ML for Demand Planning: Breaking Old Habits

Changing rooted behaviors is one of the hardest jobs for leaders. And what about trying to bring innovative technologies and novel approaches to accomplishing tasks tied to those rooted behaviors? Well, that's a lot like pulling teeth. But those changes pay off, for they position the organization to reach higher levels of value.

This blog post introduces machine learning for demand planning by looking at how car safety relates to demand planning, and exploring the obstacles organizations might experience with adoption and how to overcome them.

How Car Safety Relates to Demand Planning

When thinking about car safety, most people likely picture one or more of the images below:

Volvo first comes to mind first. And indeed, Volvo pushed for standardizing the 3-point seatbelt in the car industry, so the company can proudly take credit for saving millions of lives [1] since the 1970s. But adoption didn't come easily. Why not? Well, it's simple. Safety belts meant a radical change in people's habits with a difficult trade off: losing comfort for safety in the supposed event of an unlikely accident [2]. Using safety belts meant a behavioral paradigm shift for drivers and passengers alike that even today faces resistance from some motorists. And to make true change possible across the industry, Volvo had to open up the patent for safety belts to competitors to accelerate adoption.

Similarly, just like drivers and passengers with the introduction of the seat belt, demand planners are going through a behavioral paradigm shift with the introduction of machine learning (ML). Why? Because old habits die hard! Let's dive into it.

Demand planning work is usually a manual activity grounded in a low-accuracy system-generated baseline. To get the accuracy right, several external and internal source inputs enrich and adjust this baseline multiple times. Spreadsheet wrangling, reconciliation and error are thus inevitable fallouts of this approach. Yet many planners prefer it. Why? Because they feel comfortable with it. That also means they refrain from learning new technologies and methods despite – as is the case with ML – the game-changing benefits. Using machine learning for demand planning drastically improves accuracy and exponentially increases the number of forecasts run.

Fortunately, not every organization is resistant to changing the status quo. Some companies are trailblazing the adoption of machine learning to improve forecasting accuracy in demand planning. One such company is Autoliv, a tier-1 automotive supplier of safety components for major carmakers in the world.

Better Demand Planning Fuels Profitability

Supplier relationships in the automotive industry are based on a pull system that gives car manufacturers strong leverage on pricing. Consequently, margins can be razor thin for suppliers, and the risk of falling into loss is high. This axiom is valid in other industries as well, including in transportation, retail, wholesale, consumer products, and more.

While margins can be improved in many ways, a robust approach is needed to better understand and plan demand effectively. Why? Because an organization that appreciates the business drivers that shape future demand can better draw sales projections. Additionally, the organization can better adjust inventory levels, avoiding stock outs and breaches of service levels.

Autoliv offers a good real-world example of putting this robust approach in action. How? The company successfully embarked on a transformation journey to have a single view of profitability across Sales, the Value Chain and Finance. Autoliv also knows that – in the automotive industry – understanding demand is key to protecting and growing profit margins. Accordingly, the company is exploring the use of Sensible Machine Learning to improve demand planning. You can read the Autoliv case study here.

Dodging Obstacles Along the Way

Many organizations are using artificial intelligence (AI) or machine learning (ML) in the business in one way or another. One of the most used approaches is to build a data lake and apply ML algorithms. However, this approach does not always work well for planning use cases. When dealing with ML for demand planning, organizations may encounter the following challenges that hinder adoption:

• High complexity, low generalization. Many purpose-built applications in the market are complex. They often require additional programming skills and the transfer of sensitive data outside the module. Conversely, in-house applications are highly customized to only serve a specific use case. As a result, when business conditions change (and they change a lot!), in-house applications must be re-programmed.
• Lack of talent, lack of focus. Data scientists have one of the most sought-after skill set in the job market, in any industry. So they're not only expensive profiles, but also ones that are hard to find and retain. Often, resident data scientists work on a wide variety of use cases. The problem with such positions is that the data scientists lack the business context needed to build solutions, without tedious interaction with functional roles.
• The black-box effect. Demand planners often perceive ML planning solutions as a black box. Planners get the results from the algorithm but know little about how the solution handles the data – perceiving lack of transparency that ultimately diminishes trust in the results.

Not to mention, resistance to change can be high, and only 13% of standard ML projects make it into production. What's the point of producing an ML forecast that no one uses? [3] Luckily, there is a way to deal with the obstacles along the way.

Keeping the Eyes on the Road

Having the expected benefits clear from the start is key when considering machine learning for demand planning. Does it need to underline new patterns? Should it address variability? Can it produce a high volume of forecasts at speed? The ultimate litmus test is that the solution delivers more forecasting accuracy and that planners are trusting it.

Many organizations hold a vast amount of data, but it is pocketed in different systems and databases. When a lot of effort goes into preparing the data for the ML engine, organizations may lose sight on what's important. A solution that can ingest volume and disparate datasets is therefore key for demand planning use cases. For that reason, the following key attributes must be considered when looking for an ML solution for demand planning:

• Automated set-up. Data quality, ingestion and preparation should be easy, so the laborious tasks done by data scientists are reduced or eliminated.
• Responsive to market trends. The solution must be able to adjust quickly to demand and supply signals.
• Better accuracy. The initial pilot implementation of Sensible ML for Autoliv improved forecast accuracy by 7% compared to the controllers' manually adjusted forecasts.
• User adoption. Users get full transparency on how the solution produces the results and can use these results quickly to perform analysis and gain insight.
• Responsive to market trends. The solution had to be adjusted quickly to demand and supply signals.
• Scalability. The solution can produce hundreds if not thousands of forecasts with speed and at a fraction of the cost, augmenting the controllers' capabilities.

Beyond considering the necessity of the above attributes, organizations must also stay focused on the business outcomes they expect.

It Is Not the Destination But the Journey

When an organization has clarity on business outcomes, ML and other technologies become an enabler to accomplishing those outcomes. This clarity on goals helps organizations decide between costly and lengthy home-built ML solutions and market-leading planning solutions with built-in ML services. The latter will help better break old habits, enable enterprise-wide adoption and accelerate the time to value.

Learn More

Ready to find out how to break away from old habits on demand planning?

Discover OneStream's Sensible ML for Demand Planning, the only solution that helps demand planners and finance teams embrace machine learning with trust and full transparency into the data and results.

Download the Solution

[1] CDC Road Traffic Injuries and Deaths—A Global Problem[2] Read Volvo's amazing story here[3] VentureBeat. Why do 87% of data science projects never make it into production?