Demand forecasting and cost optimization for a chain of 8,000 convenience stores

Our client is a chain of small and medium-sized modern convenience stores in Poland. Over 20 years of trading, they had earned significant market coverage and brand awareness. The company had established its stable, enduring position primarily through aggressive expansion, focusing on being a leader in the modern convenience stores segment through investments in innovation and artificial intelligence.

First of all, we reviewed the existing approach, dug into the business problem, and recommended optimizations. The main challenge in tackling the predictions was the level of detail expected from the outcome: each product for each store required a separate prediction per time interval and day.

We built a solution that satisfied the business requirements within the suggested changes. We also proposed several new approaches to tackle problems that might occur in the future (but were yet to materialize).

The existing solution did not meet industry standards. Therefore, we rebuilt it using industry best practices. Starting from data loading and preprocessing, we were able to fix optimization issues and process years of historical data for the client’s entire pool of stores. 

Within that part of the pipeline, we reviewed potential data sources and supported the process of analyzing and gathering them. As a team, we balanced the model and data-centric approaches used in ML (we always believe effort invested in both streams is the most efficient way to do ML projects).

Starting with a model-centric approach, we achieved a solution that all stakeholders were satisfied with, but there was still scope for improvement. And we temporarily switched to a data-centric system after confronting a low slope of improvement (based on invested time). We then hit upon several significant breakthroughs by acquiring new data sources and fixing data quality issues within the whole dataset. 

We also introduced other essential components, including: 

• Model tuning
• Custom loss functions (to tackle business requirements to penalize underestimation more than overestimation)
• Experiment tracking
• Model evaluation
• Model explainability

Within these components, we achieved a fully-automated pipeline that hit all the required validation and monitoring aspects. Moreover, the client can now easily track, replicate and validate all experiments. 

In maximizing the model’s effectiveness, we used a black-box algorithm, which can cause business users to question a model’s output. To tackle this challenge, we introduced explainability, which benefited our development team as well as our client. 

Our developers could precisely validate the results, understanding patterns and identifying weak spots — while the business had confidence in the solution and could explain predictions to stakeholders. The solution enables users to describe the general model alongside batch and single predictions.

Even though the project’s scope was to handle a defined set of product groups, the solution we created was widely applicable. The code allowed us to quickly introduce the next group of products from different product branches.

In the next phase, we considered deployment. To preserve the solution’s integrity, we stuck to Microsoft Azure: the client’s existing cloud provider. Besides deployment and automated scheduling, cost optimization was another significant factor to consider. 

Thanks to the clean codebase and the quality of the architecture, deployment was relatively straightforward. We wrapped the solution as a Docker Image, which we also used to build an Azure ML environment. Within Azure ML, we defined the whole processing graph containing data loading, model training with evaluation, and registering a new model version. We also created a similar pipeline to perform the model’s inference. 

Each pipeline was easy to control in terms of configuration, including business assumptions, scheduler, and resource usage. And thanks to comprehensive testing and choosing the most accurate resources, we minimized the costs of the whole cloud infrastructure. Moreover, each pipeline was executed in a serverless way. Hence resources were released after the process finished to avoid additional costs. 

Computation time optimization was achieved by parallelization, with each product processing individually and independently.

Our collaboration with the client during the development and testing phases led to the delivery of dedicated workshops for the client’s data science team. We shared our team’s and general industry best practices across machine learning and software development during these sessions. 

We used the weekly workshops to explain the end-to-end pipeline in detail, giving the client a clear understanding of the methodology used and the components in each project. 

Moreover, we delivered several hands-on workshops, including live coding sessions, which allowed us to develop a project together and introduce recommended technology stack and code standards.