David Hoyle, Research Data Scientist Specialist at dunnhumby, discusses how Operational Research can help data scientists move from prediction to better decision-making.
DAVID HOYLE, RESEARCH DATA SCIENTIST SPECIALIST AT DUNNHUMBY
In a recent webinar for The OR Society, David Hoyle, Research Data Scientist Specialist at dunnhumby discussed the intersection between data science and operational research (OR) through the lens of grocery demand modelling at dunnhumby.
Data science has transformed how organisations understand the world. From predicting customer behaviour to forecasting demand, data scientists are skilled at building models that tell businesses what is likely to happen next. But increasingly, that’s not enough.
Organisations are no longer satisfied with insight alone. The real value lies not only knowing not just what will happen, but what action to take. This is where a large gap exists that Operational Research (OR) is uniquely positioned to fill.
OR uses mathematical modelling and optimisation to support better decision-making. Where data science often focuses on prediction, OR focuses on prescription and identifying the best course of action within a complex set of constraints.
As businesses move toward automated, real-time decision-making, that distinction is becoming important. If data scientists want to remain effective in a world increasingly driven by automated decisions, they must move beyond prediction and embrace optimisation techniques.
The complexity behind demand
dunnhumby is a global customer data insight company helping retailers and brands perfect the science of shopping. Working with clients including Tesco and Walmart Data Ventures, dunnhumby helps businesses turn customer data into insights that drive growth and improve customer experiences. At the heart of this work lies demand modelling, a cornerstone of retail decision-making.
For a large retailer like Tesco, even small improvements in sales forecasting can deliver significant commercial impact, yet demand is shaped by a complex interplay of factors, from price and promotions to seasonality, marketing activity and competitor products.
Increasing the price of a product will typically reduce sales, but the scale of reduction depends on how price sensitive customers are. A promotion may boost sales, but some uplift may be borrowed from competing products, a phenomenon known as cannibalisation.
Layer on top regional differences and shifting consumer trends and the modelling challenge becomes even more complex.
Retailers manage thousands of products across multiple regions, customer segments and price zones. Data quality is rarely perfect, and models must cope with everything from predictable seasonal patterns to sudden structural shifts, such as viral Tik Tok trends or new product launches.
Importantly, these models must work in the real world. They must be scalable, automated and continuously updated, running reliably in production without constant human intervention. When decisions are being made at scale, there is little tolerance for fragile systems or human fixes.
Why prediction isn’t enough
Even the most advanced demand model is only the starting point as the real value comes from what it enables. These models can reveal how sensitive customers are to price or allow businesses to test “what if” scenarios, exploring how different strategies might play out. They can support forecasting and planning, helping retailers manage inventory and supply chains more effectively.
But insight alone does not drive value. Decisions do.
This is where Operational Research comes in. By combining demand models with optimisation techniques, businesses can move from understanding outcomes to shaping them. Instead of asking what will happen, they can ask what they should do. This is a fundamental change in the role of data science. It is no longer just about generating insight, but about enabling better decisions at scale.
The rise of automated decision-making
Despite the growing overlaps OR is rarely part of a data scientist’s toolkit. Most are trained in statistics, programming and machine learning, but not in optimisation or decision science. Many become accidental practitioners, picking up techniques as they go without a broader framework.
Even for dunnhumby, which has been building expertise in integrating retail data science and OR over many years, there are still opportunities to improve processes. However, for many organisations, this division of labour, with data scientists building models and OR specialists handling optimisation still exists. It has been workable in the past, but is increasingly unsustainable.
Businesses are rapidly moving towards closed-loop, model-driven decision-making systems. In these environments, data is continuously collected, analysed and fed into algorithms that make decisions, often automatically. Those decisions are fed back into the system, allowing it to learn and improve over time.
This is happening in pricing, supply chains and personalised marketing, and the pace is accelerating with advances in AI. In this context, the line between prediction and decision-making begins to blur. Models are no longer passive tools that describe the world; they are active components in systems that shape it. This raises the stakes.
If a demand model feeds directly into an optimisation engine that sets prices in real time, then any flaw in the model or the optimisation framework can have immediate and significant consequences. Just as importantly, the way a decision problem is defined such as the objectives, the constraints, the trade-offs, can be as influential as the accuracy of the predictions.
Yet many data scientists that work inside commercial organisations are not trained to think in these terms, and that gap is becoming a business risk.
From model builders to decision designers
For now, there is often still a human in the loop. Algorithmically generated decisions may be reviewed and approved before being implemented. But this safety net is at risk of disappearing from within organisations. As organisations demand faster, more responsive decision-making, they are moving towards fully automated systems where decisions are made and executed in real time.
In this world, there is no opportunity to sense-check outputs after the fact. The entire pipeline, from data to decision, must be robust by design. That requires data scientists to understand not just how to build models, but how those models will be drive decisions in live systems.
Closing this gap does not mean turning every data scientist into a fully trained OR specialist. A more practical approach is to integrate OR thinking directly into data science. That means equipping data scientists with the tools to frame decision problems properly, to understand constraints and trade-offs, and to design systems that optimise outcomes, not just predict them.
It also requires a shift in mindset. Data scientists must begin to see themselves not simply as model builders, but as designers of decision systems.
The next evolution of data science
As organisations seek to move from insight to action, the ability to optimise decisions is becoming just as important as the ability to predict outcomes. Retail demand modelling is just one example, but the same principles apply across industries.
For businesses, this is about competitive advantage. The organisations that succeed will not be those with the most accurate predictions, but those that can turn those predictions into better decisions, faster and more consistently.
To remain relevant in an increasingly automated world, data scientists must learn to optimise. Operational research is not a niche discipline or an optional extra but the missing link between insight and action, and it is set to redefine what it means to be a data scientist.
David Hoyle is Research Data Scientist Specialist at dunnhumby. For more about David go to: https://www.linkedin.com/in/davidchoyle
Empowering employees and creating cross-functional communities around data encourages trust and collaboration, helping to break down silos and support better decision-making. Organisations that invest in data literacy and shared accountability are far more likely to see measurable impact from AI adoption.
OR Society members can watch David's webinar here:
Watch Here