Skip to main content

Tech Guest Viewpoint: Avoiding Time-Series Demand Forecasting


Silicon Valley’s Winchester House confounds visitors from around the world. An eccentric heiress spent decades adding endless rooms and hallways, doors that lead to nowhere, and random structural additions — turning a home into an inscrutable, imposing and meandering oddity.

It’s a prime example of add-ons gone wild — and a good analogy for the state of retail forecasting today. Many of today’s retail forecasting systems have been added to and tweaked over decades until they have become imposing and complex, leading to a burdensome, difficult to maintain and a vastly different expression of their original intent — to provide retailers with accurate forecasts that lead to greater insight on the drivers of consumer demand, improved sell-through, successful promotions, and greater profitability.

The time-series forecasting methods still used by most retailers today date back to the early 1960’s. Over time, generations of developers added here and tweaked there, and today the science that data analysts rely on is often older than they are. Despite decades of effort, the cost exacted on retailers is sobering: IHL Group estimates that retailers worldwide lose $220 billion annually due to inaccurate forecasts.

The time-series model hinders accurate forecasting in several ways:

• Time-series forecasts degrade over time. As scientists add on and manipulate these traditional forecasts, the models can exponentially degrade and become increasingly inaccurate and bulky. As a result, it becomes difficult to harvest high-performance output without requiring a great deal of data scientist input, turning an automated process into one that’s highly manual.

• Traditional forecasting methods are unsuited for new products or stores. Time-series forecasting looks into the past and projects into the future, so what do retailers do with new products and stores? Because so many retailers are centered on a time-series approach, they’ve adopted a “like SKU” or “like store” tactic in an attempt to identify the history of a similar past SKU or for other markets, but what happens when you introduce a completely new category? or a new set of attributes? or enter a new market? or when (inevitably) fickle consumers change? Time-series methods just don’t cut it, while machine learning-based forecasters that ask for a forecast on a new product, get it — naturally and with no need for specific algorithms or user input.

• Outdated time-series models constrict retailers’ abilities to forecast at a granular level. Situations such as new stores or SKUs, complex promotions and regional events are highly relevant, frequent in general yet infrequent in terms of specific scenarios, and are beyond the abilities of a time-series based model. Retailers that want to shape demand and accurately forecast promotional activity or do so at a local market level can expect a cumbersome output that requires their data scientists to spend time manipulating the inputs and outputs, with consistently poor results.

Clearly, it’s time to tear down time-series forecasting, and make room for a new paradigm to take its place. Machine learning technology, the modern underpinning of the innovations used by Netflix, Amazon and Google to understand their customers at great depths — has the power to transform demand forecasting in retail by setting better prices and creating better promotions.

Machine learning uses causal factors such as product descriptions, competitors, promotions, regional events, and trending styles and colors to determine exact demand drivers based on the many attributes involved, for every possible SKU-store combination. And if an attribute provides no value or inherent impact on demand, the machine learning forecast models are smart enough to discard it where it is irrelevant, and to use it when it offers value in a different situation. There is no such thing as too many factors with machine learning models.

Don’t let your retail enterprise get lost in its own Winchester House maze. Machine learning-based forecasting delivers a new and better “house” — where they can understand and act on the true drivers of demand and enter the 21st century of forecasts with accurate context that drives specific and actionable insights.

Ron Menich is Ph.D., EVP and Chief Scientist, Predictix.

This ad will auto-close in 10 seconds