The case for integrating artificial intelligence and machine learning across Supply Chain.
Supply chains rely on accurate, evidence-based demand plans to operate efficiently. Achieving, or nearing, forecast accuracy benchmarks is a significant milestone for most organisations. But in an increasingly volatile and interconnected market, the real question is: is hitting the benchmark enough?
Forecasting in Demand Planning: Are we doing enough?
Demand planning sits at the core of supply chain performance. It helps inform replenishment planning, production scheduling, inventory management and the allocation of working capital. When demand plans are right, shelves stay stocked, delivery commitments are met and service levels remain high. When they’re wrong, the consequences can ripple across the entire Supply Chain. Effective forecasting is therefore critical. The forecast algorithms that underpin demand models play a central role in mitigating risk and enabling long-term operational success. This is by ensuring customer demand is met in the most cost-effective way possible.
Many organisations already use forecasting techniques within their planning landscape. But an important question often remains unanswered: are those methods the right fit for every product, channel and region? And are they capturing the demand signals that truly matter?
What is Forecasting in Demand Planning?
Forecasting in demand planning is the practice of predicting future sales or demand using historical data, related data and business context (subject matter experts, product owners and marketing teams). The goal to achieve is simple – achieve the most reliable demand signal possible, enabling organisations to position the right inventory, capacity and working capital to meet customer demand.
In theory, forecasting can appear to be a straightforward task. A demand plan can be created by reviewing past sales, creating a trend line and projecting it to future periods. However, demand signals can be noisy, irregular and influenced by several factors which aren’t discernible on a graph or simple algorithms.
What techniques have been historically used to predict?
Before exploring the advanced techniques increasingly used by leading organisations, it is useful to revisit the foundational forecasting approaches that have shaped demand planning for decades. Many of these methods rely heavily on historical data and relatively simple statistical algorithms to generate predictions. While they remain widely used, their effectiveness can vary significantly depending on the nature of the product, the volatility of demand and the complexity of the market being served.
Traditional forecasting in demand planning has depended on well-established statistical techniques designed to generate forecasts using limited inputs. Approaches such as moving averages, exponential smoothing, and proxy or analogue forecasting have become widely adopted due to their ability to be implemented quickly whilst maintaining transparency. While relatively simple, these methods provide some effectivity in achieving accurate forecasts, due their ability to identify simple trends, for example:
Moving Averages:
- Useful for products with stable demand
- Short-term forecasts where demand volatility is low
- Treats each product individually and only analyses its historical volume
Exponential Smoothing:
- Demand with trends or seasonality components
- Effective where historical data has clear demand signals
- Reacts to changing patterns in trends
Proxy / Analogue:
- Use an existing products historical volume as the forecast for a new product
- Handle New Product Introductions (NPIs) with ease, even with limited or no historical data
- Useful when expanding into new regions or channels
- Products with strong similarities to existing SKUs
While these approaches remain valuable, they demonstrate clear limitations in today’s increasingly complex demand environments. The underlying algorithms mentioned are robust but relatively simple, relying primarily on historical demand patterns and limited statistical relationships. As a result, they fail to capture more complex signals, shifting consumer behaviour, non-linear trends or interactions between multiple drivers across regions, or channels. This can mean some organisations are giving ground to competitors by not only inaccurately capturing demand, but also not truly understanding the nature of how their products perform. The good news is that modern tools allow us to capture these more complex trends or patterns than simple models may miss – and this can be automated to reduce time required from demand planners allowing them to focus on what really matters.
What are Supply Chain leaders using to predict?
As supply chains become more data-rich and markets more volatile, relying solely on these traditional forecasting methods can leave valuable insights undiscovered. This not only impacts supply planning but can seep into other areas of the business such as Finance, where inaccurate demand plans could materially impact revenue forecasting. The question then becomes: how can organisations build on their existing forecasting processes to uncover deeper demand signals and generate more reliable forecasts?
Advances in forecasting techniques help organisations address this challenge by introducing machine learning, a component of artificial intelligence, into the demand planning process. ML approaches improve demand forecasts by identifying complex patterns in historical and related data, unlike simpler traditional methods. Some of these approaches are outlined below, together with explanations of how they address limitations of more basic methods:
SARIMAX:
- Useful for demand with strong seasonal patterns and clear historical structure which have repeated patterns over time (weekly or yearly cycles)
- Incorporate external variables such as promotions, pricing and macroeconomic drivers
- Looks at past trends and patterns to predict what might happen next, whilst considering external influences
- Thinking of how this could fit in your product portfolio? Here’s some examples where this algorithm may be useful in particular:
- Retail products which experience increased demand during holiday periods (impacted by promotional drivers)
- Airline ticket sales (impacted by fuel prices and holiday drivers)
- Grocery products (impacted by seasonal holidays, weather and discount cycles)
LightGBM:
- Able to handle large numbers of products which are influenced by a few different factors at the same time
- Analyses how variables interact with each other at any given time to impact demand for products
- Shows which factors matter most for forecasts, helping you understand what’s truly driving your demand
- Thinking of how this could fit in your product portfolio? Here’s some examples where this algorithm may be useful in particular:
- E-commerce where volumes may be impacted by a combination of price, competitor activity, advertising spend and seasonality
- Fashion items where demand is influenced by promotions, trends, pricing changes and world events
TimesFM:
- Use a model pre-trained on large scale datasets which can handle products with varied characteristics
- Generate quick, accurate and reliable forecasts even with limited historical data
- Useful for analysing large portfolios across multiple markets, channels or regions
- Thinking of how this could work for your product portfolio? Here’s some examples:
- New product launches such as own-brand items with little to no actual volumes, but like many existing products where patterns can be shared.
- Long tail marketplace products, which may display low historical volumes, but collective wider market patterns help improve predictions
DeepAR+:
- Utilises components of Machine Learning (ML), a subset of Artificial Intelligence (AI), to learn patterns across the entire product portfolio, not just one item at a time
- Performs well even when data is limited, inconsistent or noisy
- Produce a range of possible demand outcomes (not just a single number), helping you understand uncertainty and plan for different scenarios
- Thinking of how this could work for your product portfolio? Here’s some examples:
- Pharmaceutical products where demand can be irregular, data can be sparse for SKUs and understanding uncertainty is critical for avoiding shortages.
- Subscription based services where demand can fluctuate and planning requires an understanding of a range of possible outcomes.
- Electricity or fuel demand which can be influenced by many related time series such as households, substations, or cities.
Historically, these advanced forecasting models required specialised infrastructure and data science expertise, making them difficult for many organisations to adopt. Today, advances in cloud technology and integrated planning platforms have significantly lowered these barriers. As a result, machine learning algorithms can now be embedded into existing demand planning processes far more easily.
Building a modern demand planning process?
In this article, we have outlined two key capabilities that modern demand planning solutions must support to enable a more effective and responsive supply chain process:
- Leveraging established forecasting methods and planning processes that organisations already rely on.
- Incorporating advanced machine learning techniques to analyse data more deeply and uncover demand patterns that traditional models may miss.
The challenge for many organisations is not recognising the value of these capabilities but understanding how to introduce them without disrupting existing planning processes, governance structures or operational workflows.
In our next article, “Anaplan & Demand Planning”, we explore how organisations can bridge this gap by examining practical use cases for machine learning algorithms in forecasting and how Anaplan can help connect demand planning with the wider supply chain to create a more responsive, data-driven planning process.