Did you just look back? Supply chain management, Artificial Intelligence & Forecasting
Have you ever heard these comments about forecasting in Supply Chain:
- “Forecasting in our context is impossible, and we don’t want to invest in developing it right now”.
- “We just get too poor input from other functions, so we simply copy past data for the future”.
- “We cannot take the risk of trusting sales plans – We have our own static way of dividing annual budget into product level forecasts in SCM.”
These are more than typical first responses, and often the starting point of forecasting project initialization with our customers. The comments reflect well the challenges of data-driven decision-making and can be divided into at least three categories: Vision, trust, and technology. These are all connected, but especially the technical tools category keeps bringing new possibilities at an increasing pace.
The world of Artificial Intelligence (AI) is touching us all. It is both transforming industries but also changing individuals’ work. Besides all the ongoing marketing speech of limitless AI opportunities, many businesses have started to lean forward to crystallize their plans and take concrete actions to introduce new technology systematically.
The same applies to the supply chain management (SCM) context too - AI is not only a future concept. By combining today’s tools and technology, we can support SCM area interests in many ways, varying from building global strategies to boosting the efficiency of a single process in a dynamic environment.
When it comes to applying AI successfully as an integral part of business processes, it is easy to place too much attention on technological capability development alone and leave the organization and people-related questions unanswered. At best, AI can be seen as an integral part of a comprehensive transformation program for companies, where change management and cross-organizational collaboration are fundamental parts of success.
Forecasting and AI in real life
For years, forecasting as a theme has taken leaps forward as machine learning algorithms and AI models develop. Conventional simple statistical methods, while useful in certain circumstances, may not anymore suffice to keep up with the fast-paced and complex nature of modern supply chains. The rapidly changing environment and recent technological advancement provide interesting opportunities and a window for development actions.
Where well-designed forecasting AI applications stand out, is their ability to crunch large volumes of data, combine diverse information and identify significant signals. Especially in the SCM context, a lot of transaction data can be collected. Gathering big data – that is more data that can be dumped into memory at once - is very easy.
However, deriving the signals of information from that data requires a well-defined plan and investment in the data architecture, analysis, feature engineering, model building, and many other steps. Getting to the point where the information is trusted within the organization, brings another layer of complexity. In many AI projects, the most challenging part is to build an efficient and consistent feedback loop between AI developers and business stakeholders. AI tools are rarely a one-off development sprint but require constant iteration and development.
Often the best way to start introducing AI into the forecasting process is to map simple univariate models, and gradually enrich them with high-quality data sources and more advanced models or stacked models. Due to constantly increasing computational power and developing AI models, the additional data sources can enable more and more complex linear and non-linear insights, picking the signals from multiple domain-specific data sources.
Data is the key, but so is the models’ capability to pick the relevant signals too. Usable data sources may vary from market trends, indexes, and weather forecasts to solely enriching data from internal process insights. Often the best results follow from a combination of multiple cleansed and well-combined data sources. However, finding the right signals is not easy. It does not happen automatically but requires a human touch starting from business- and context understanding.
Building forecasting capabilities
When building new capabilities and increasing AI's influence in forecasting, it most often can be seen as a gradual development program with multiple connected projects, that all together enable increased accuracy in each step. Although a hundred percent accuracy in predictions is often impossible to achieve, today’s technology provides tools to significantly narrow down the error. Also, AI-assisted forecasting unlocks the potential to quickly draw a broad set of future scenarios that in turn help with preparing for the future and speed up decision-making.
Below is a demonstration of how initial targets for forecasting accuracy development can be set. In the example, the forecast error has been +-10% from the realization, and a sensible goal to set would be to start running models that aim to reduce the error and develop the model step by step.
A variety of algorithms and models exist that could be applied and stacked to process the data to reach the targets described above. For someone starting from scratch, it can feel like an overwhelming task to get a grasp of the strengths and weaknesses of each of them. If you just set a good-sounding ML model running, e.g., autoregressive integrated moving averages, it naturally takes you forward on the learning path, but the problem is that this kind of application, placed over uncleansed raw data sources, typically alone does not bring you closer to the goals. A big risk is that it will lack your business DNA and have a huge underlying inbuilt bias, which causes frustration and does not help to get closer to the goals.
To be able to build useful tools, the right steps before applying any model or technique should be taken. No matter whether it is the lack of vision and targets, too low level of business or technical understanding, low quality of data, poor internal communication, or project management, the result will be poor.
A well-run AI program typically surfaces many new development possibilities too: how to cure the root causes that lead to forecasting accuracy errors? Sometimes the answer lies in the network’s transparency, common internal stakeholder and external partners’ game rules, or adding in information system capabilities, such as automation.
Conclusion: The Time to Lean Forward is Now
Integrating AI in SCM is a pivotal shift, not just a hype or temporary trend. Those who take AI in their SCM development roadmap will be the leaders in the years ahead. It's not just about adapting; competitiveness requires being at the forefront of change. With AI, the future of SCM is not only exciting but also full of possibilities.
Despite the high technical learning curve and the need to challenge traditional ways of thinking when applying AI to business processes, the reward is high: a future-proof organization, processes, and tools. Because the rabbit holes are deep in the domain of AI, finding the right-minded and skilful partners to speed up your progress and avoid the typical pitfalls in transformation programs and technical applications will be a big help.
When we carry out AI projects with our customers, we promise to provide a unique set of business understanding and technology expertise, which we believe is the winning concept. Should you ponder your next step with working on forecasting process development, SCM development, or AI development overall, please contact me and my colleagues at Knowit - Let’s discuss more!