Life in a digital world

The power of big data and machine learning in logistics

Gay Sutton cuts through the hype and erroneous assumptions around big data, artificial intelligence and machine learning to discover how the vast swarm of continuously flowing online data is being channelled cost-effectively to enhance logistics and supply chain management.

Digitalisation of almost every aspect of human life can leave us wondering if we live in a virtual world where human interactions, purchases and pleasures take place online and no longer face to face. This trend has created an overwhelming deluge of data.

For many years, data generated at the point of sale (POS) has been utilised to improve forecasting and to optimise logistics and the supply chain. Investments are now being made in GPS (global positioning systems) and RFID (radio-frequency identification) to track products, processes and deliveries. Machine sensors and webcams are monitoring transport vehicles and production equipment, enabling predictive maintenance to improve reliability. But this is not even touching the tip of the data iceberg.

A vast amount of data is now being generated through our online browsing and buying habits and from unstructured data sources such as camera and surveillance footage, social media images and posts, and much more.


The new world

In the past, analysing this overwhelming and non-stop flow of data would have been an insurmountable challenge – but not anymore.

“We have seen enormous advancements in computing power in recent years and can now process these vast quantities of data much more cost-effectively, and in a meaningful way,” says Andrew Fowkes, Head of Retail Centre of Excellence at analytics company SAS. “Analytic runs that a few years ago would have taken days can now be completed in seconds.”

Moreover, this can be done in the cloud, so those huge investments in on-premises data storage and computing power are becoming increasingly unnecessary. Software usage can be dialled up or dialled down in the cloud according to business need. As the technology matures, this could make top-quality analytics possible for companies of all flavours and sizes.

Retail is an excellent example of where improvements are already being made. Traditionally, POS sources have provided vast quantities of data for analysis in time series (time order). Today, a wealth of additional data can be factored in, such as weather events, browsing traffic, what your competitors are doing, and the unique attributes of the individual product, location or even the customer (see box). Much of this is text-based. It can vary in applicability over time and can also be affected by other contextual things that are happening in the marketplace such as promotions and fashion. So how can today’s analytics combine these awkward variables and make sense of them in a meaningful way?


Andrew Fowkes, Head of Retail Centre of Excellence at analytics company SAS

Enter the intelligent machine

“We do this and get a much more accurate demand forecast by using a machine learning model or combining a machine learning model across a number of products with a time series approach,” Andrew answers.

Companies can tap into the trail customers leave online to mark their browsing and purchasing history, which may even include products that are added to the basket but not bought – a very subtle indicator of taste and demand. “So, you no longer have to assume why sales are behaving the way they do. You give the machine the data in near real time, and it uses the continual stream of data and other attributes to work out what is happening and why.”

The outcome is a more accurate forecast that can be flowed throughout the supply chain, resulting in improved materials ordering, production levels and order fulfilment, and leading to better inventory and fulfilment optimisation.

The figures Andrew quotes are impressive. “We have customers that have deployed machine learning models and have seen improvements of between 3% and 10% in their stock forecasting accuracy. That yields enormous savings in terms of delivering merchandise in the right volumes to the right locations and eradicating redundant stock.”

One key attribute of the machine learning is… learning. Once deployed, the machines are continually self-learning and improving. They work out, for example, how to value and weight the influence of various data elements such as product attributes, and then continue to hone that knowledge and improve performance. Then, if a human intervenes to override a forecast, the machine learns from that too. “We are finding that the machines themselves can evaluate the human interaction effectively and apply the understanding to the next situations. In other words, they self-correct.”

“New forecast models can deliver something that we’ve not seen before – a really robust forecast for a new product that has no sales history. That's exciting news for any business”

In the pipeline

This brave new world is moving fast and there are many developments in the pipeline. We are already seeing computer vision in action. This uses machine learning to automatically recognise and analyse digital images, and then make a decision based on that. On a production line, for example, it recognises faults in the product and decides, there and then, whether to take the defective part off the production line and discard or repair it. If a human then intervenes to change that decision, the machine learns from that and makes a better decision next time. Meanwhile, in the logistics sector, DHL is developing a computer vision application to optimise the loading of planes and trucks by recognising the characteristics of packages and deciding on the best way to stack them.

Even more exciting is the concept of analysing social media images and posts to learn about changing customer tastes and trends and then apply them to the demand forecast. Twitter comments, for example, often express reactions to a product or event while Instagram images can be used to identify trends. What colours and styles of clothes are people wearing most? What foods are they interested in trying? The options are endless, and the resulting data is invaluable for accurate forecasting. It can also be fed into other applications. For example, there are plans to monitor the internet and social media for anything that could signal disruption in a delivery cycle, so that the service can be very quickly restored to normal.

“We have now gone beyond the AI hype and understand that machine learning is quite practical,” Andrew notes. And as the machines learn, more exciting capabilities are emerging. “The new forecast models can deliver something that we’ve not seen before – a really robust forecast for a new product that has no sales history. Now, that’s exciting news for any business. It removes the guesswork, and you don’t have to predict demand by attributing it to another product and hoping for the best.”

In the future

This is an interesting moment in the development of this technology but, as Andrew remarks, “Like most things at this stage in their cycle, it will soon become normal for us to use all the data that is available to us. So rather than evaluating data in its raw form and then safely storing it for use later, more and more organisations will be letting the data just flow. Machine learning algorithms can evaluate it in almost real time, giving us a much more accurate and immediate result.”

He concludes, “The machines will be ahead of us by design, and this supports the thinking: ‘We’ll send you the product before you have even ordered it.’ I think that’s where we’re headed.”



Transaction data:

Real-time basket transactions through a physical till or online shop

Browsing data:

  • Online shopping journey, including browsing history and products put into basket but not purchased
  • Click-through data (from online ads)

Unstructured data:

Information that can’t be stored in a relational database:

  • Photos, social media and webcam images
  • Messages and social media posts
  • Music, videos etc.


  • Physical attributes: product type, brand, pack size, colour, weight, flavour, material, style etc.
  • Aspirational attributes: how fashionable, ‘stars’ or role models that have worn or used it, price point range etc.
  • Location attributes: what type of store, whether it has a car park, details about nearest competitor
  • Customer attributes: characteristics of the customer, profile type, age, gender etc.