We started this series by describing how keeping your data and analytics current, accurate and relevant can help navigate uncertain and chaotic times such as the COVID-19 pandemic. In the second installment, we described how artificial intelligence (AI) enables flexible forecasting built on models and data that change as the world does.
Our clients are telling us how necessary and urgent it is to update their analytics processes amid the pandemic. In a recent survey that we conducted in North America, three-quarters of respondents said that their companies had changed their data management and analytics/models during COVID-19. Respondents at companies with more volatile revenue swings were much more likely to have made major changes in data management than those with less volatile revenue.
In our final installment, we describe agile analytics in action and then reveal three essential elements for success.
Shopping, pandemic style
Retailers have had to adapt to two strong shifts in consumer behavior during the pandemic: Fluctuating demand for goods and a move to e-commerce. For example, as governments imposed social distancing rules in early 2020, a leading global convenience store chain found the “basket” of top-selling items changing virtually overnight. With sales falling sharply as customers avoided going out, the chain needed to quickly adjust its inventory and store displays to maximize sales and profits.
To remedy this, we created a robust cloud-based data and forecasting foundation, and then within five days created new models to enhance the convenient chain’s understanding of market dynamics to respond to daily changes in a wider range of parameters. This helped decision-makers see changes like the shift toward bread, milk and eggs over the previous priorities of coffee, cigarettes and muffins. This updated system continued to provide new insights throughout the volatility of the pandemic. Managers responded by making extra efforts to keep those items in stock and placing them near the checkout so customers needed to spend as little time as possible in the store. As a result, our client said average per-customer purchases of these products rose about 25%.
Fine-tuning supply chains amid demand shocks
A global consumer technology provider had well-tuned planning models to account for how next-generation releases affected its complex supply and distribution chain. It lacked, however, models that took into account external shocks such as COVID-19 lockdowns and their impact on economic activity and product demand.
We worked with this company across its global markets to understand how changes in orders from its distribution partners helped explain how COVID-19 was affecting sales. Our work is generating frequent, country-specific alerts to management on the effects of COVID-19 so that they can devote more energy to understanding and responding to these changes.
Our AI team is working with the Xprize Pandemic Alliance on ways to use AI to better predict the spread of COVID-19 and the effectiveness of various countermeasures such as lockdowns. We are developing, refining and evaluating AI models to predict the spread of infections and the effect of countermeasures such as lockdowns to reduce infections. The model is updated daily with new data about both infections and the state of non-pharmaceutical interventions (NPIs) such as lockdowns, travel bans and restrictions on public events in countries all over the world and every state in the U.S.
Our work includes continually assessing the quality of various types of data and its effectiveness in predicting future infections and the role of certain countermeasures. We have found that changes in even unreliable data over time can provide useful insights into underlying trends. In addition, we found that we are able to automatically retrain this model every day with the latest data to help our algorithms adapt to new information and provide new insights on the progress of the disease, allowing them to make better decisions and to increase accuracy continuously.
The lessons learned
Our work with Agile analytics clients across the globe reveals four essential components of an Agile analytics strategy.
Update loops for your data and your models
Too many companies still implement AI as one-time projects. As a result, the analytics models are fed stale data that doesn’t reflect the latest sudden change in the business landscape, whether that’s a new COVID-19 lockdown or a competitor’s introduction of a popular new product. Running COVID-19 analytics models frequently — even daily — trains models in the underlying reasons for changes in the real world and helps them better predict the future. Doing so requires investments in staff, skills and processes to regularly review and update both the data and models.
Analytic Sprints that target quick returns
For the global convenience store chain, we created an Agile process like that used in software development. We used each Sprint to describe in rough form our proposed analysis, with examples of the simplified dashboards in which we would present the results. This allowed us to perform more steps in parallel (such as requirements gathering, building a data integration pipeline, and developing prototypes), reducing the time to business benefit and ensuring that we met the retailer’s needs. This iterative process also allowed us to deliver quick results with clear business benefits, building support for future analytic efforts.
Elements of an Agile data infrastructure
Allow the creation of data lakes that can rapidly store and make accessible any type of data that your AI applications need, such as from social media or edge devices on the Internet of Things.
Use AI-enabled data pipelines to automate data discovery, ingestion, cleansing and validation, inferring the content of data and its quality and propagating it into an enterprise data domain to support access by users.
Make data access and analytics easily available and reusable with a searchable marketplace of data and models.
Use metadata management to more easily understand where data originated, its quality and how it has been changed to ensure that AI models are analyzing the right data.
Archive perishable, less critical data onto lower-performance, less expensive platforms to reduce the costs of both data storage and of training models on less useful data.
Leverage cloud-based data management and analytic platforms to minimize cost while maximizing flexibility and scalability.
Extend existing security and data protection policies across hybrid cloud/in-house environments.
Continuous AI-enabled learning to prepare for tomorrow’s shocks
AI platforms such as our Learning Evolutionary Algorithm Framework (LEAF) system generates diverse model candidates to distinguish which are best suited to solve a particular problem. Through masses of generations of models, the system gradually converges on those best suited to solve a specific problem.
LEAF also allows continuous learning and optimization on a surrogate model of the real world, allowing substantially faster learning through simulation. This simulation-based approach fine-tunes the models by considering the outcomes of all possible future states without waiting for actual real-world data — which may come too late to prepare for that future state.
We also recommend analyzing analytic models at the level of their individual components such as variables and parameters to see which are most and least causal and accurate based on changes to the data since they were developed.
Agile comes to analytics
The days of one-off, “fire and forget” analytics are gone. Just as software must constantly evolve in iterative, analytic sprints to meet business needs, so must the analytics we use to understand and navigate dramatic changes.
As one of our study respondents said, “The way our customers interact now has changed completely, so we are having a hard [time] projecting the new behavior because it's nothing like we have seen before.” AI can help meet such needs by not only analyzing current trends, but also continually refining the data and models used in that analysis.
This article was written by Bret Greenstein, SVP and Global Head of Data & AI; Babak Hodjat, VP of Evolutionary AI and Jason Kodish, Global Guild Lead, Data & AI.
For more insights, visit the AI section of our website or contact us.
Putting Agile Analytics to Work: Three Essential Requirements (Part 3)