Join Adam Aspinall, Vista DnA Senior Data Product Manager, as he shines a light on the technical and relational alchemy of modern forecasting in the Sales and Operations Planning domain (S&OP).
Back to the future
Forecasting has always been an Olympic effort to get right. Teams had to deal with the legacy of arduous manual processes, but the biggest hurdle both then and now is the nature of the business. Vista offers a range of products via e-commerce with mass customization. We’re home to more than 50,000 product SKUs (stock keeping units). Add huge volume to order variability in a manufacturing environment, and Vista’s race to become a customer-centric tech company makes sense – in forecasting and beyond, it’s a key that unlocks complex challenges.
Let’s hone in on mobilizing the endless assortment our Vista customers love. Each SKU needs to be forecasted so fulfilment teams can plan for machine capacity, labour needs, and raw materials. Plus, daily demand can fluctuate wildly. My team needs to be creative and deliberate about how we think about forecasting SKUs that have seasonality, for example – one of many topics keeping us on our toes 365 days a year. Similarly, in our new world where data products, data mesh, domains and chapters are made possible through DnA, Vista’s data-and-analytics organization, being intentional matters. We’re balancing myriad needs at any moment, from front-end product and marketing functions to back-end manufacturing and supply chain side (MSC).
Getting down to business
What are we aiming for? Our job is to support S&OP processes with data products built on leading data science and engineering. We’re creating a semi-autonomous global forecasting data product that requires minimal intervention. In parallel, the nut to crack is being efficient with who gets what level of forecast. Marketers planning big categories won’t need fine-grain forecasting; MSC teams procuring raw materials for specific machine capacities will. If that sounds transactional, it isn’t. We really like to co-create with our stakeholders. We’re building a model that makes predictions about future demand, yes. But we don’t own Vista forecasts. It’s about enabling execution across the business.
To get there, the goal is to generate 40% more accuracy than historical models. In development mode, we’re laying the foundations of new and best time-series forecasting to include rock-solid data quality and testing, autonomous data reviews, and automated alerts of changes in model performance, for instance. In tandem, we’re expanding coverage and are two months away from being in all regions. Of course, there’s a spotlight on future capability. The current solution relies largely on historical customer orders, but what about leading indicators of demand changes? Strategic objectives need to be added as a new model regressor to make better predictions. Imagine if Vista planned to ‘go big’ on signage design in 2023, as a quick illustration.
Success so far
Fine-grain forecasting was rare in addition to being a brittle, by hand process with regional variation. Cut to today: forecasts are robust and daily, with automatic weekly updates that include the most recent customer demand trends and respond to upcoming events like holidays and promotions . The S&OP model forecasts with a 90% confidence interval rather than point predictions. In other words, we’re 90% sure we’ll sell 40,000 to 55,000 units of SKU ‘XYZ’ on the 1st of June. The 90% represents a balance between model performance and usefulness of the data product – if the interval is too wide, we can reach (theoretical) 100% confidence, but insight for stakeholders is negligible.
We achieved a 26% improvement in our last quarterly review. It’s significant while leaving us with opportunity to hit our 40% goal. Of course, the eternal mission remains alignment of supply and demand. In an extreme ‘100% more correct’ example, we’d know exactly how many team members we needed (and with what skillsets), how many machines to run, and how much raw material to purchase to the last sheet. The closer the accuracy, the less ‘just in case’ is needed to meet demand, the more savings we make on labour, materials and shipping. Millions of dollars are on the table.
A multitasking data product
We can be a lot more effective in aligning supply and demand by bridging terminology gaps between front-end and back-end teams. Product labelling on our website can differ from how we classify items in MSC; merchandized products use language that customers will understand – like a choice of ‘premium or premium plus’ paper thickness, versus supplier parlance of ‘16pt or 18pt substrate’. The solution is a translation component or ‘MSKU conversion’ for our model.
This forecasting data product is helping to build connections by creating a common language of ‘apples to apples’ at a crucial point in time. Relationships across domains are even more critical as the organization evolves. I’m finding a new awareness of forecasting and our objectives and dependencies, which helps to solve business-wide problems. The decentralized format of DnA is a great enabler, with its cross-functional data product teams and Snowflake as our data platform. The path to standardizing how data should live, breathe and be accessible is bringing teams and systems together – the core of any sales and operations planning process.
Bigger than measurement
Measuring error is propelling us forward. Techniques like mean absolute percent error (MAPE) and mean absolute error (MAE) are fundamental to success so far, but it’s one part of a more complex story.
Forecasting is a long-term journey, more truthfully navigated through rich rather than transactional relationships with stakeholders because it’s the adoption of data products that benefit all. What good is an unused forecast sitting on a virtual shelf collecting dust? The data product by itself unlocks no value. In the end, we’re business partners. Here to smooth, polish and power better execution.
Discover more Vista transformation stories on our blog or explore current job openings at DnA.