Manufacturer Optimises Supply Chain Analytics With Azure Machine Learning

Pandemic-driven price spikes highlighted long-standing challenges

When the COVID-19 pandemic struck in 2020, this global manufacturer of primary batteries and portable lighting products was among the many organisations impacted by shipping and supply chain challenges. Compounding the issue was the fact this client was using an Excel spreadsheet  approach as its primary shipping cost analysis tool. While ocean freight costs and lead times had been steady prior to COVID, the pandemic drove increases in costs, which leapt up to 10 times prior charges, lead times, which jumped from three days to more than 15 days, significant port backups and complications with inland transportation .  The client realised its previous method of tracking costs was not sustainable considering these rapid industry changes, had an immediate need to improve the accuracy of its ocean freight forecasting to understand the impacts to its business. This was not unusual for any consumer goods company; ocean freight had been so steady over so many years that this sudden increase impacted everyone.

Initial assessment

An initial analysis involved manually processing a considerable amount of data around hundreds of components including destination ports, types and sizes of shipping containers, etc., all Excel-based. Protiviti leveraged the Microsoft Azure platform to perform an initial cleaning and sterilising of all the data, with a goal of ensuring that every source and destination combination would be at a high level of accuracy. This would enable the client to operate with the confidence that its 30, 60, 90 day and beyond forecasts would be accurate as possible.

Protiviti used Python within Azure to accomplish the bulk of the data ingestion, data cleansing, filtering and built preliminary models. Once a predictive model had been developed for each shipping lane, the team leveraged Microsoft Auto Machine Learning (ML), which allows for hyperparameter tuning, in which the Auto ML tool reviews algorithms and parameters within those algorithms to identify the best configurations for each model to maximise predictability.  This was eye-opening as it enabled rapid automated analysis.

Partnering with Protiviti to analyse the data, the client worked to standardise the nomenclature of its data. During that process, the company realised that data governance could be applied to other downstream processes to better define its shipping lanes while simultaneously improving container and unit definitions. This enabled the forecasting team to inform the downstream teams as to how to improve/cleanse their data for regular consumption and other data science activities.

Introducing machine learning models

For this client, the immediate solution was to focus on its data and to build machine learning models that would look at data points to predict what costs would be going forward. Challenges going into the model building included:

• Not all data was available
• Uncertainty around which data should be considered
• Uncertainty around whether the models would predict any usable information towards reducing costs
• Significant training required to support a new platform and data science methodologies

Supported by Protiviti’s predictive analytics experts, the client team built models around particular shipping lanes (for example, Shanghai to Los Angeles, which accounts for 50% of this company’s shipping). Using their own data plus industry-standard indices, they developed hypotheses on the components that could impact costs. There was much exploration into the data, as it had not been looked at in this way before. Each model was built and then trained to be more accurate with each iteration. The goal: no surprises while improving expense management and visibility.  

Interactive training and hands-on model building

One of the additional challenges along the way was getting the client team familiar and comfortable with the Microsoft Azure platform, specifically the ML studio. At the client’s request, Protiviti and Microsoft shared best practices for maximising the platform while teaching the client team new data science methodologies. Considerable time was devoted to interactive training exercises where test models were built and evaluated as the team members learned what could be done differently to improve performance results. The secret to success was the hands-on model building and learning. Part of the data science process and methodology is being able to take some of those test models and understand what the models were saying from a predictability standpoint and why some models were more successful than others, then go back and massage the data again to improve results.

A pioneer in the consumer goods industry

The ability to quickly respond to a shift in macro-economic business conditions highlights the power of their data programme. This problem that surfaced during and after the pandemic for everyone who ships goods – the deafferenting behavior is the ability to respond faster than competitors. The client team wanted to use their data more intelligently and introduce more modern methodologies to more accurately predict future ocean freight costs. This proved to be a game-changer as this company now regularly incorporates the building of machine learning-based models and predictive analytics, along with industry indices, to improve cost predictions. In doing so, this company aims to become a cost predictions leader in both the consumer goods and shipping industries.

Moving from a rigid, manual approach to a more flexible and agile machine learning-based approach was a significant step for the organisation. Transforming its pricing predictive models using data science processes and methodologies allowed it to both manage data more effectively and respond nimbly to rapidly changing market conditions.

Protiviti has become a strategic partner for this client. Our predictive analytics team was able to teach critical aspects of data science methodologies, explaining the “why” behind each step of the model building process. Working meetings became brainstorming sessions where teams review test results and learn how analytics contributed to the test’s conclusions, helping grow advanced analytics capabilities. This new way of doing business is now an established part of the client’s daily operations.

Impact by the Numbers: