Skip to main content

Select your location

How can Google Cloud predict the demand for my pair of jeans?

Mo Joueid
Boxes on warehouse

Why should retailers include price optimization and demand forecasting at the core of their digital transformation? Mo Joueid, the Practice Lead for Google Cloud Platform at Kin + Carta, shares some insights on how the Google Cloud team at Kin + Carta are supporting supply chain optimization journeys

In Part 1 of this Supply Optimization series, we looked at the story behind how the price of my pair of jeans was being influenced through Machine Learning (ML), and how this is increasingly becoming one of several new ‘tools’ in what is increasingly becoming a very competitive retail landscape. However, any element of price optimization is likely to impact other areas for a retailer, particularly demand forecasting. Whilst ensuring that they are able to offer goods at the right price, retailers also need to be able to accurately anticipate customer demand, and ensure that goods are in the right place at the right time. Out-of-stock goods/products account for $634 billion in lost sales worldwide each year, while overstocking results in $472 billion in lost revenues1. Modern retailers face the challenge of being able to perform demand planning at scale.

In this Part 2 of our Supply Chain Optimization series, we look at how effective demand forecasting enables better planning decisions that help retailers minimize stock outs and excess inventory.

Putting Supply Chain 4.0 to work

Back in 2011, the German government coined the term Industry 4.0. Whilst initially intended for manufacturing, the idea was simple - smart machines would revolutionise swathes of industries through a network of inter-connected devices, exchanging information and facilitating decision making in an automated manner. The Cloud has made a significant contribution to this realization, with many signaling that we are well on our way to Industry 5.0 through the use of Artificial Intelligence (AI) and ML.

McKinsey originally extended this notion by introducing Supply Chain 4.0 (SC 4.0) back in 2016, before looking at the application of SC 4.0 to jump-start performance, and customer satisfaction in Consumer Packaged Goods in 2017. This new perspective on the supply chain, would allow retailers (and consumer goods manufacturers) to bring about digitization that would allow them to become faster, more flexible, granular, accurate, and above all efficient. Unsurprisingly, this hinged on leveraging data and ML to deliver predictive analytics in demand planning, amongst other things. We have since seen significant momentum behind SC 4.0 as we forge a new way forward in the post-pandemic world around us.

The many dimensions of demand forecasting

Retail businesses understand the value of demand forecasting— using their intuition, product and market experience, and seasonal patterns and cycles to plan for future demand. However, where there is value there will no doubt be challenges, and retailers face a number of these when implementing demand forecasting. What is really driving demand for their product, how is this demand shifting with consumer behaviors, why is consumer behavior shifting? How does this vary by region and can this be modeled? Can recent sales history be relied upon given global events over the last few years? How do we address new products, rapidly changing SKUs or short-life cycle products for which there is little data around? And many more questions that need to be faced down to build an accurate and reliable model on which to base demand.

The different approaches to demand forecasting

When looking at this business challenge, retailers will often give thought to either purchasing a full end-to-end demand forecasting solution, which takes significant time and resources to implement and maintain, or leverage an all-purpose ML platform, which may require deep experience in both modeling and data engineering.

Google Cloud has become a great alternative for this data-related challenge, due to the rich background that the platform has to offer in the way of AI and signal data (or demand drivers as they are often referred to). When looking at both price optimization and demand forecasting, demand drivers are key - both in terms of enriching your points of reference, but also ensuring that these do not unintentionally distort decision making. Google Cloud allows us the opportunity to augment models with additional demand drivers coming from across the Google ecosystem, and those outside of the Google ecosystem without the need to bring your data to Google Cloud - something that we will revisit later in this article.

The Kin + Carta approach

When tackling demand forecasting with clients, we use statistical or deep learning models to automatically recognize patterns in data and make predictions. There are namely two different approaches that Kin+Carta consider:

  • How do we quickly and efficiently demonstrate valuable predictions through a variety of custom or pre-built ML models,
  • How do we scale and productionize the demand forecasting offering i.e. Vertex AI Forecasting.

In Part 1, we looked at the Data Platform that we help retailers build, allowing them to move from simply capturing data, to consuming that data to employ data-driven strategies for revenue growth. In the next few sections, we look at how we re-use this platform in the context of the above approaches.

Custom build forecasting ML models

There are tried and tested approaches to tackling demand forecasting, primarily through quantitative or qualitative means. In the case of my pair of jeans, this could be as simple as using statistical models and historical (time-series sales) data to predict future sales. This time-series data is pre-processed to get the data into the correct format, curated and then made ready for analysis within BigQuery (in the Analysis/Curated Zone from Part 1). Once this data is available, BigQuery ML can then be used to run and train a time-series model. Once the model has been trained and evaluated, this can then provide forecast predictions for the pair of jeans (and the many other products the retailer may have). Retail Lines of Business (LOB) can then make use of a dashboard to visualize the forecast predictions using a Business Intelligence (BI) tool like Looker, and integrate these predictions into their overall retail analytics and supply chain management.
Steps for using statistical models and historical (time-series sales) data  to predict future sales

To equip retailers with the ability to optimize over time and identify new areas of opportunity, a regular cadence for the time-series modelling is essential. This enables the models to adapt to changes and maintain performance over time, and the ability to incorporate unprecedented events into their predictions. This would be an integral function of an operationalized, living ML system - which is beyond the scope of this article.

There are some notable challenges with the approach described above:

  • With a reported 87% of data science projects never making it into production, how do we show value quickly and help businesses become comfortable working with data?
  • There will be many demand drivers that will influence demand beyond just historical sales data. How do we identify/qualify these, and incorporate them into demand forecasting?
  • How do we predict demand for a new product (that has no sales history)? Or a product which does not have stable demand? Or a product for which previous sales is not a good indicator of future demand?

To overcome some of these challenges and provide retailers with richer, more accurate forecast predictions, at Kin + Carta we incorporate a qualitative approach that leverages other data signals. These include, but are not limited to:

  • Past sales, inventory, product catalog,
  • Pricing and promotions,
  • Web history, marketing, store and location data,
  • Loyalty, and
  • Third-party.

To help retailers quantify the potential benefit prior to extensive investment, we begin by validating the signal in candidate data sources and iterating on exploratory models using Kin+Carta’s Octain™ (running on Google Cloud). This takes into account variations for specific product demand i.e. consistent demand with spikes and troughs, flat demand, seasonal demand, how demand varies across a product’s lifespan from campaign launches to being discontinued, predictable events e.g. Six Nations Championship, and allowing room for errors. We find that through this combined approach, retailers begin to see value in hours rather than months.

The above models are deployed with Vertex AI (as introduced in Part 1 of this series). Vertex AI receives small batches of data to the service, and returns forecast predictions in the response. It is optimized to run the data through hosted models with as little latency as possible, rendering the predictions in a BI tool like Looker.

In a recent case, Kin+Carta helped a retailer (who had experimented with alternative technology) to establish demand forecasting, providing them with a competitive advantage over two specific competitors. In doing so, Kin + Carta were able to demonstrate up to 12% improvement in unit forecasting versus the alternative ML methods evaluated, resulting in a multi-billion dollar cost saving:

Demand forecasting alternatives comparison

The client was provided with a consolidation of all supply and demand data into a single interface, enabling the retailer to:

  • Continually optimize their forecasts throughout their challenging supply chain,
  • Determine what products are needed where, and manage inventory to minimize the time it takes to receive goods/products,
  • Determine the optimal location from which to ship i.e. distribution center, store, drop-ship,
  • Leverage the value of shared stock and promote sales of slow-moving inventory,
  • Identify opportunities to increase supply and exploit the demand opportunity of fast-selling products.

The above is a perfect demonstration of how demand for a fast-selling product may subsequently be used as a data signal prompting a price adjustment, demonstrating the combined power of demand forecasting and price optimization working in unison to increase revenue and margins.

Vertex AI forecasting

In January, we saw Google announce the release of the Vertex AI Forecasting feature (previously Demand AI). Launched ‘as-a-service’, this extends the previous capabilities covered above, by bringing with it demand forecasting ML algorithms for both long horizon order planning and short horizon use cases, enrichment of demand signals, and specific insights for retailers.
Vertex Forecast steps: Prepare Dataset, Train model(s), Forecast and Visualize/Integrate

The Data Platform introduced in Part 1 provides all of the building blocks needed to capture, process and prepare (curated) retail datasets. With Vertex Forecast, we extend the capability of this platform further by being able to include up to 1,000 different demand drivers (color, brand, promotion schedule, e-commerce traffic statistics, and more). This opens up the platform to myriad additional use cases which can benefit retailers, including staffing distribution centers/stores, targeted marketing, and driving other campaigns across both physical and digital channels.

Google Cloud Platform addresses retailers' need for demand forecasting. At Kin + Carta, we unlock the power of Google Cloud ML to achieve high forecast accuracy that has the potential to not only help retailers achieve significant cost savings, but also move towards a smarter, more sustainable supplier chain. This has the potential to positively impact so many business functions of a retailer ranging from order planning, to replenishment and allocation planning, to workforce and space planning.

Be on the lookout for Part 3 of our Supply Chain optimization series, in which we explore B2B Price Optimization and Demand Forecasting.

Contact us

Want to know more about how Kin + Carta can help you on the road to Supply Chain Optimization?


Share this article

Show me all