By Erik de Vos, Bregje van der Staak & Wout Olde Hampsink
In today’s globalized market, companies face complex challenges in managing their supply chains and ensuring seamless operations across diverse regions. Accurate forecasting has become a critical factor in driving efficiency, reducing costs, and improving decision-making. Demand for products and services is often heavily influenced by various internal and external factors that result in an up- or downward shift of demand. By incorporating these factors into a machine learning model, these companies can achieve more accurate forecasts, allowing them to anticipate demand fluctuations, optimize inventory levels, and align their end-to-end value chain.
Machine learning is all about using machines to improve your demand forecasting capabilities. So how do you set up the machine to do part of the job for you? Is it as easy as organizing data and go?
The challenge
Machine learning in demand forecasting, where to start
In previous blogs, we have discussed the importance of data availability for your forecasting approach. Having data quality processes in place is the first step, but how do you get from data to insights and then to a solid forecast?
The challenge of extracting valuable information from your data is underestimated, and there are often misplaced expectations about which specific data elements will add value to your forecasting models.
Furthermore, an incorrect focus on specific data elements often leads to the wrong behavior in model optimization. It is a misconception that machine learning models can only learn correctly from an information overload of input. Accelerating your forecasting approach with advanced models ultimately leads to the following challenges:
- What are the drivers that affect the variability of my demand? Having the data is one thing, but finding the right predictors for your forecasting engine is another.
- How to fit in your predictors and optimize your forecasting model. After finding the best predictors, fitting them in to optimize your model is the second step, with the risk of overfitting.
Let’s take the example of finding the best predictor for your model. If you want to predict ice cream sales in the summer, it seems logical to include the weather forecast in your forecasting model. However, the weather forecast is only accurate enough for the short term, which does not allow you to make purchasing decisions for the long term. You would need a different weather-related data set to potentially add value to your long-term forecast.
Consequently, challenging your data with assumptions that are only valid for a subset of your scope will result in overfitting of the model, while leaving important assumptions out, will result in underfitted forecasting models. Ideally, you want to validate your model on an unseen data set to see which approach adds the most value.
The product
EyeOn Proof of Concept for accelerated improvement
When the key challenge is to identify and validate the value of data to your business, the EyeOn Proof of Concept (PoC) comes into play. The PoC approach is a three-step process designed to optimize the information contained in your existing data:
- First, we start with an intake and data discovery step to validate and understand your key demand drivers and detect where there are connections to driving your forecast.
- The second phase involves iterative machine learning modeling to determine the impact of drivers on forecast accuracy and bias. During this step, model parameters are also evaluated and optimized.
- The final step focuses on the advice on how to proceed with your digital transformation to unlock more value of the data with machine learning.
When your forecasting process is ready for the next step, EyeOn Proof of Concept will help you determine the right direction with an extensive elaboration of your forecast drivers. Also, you will get quantified confirmation on the forecast potential with the following deliverables:
- Insights into the key levels for forecasting and the importance and usage of demand drivers
- A detailed and tested performance benchmark of the machine learning demand forecast against your current forecast
- An operationalization approach (roadmap, work packages, go-live approach)
Appropriate for
Machine learning to level up the demand forecasting game in various situations
With the EyeOn Proof of Concept, we take the next step in forecasting automation: what are the factors that affect your business and how do you automatically account for them in your forecasting engine? This is especially beneficial in cases where:
- You have several known reasons to manually enrich your forecast
This is a good starting point to investigate which factors influence your demand as these can then be used in your forecast engine in a more automated way of enrichment. With the same data and less effort, you may potentially be able to achieve higher forecast accuracy when using machine learning.
- You have a lot of business data that you want to use for forecasting
Do you recognize the following: you have more data than just your demand time series. If so, there is an opportunity to explore the potential of these data elements for your demand fluctuations. There may be more valuable information in the additional data that you can unlock for forecasting.
- You want to improve your forecast accuracy with new strategies
Your company uses a statistical forecasting process and is ready to evaluate the power of machine learning. You have discovered that part of your portfolio has volatile time series that are no longer correctly captured by standard classical statistical models, and you want to see what machine learning can do for these challenging products.
- You want to level up your forecasting capabilities
Your company is ready to take its forecasting methods to the next level, but you do not know if your data is ready for this step. You need an assessment of the potential of your demand driver data and a strategy for the next steps in your development.
Improving your traditional forecasting setup requires attention, especially in gathering the right additional data to improve the forecasting engine. Our EyeOn specialists can help you identify potential internal and external drivers of your forecast. Our in-depth market knowledge allows us to focus on the best potential drivers in your industry. With an optimized set of drivers and a tested, improved model, our final step is to define an operationalization approach to implement the new engine.
Results
Focus on the best models
Demand forecasting with machine learning can accelerate your forecasting performance in many cases, but it is important to select the best model for each individual time series. In many cases, like when sales are stable, statistical methods outperform advanced machine learning techniques. In more complex situations, for example when external factors are involved, machine learning techniques do outperform statistical models. It all comes down to the right combination of models. EyeOn’s proof-of-concept approach confirms which model is best for each time series. The following examples give a clear direction of our proof-of-concept capabilities:
- For a company in the vegetable industry, our proof of concept analyzed the use of machine learning as a complement to standard statistical approaches to improve performance in selected time buckets. After several discussions with their experts, it became clear that promotional information was available in different time buckets. The use of exploratory data analysis convinced us that there was potential to apply machine learning to selected periods of the forecast horizon. Ultimately, the results of the PoC showed the performance gain with an ML approach, which is now being used successfully.
- Another example of EyeOn’s proof of concept was a large pharmaceutical company that was able to gain efficiencies by automating forecasting for a large portion of its portfolio. Previously, planners were working with one set of market information as input for their manual forecast enrichment. Together, this set of market information was transformed into different features and used as input for a machine learning model. The results of the PoC assessed the portfolio where efficiencies could be gained through automation, freeing up time to focus on business improvement.
- Finally, for a company focused on juvenile products, the introduction of demand drivers in the PoC was analyzed to improve forecast accuracy and accelerate their digitalization journey. In evaluating the results, it became clear that not all drivers were effective in improving forecast accuracy. In fact, the machine learning model with only a selected group of drivers resulted in the best forecast accuracy. In fact, the challenge is to find the right subset of features that add value to your forecast, and then organize yourself properly to unlock the data and the value it brings to your forecast.
EyeOn proof of concept in a nutshell
The EyeOn Proof of Concept provides insight into the drivers of your demand. It helps you take the next step in your smart-touch automation journey.
- Get quick insight into the main demand drivers and how they contribute to the forecastability of your business
- Underpin the connection between data, techniques, and the forecast quality you could achieve
- Identify areas for development to unlock full proof of concept potential
Are you ready to discover how machine learning can help you take the next step in forecasting automation? The EyeOn Proof of Concept identifies what is (not) driving your demand. It also automates the incorporation of your relevant drivers into your forecasting engine. It reveals the highest possible forecast accuracy that can be achieved by combining statistical forecasting (for predictable products) and machine learning (for unpredictable products), and identifies key opportunities for improvement.
Are you ready to take your forecasting insights to the next level? Discover more about our Proof of Concept or get in contact with our forecasting specialists Wout or Bregje.