FloraCast is an end-to-end forecasting solution that demonstrates how to build production-grade time series prediction workflows with ZenML and Darts. It provides reproducible training, automatic model versioning, and scheduled batch inference with rich visualizations to support demand and sales forecasting use cases in retail, e-commerce, and supply chain.
What It Does
-
Trains advanced forecasting models (Temporal Fusion Transformer by default; ExponentialSmoothing fallback)
-
Evaluates models with SMAPE and produces clear visual reports
-
Registers and promotes the best-performing models using ZenML's Model Control Plane
-
Runs automated batch inference to generate future forecasts on a schedule
How It Works
-
Ingests ecommerce sales data and converts it into Darts TimeSeries objects
-
Applies standardized preprocessing, including train/validation split and frequency alignment
-
Trains a configurable TFT model and computes evaluation metrics
-
Stores artifacts with custom materializers, including timeseries visualizations
-
Loads the production model for batch inference and exports predictions to CSV
Architecture
The architecture connects data ingestion, preprocessing, model training, evaluation, and scheduled batch inference through ZenML pipelines, leveraging the Model Control Plane for lineage, versioning, and promotion.
Gallery
