The future of demand planning

Can we predict the future? For centuries, humans have tried, and we're finally getting better at it. As machine learning and artificial intelligence not only evolve, but begin to replicate human cognition, we will be able to see into the future as never before. And for supply chains, that's going to revolutionize demand planning.

Many factors influence variability and volatility, which are two critical measures for understanding demand patterns. Shorter product life cycles and high SKU proliferation are adding complexity to unpredictable markets. New ways of shopping pop up virtually daily, along with corresponding omnichannel marketing efforts. And as consumer behavior becomes more fluid, promotions must become more specific and targeted in order to influence buying decisions – and that, in turn, contributes to demand volatility.

New technologies like the internet of things (IoT) and point-of-sale tools are helping us capture, harvest, and harness masses of data and information. But if collating and correlating this material takes a great deal of effort, making sense of it is an even bigger challenge. Traditional forecasting techniques simply don't cut it in this context.

That's where artificial neural network (ANN) models come into play. These differ significantly from traditional forecasting methods because they don't rely solely on a demand planner's expertise and experience to identify relationships and patterns in variables. Instead, they self-learn using observational data to identify regularities, relationships, and patterns between variables.

ANN models replicate one of nature's most complex and agile creations – the human brain. Much as our brains analyze electric impulses from our eyes, ears, and other organs to make sense of the world, ANN models evaluate multiple input variables to deliver output. The process has four key components (Figure 2):

1. Input signals (independent variables) are received by neurons in the form of signals originating from a source or from other neurons
2. Synapses (weightage) transfer the input signals to neurons
3. Neurons (activation function) receive the input signal depending on its strength (weightage) and process it
4. Output signal (dependent variable/demand forecast) process information from neurons

Previous demand forecasting tools based their predictions on the theory that historical drivers and patterns will repeat – and make a similar impact – in the future, based on the assumption that data is linear. For demand planners using these causal methods, the challenge is to identify factors that influence demand and build correlations between them.

By comparison, ANN models:

• Thrive on input variables: ANN model accuracy increases with every external and internal factor influencing demand that is fed in, resulting in a better grasp of demand variability
• Process non-linear data: The models combine multiple input variables – promotion duration, promotion type, competitor pricing, and more – to better evaluate volatility
• Measure the impact of special events: They take promotions, supply crises, and other one-offs into account
• Unearth, build, and model complex relationships: They identify complex patterns, trends, and significant relationships – for instance, how customer reviews for a product in one category influence sales in another
• Automatically learn and adjust with time: The more data input, the better ANN models get at assessing relationships between independent variables

Figure 3 shows the approach for developing ANN models, which depends in part on a company's digital maturity.

What can ANN deliver?

The results from implementing ANN models into demand planning are pretty impressive:

• Up to 40% improvement in forecast accuracy
• Up to 70% productivity benefits from touchless forecasts with fewer forecast exceptions, less need for data consolidation, and zero feedback required from sales and marketing teams

The future for ANN

The transition to more advanced demand forecasting techniques like ANN has already begun in the consumer packaged goods industry. So far, implementations have been piecemeal, specific to particular product categories, geographies, or markets, and undertaken in partnership with specialist machine learning companies. But in the foreseeable future, the success and value associated with introducing ANN and similar models, as well as the development of a subsequent ecosystem, will ensure high adoption rates across a wide range of industries.