Flieber
Flieber modernized its forecasting offering with a custom TimeGPT model to generate daily demand forecasts up to one year ahead across more than 20,000 product-level time series, improving accuracy and scalability while reducing operational complexity.
34%Reduction in Forecast Error
Reduced forecast error by 34% compared to its previous approach, measured by MAE% and Bias%.
365 daysForecast Horizon
Generate forecasts up to 365 days ahead across over 20,000 product-level time series.
20,000+Time Series Forecasted
Operate a unified, production-ready forecasting pipeline with lower maintenance overhead and faster client onboarding.
<5 minForecasting Runtime
Shorten forecasting inference runtime from five hours to under five minutes.
Overview
"Nixtla has become a strategic partner for Flieber. Together, we bring our customers cutting-edge predictive capabilities that set a new standard in demand planning and inventory management." - Fabricio Miranda, CEO of Flieber
Flieber is a multichannel inventory planning platform that helps retail brands streamline supply chain decisions. Its platform integrates sales, inventory, and supply chain data across multiple channels to support demand forecasting, inventory optimization, and replenishment planning.
Accurate demand forecasting is a critical part of Flieber's offering. Forecasts directly inform inventory planning, helping customers balance product availability while preventing stockouts and overstocking. This requires forecasting models that can operate reliably across large product catalogs with thousands of SKUs and long planning horizons.
The Challenge
Flieber's forecasting problem involves several layers of complexity. Before adopting TimeGPT, Flieber relied on a large ensemble of statistical models supported by extensive preprocessing and postprocessing steps. While flexible, this approach introduced significant complexity. Model selection and maintenance were resource-intensive, and end-to-end forecasting pipelines could take several hours to run.
- · Holiday and event-driven demand. Fixed holidays such as Christmas and New Year's, along with major commercial events like Black Friday, Cyber Monday, and Amazon Prime Day, can cause sharp demand spikes.
- · Long forecasting horizons. Customers require daily forecasts up to one year ahead to support procurement and inventory planning.
- · Sparse and intermittent demand. Many products exhibit low, irregular sales or limited historical data.
- · Operational scale. Forecasts must be generated across tens of thousands of product, seller, and tenant combinations.
The Solution
Flieber adopted a custom implementation of TimeGPT as a unified forecasting engine, replacing its previous ensemble-based pipeline. The system generates forecasts across multiple levels of granularity, including tenant, seller, and product combinations, covering more than 50 tenants, more than 200 sellers, and more than 12,000 products. In total, this represents more than 20,000 time series requiring long-horizon forecasts, each with unique characteristics such as level, trend, seasonality, intermittency, and varying amounts of historical data. Working closely with Flieber's team, several key features were tailored to address Flieber's specific requirements:
- · Holidays and events modeling. The system captures the impact of both fixed holidays, such as Christmas and New Year's, and moving commercial events, such as Prime Day and Black Friday. Because these events occur on different dates each year, the model learns their historical impact on demand and transfers that learned effect to future occurrences.
- · Long-horizon forecasting with temporal reconciliation. To support daily forecasts up to one year ahead, the system uses temporal reconciliation across multiple time granularities. This improves long-horizon accuracy by combining stable longer-term patterns with higher-frequency demand signals, while keeping forecasts coherent across time scales.
- · Integration of domain knowledge. The model incorporates Flieber's supply chain expertise, including adjustments for stockouts, price changes, and outliers caused by unexpected events, enabling forecasts to better reflect true demand rather than raw sales signals.
Business Outcomes
[01]
Improved Forecast Accuracy
- Compared with its previous approach, Flieber improved forecast accuracy by approximately 34%, as measured by MAE% and Bias%.
[02]
Faster Forecasting Cycles
- Forecasting cycles that once took five hours can now be completed in less than 5 minutes, enabling faster iteration and more responsive planning.
- TimeGPT also reduced operational complexity.
- By consolidating a large collection of statistical models into a unified forecasting system, Flieber simplified maintenance, reduced compute requirements, and improved reproducibility.
[03]
Better Outcomes for Customers
- For Flieber's customers, these improvements create clear business value.
- Better forecasts support more efficient inventory allocation, lower the risk of stockouts, reduce excess inventory, and help improve service levels while lowering inventory costs.
"The TimeGPT engine enabled us to build a much more powerful and accurate model that cut our runtime from five hours to under five minutes."
CEO, Flieber