Omlet designs, manufactures, and distributes pet products directly to consumers across multiple markets. As sales scaled, their legacy forecasting approach, built on linear regression and moving averages, couldn't capture the complex demand patterns of modern e-commerce. New Gradient built an AI demand forecasting system using the Temporal Fusion Transformer architecture, delivering 46% more accurate predictions and enabling intelligent fulfilment decisions from one week to twelve months ahead.
Legacy forecasting fails at e-commerce scale
Omlet's existing system relied on traditional statistical methods that struggled to capture the non-linear relationships inherent in modern e-commerce demand patterns. This resulted in inventory imbalances: overstocking of slow-moving SKUs whilst understocking high-demand products.
Fulfilment inefficiencies followed, with suboptimal warehouse allocation and shipping route planning. Manufacturing schedules became disconnected from actual demand signals, leading to missed revenue opportunities through stockouts during peak periods and promotional campaigns. The business needed a forecasting solution capable of learning from historical patterns, adapting to seasonal fluctuations, and incorporating multiple external variables.
Temporal Fusion Transformers for multi-horizon prediction
We designed and implemented an end-to-end machine learning forecasting pipeline centred on the Temporal Fusion Transformer (TFT) architecture, a deep learning model purpose-built for multi-horizon time series prediction. Data engineering work consolidated disparate sources into a unified forecasting dataset: historical sales across product categories and customer segments, promotional calendars, seasonal patterns, and external variables including regional demand indicators.
The TFT model generates accurate predictions from one week to twelve months ahead, with interpretable outputs that quantify the influence of each input variable. It captures complex temporal dependencies that traditional methods miss and incorporates known future inputs such as planned promotions, holidays, and events.
Forecasting outputs directly inform inventory replenishment through automated reorder triggers, warehouse allocation through proactive stock positioning, manufacturing scheduling through demand-driven production planning, and logistics optimisation through improved carrier selection and delivery route planning.
46% more accurate forecasts, deployed in 3.5 months
Delivered in 3.5 months from project inception to production deployment, the solution achieved a 46% reduction in forecasting error compared to the baseline approach. The model now generates predictions across a full 1-year horizon, enabling long-range production planning alongside short-term inventory decisions.
The accuracy improvement has a direct knock-on effect to inventory holding costs through optimised stock levels, while protecting against stockout incidents during peak trading periods. Manufacturing output is now better aligned with actual demand signals, reducing both waste from overproduction and missed revenue from understocking.