shipments a poor predictor of future sales. 

 Automation is the answer

Automation is at the heart of any system that senses demand. The sheer scale of data at multinational manufacturers quickly surpasses human capabilities to process, let alone comprehend.  For example, a manufacturer selling 500 items through Walmart’s 3,800 stores receives almost 2 million new data points every day through Retail Link – just for its North American business with Walmart.

This requires an army of planners to manage if forecasts are updated on a regular basis. More often, planners focus their limited resources on important fast-moving items and high-profile promotions because they don’t have time to do everything. Findings from an annual forecasting benchmark study,encompassing one third of consumer packaged goods volume in North America, confirm this reality.

Forecast error for the slowest-moving products (which represent 80% of all items but only 20% of volume) is 1.6 times higher than for the fastest-moving products (2% of all items representing 20% of volume).  While understandable from a human resource perspective, this prioritisation creates less than ideal business results considering the large capital investment in inventory for 80% of a company’s items. Addressing this gap is a clear opportunity for automation. 

Unlike their human counterparts, software that senses demand using automated pattern recognition algorithms is particularly well-suited for the computationally-intense and mind-numbingly repetitive tasks of crunching data and finding correlations for the entire business. 

At some of the world’s largest companies, near-term forecasts are updated daily and published directly into supply planning systems to adjust manufacturing and deployment in time to respond to unexpected lifts or downturns in markets.  For long-term horizons, automated pattern recognition provides companies with a better starting point for their demand planning process. 

Its robust statistical baseline removes the need for planners to individually tune and curve-fit each model, reapply forecast adjustments from prior reviews and account for mathematically-complex interactions like promotion cannibalisation or similar item matching.  This frees planners to focus on important business areas that require human knowledge like strategic planning and promotional activities.

By upgrading demand prediction software to do what machines are capable of today, companies are breaking free from the traditional time-series methods that have held back industry in recent years.  It is an important step towards realising the promise of the digital supply chain and the future of demand planning – paving the way for better planner and S&OP productivity, better cash flow and return on capital with less unnecessary inventory and the competitive advantage of an agile supply chain that can quickly respond to unexpected market opportunities.