Automation through machine learning (ML) allowed Amazon.com to predict future demand for millions of products around the world in seconds. Executives at the multinational tech giant have successfully reinvented its data infrastructure to improve purchasing systems, automate inventory tracking in distribution centers, and deliver on your two-day shipment promise to customers.
With a comprehensive predictive model fully integrated in the cloud, Amazon.com uses data to make better decisions, simplify operations, and deliver compelling customer experiences.
Reinventing manual product forecasting through ML
Predicting customer demand is not an easy task in e-commerce, as delayed inventory or erroneous deliveries can be costly and disrupt the chain supply. Although 80-90% of all planning activity can be automated, many industries still rely on manual forecasting.
Ecommerce retailers sometimes have to predict hundreds of millions of products, and “no human brain can predict on this scale on a daily basis,” said Jenny Freshwater, vice president of marketing and traffic technology at Amazon.com . and former vice president of forecasting. The Freshwater team has made predictions for over 400 million products on Amazon.
Design teams, no matter how advanced, can’t do it all: assess historical trends, develop sales forecasts, and conduct independent research on such a large volume of products. Even when combined with more complex models, legacy systems like outdated computer software or manual inventory records aren’t as accurate as machine learning models.
When toilet paper sales increased 213% during the height of the Covid19 pandemic, Amazon used AI-based forecasts to respond quickly to unexpected demand signals and increase adaptability to market fluctuations. predict before COVID, but our models have quickly responded to the new trend in demand ”.
Freshwater advises retailers to re-prioritize their machine learning roadmaps to deal with the unexpected. The pandemic prompted their team to make changes and implement new macroeconomic ideas. “The data, and they’ve gone from predictions based on numbers with confidence intervals to predictions based on scenarios. Some scenarios that we had thought of before, but never got priority because they didn’t:” It doesn’t It’s not urgent at this point, “he said.” With the pandemic strike, we have shifted priorities. And a lot of things that we wanted to do in the past are now in production. “
How Amazon.com became a leader in product forecasting
According to Freshwater, Amazon’s journey to machine learning began about 10 years ago to improve predictive accuracy. “We started machine learning because our moving average models weren’t as accurate as expected.
Business leaders recognized the need to leverage data and machine learning to deliver on customer promises and deliver profitable functionality at scale. With these goals in mind, Amazon.com aims to become an AI-based leader in product prediction.
To speed up the process as demand increases, the company has partnered with Amazon Web Services (AWS) to “create machine learning models that, in terms of the amount of data, the characteristics that we use to predict the demand ”, such as the complexity of algorithms. We now use neural network models to predict demand for the products we sell on Amazon. “And the difference was day and night.”
Amazon.com uses machine learning on AWS to aggregate and analyze product purchase data and run its forecast models. Additionally, the company uses browsing data and purchase data to provide more personalized product recommendations. Machine learning enables data experiments that empower data scientists to do better to create a more personalized experience for customers.
For Freshwater, prototyping and iteration was essential to achieve successful machine learning results with our existing models and at one point we were able to improve 15 times what we had before with these “neural network models”. “. So it was a very iterative process.
Key Insights for business leaders Using Predictive Models
ML in the cloud is key to extracting insights from data and making better business decisions. Use these best practices to maximize ML modeling on your reinvention journey.
1. Trust the model. Monitoring millions of products on a regular basis requires valuable development time and resources. Freshwater says that almost all of Amazon’s forecasts are automated through machine learning models, and people and business users only interact with and override the forecast when they have information that the models could not possibly have the models run while business users are away Focus on other critical tasks. Overrides should be considered when you are sure you have more quality or trending information than the model.
2. Define a clear data strategy. It’s impossible to think about machine learning and get real value from ML models without first developing a data strategy. At Amazon, the preparation of data for ML use was an important part of the strategy. When I talk to people about our path from our old to the new model, I estimate that about 40% of the time was spent preparing the data, says Freshwater. With functions, it’s really about getting the data in the right place. Without this investment of time and effort, we’d run the risk of getting poor or biased results because the data doesn’t accurately reflect our decision-making.” Developing a data strategy that is aligned with business goals, prioritizing data cleansing, and presenting the data representative of what you want to predict or optimize is key for leaders embarking on their ML journey.
3. Know what you are measuring. At the start of your reinvention journey, it’s important to be clear about what you’re measuring and how you know you’ve improved. Freshwater recalls, “In our problem space, we expected more and more product every year. So it was not enough to look at year-to-year comparisons. We had to implement a series of benchmarking models that used a methodology to know. If compared to the benchmark, our new models were better. “Take some time up front to define metrics and criteria for success to avoid wasting time on quitting.
4. Build a data-driven culture: To get the most out of your machine learning initiatives, you need to foster a culture of innovation and data-driven thinking across your organization. Reinvention projects are really about being able to take risks and fail quickly, says Freshwater. Pick people who are okay with multiple failures before you succeed. I think that the culture should be cultivated while working on each prototype because the first results aren’t always good. If so, maybe not. When you experiment with machine learning, make sure you have people at every level of the business who are passionate about the perspective of the possibilities of ML. Freshwater adds, Since Amazon is a data-driven culture, we almost got the needle Move completely into machine learning just by looking at the fact that our models were much more accurate in terms of predictive accuracy.
