RetailForecast: Production-Ready Sales Forecasting with ZenML and Prophet

A robust MLOps pipeline for retail sales forecasting designed for retail data scientists and ML engineers.

📊 Business Context

In retail, accurate demand forecasting is critical for optimizing inventory, staff scheduling, and financial planning. This project provides a production-ready sales forecasting solution that can be immediately deployed in retail environments to:

• Predict future sales volumes across multiple stores and products
• Capture seasonal patterns and trends in customer purchasing behavior
• Support data-driven inventory management and purchasing decisions
• Provide actionable insights through visual forecasting dashboards

HTML dashboard visualization showing forecasts with uncertainty intervals

🔍 Data Overview

The pipeline works with time-series retail sales data structured as follows:

Field	Description
date	Date of sales record (YYYY-MM-DD)
store	Store identifier (e.g., Store_1, Store_2)
item	Product identifier (e.g., Item_A, Item_B)
sales	Number of units sold
price	Unit price

The system automatically handles:

• Multiple store/item combinations as separate time series
• Train/test splitting for model validation
• Proper data transformations required by Prophet
• Missing value imputation and outlier detection

Interactive visualization of historical sales patterns

🚀 Pipeline Architecture

The project includes two primary pipelines:

1. Training Pipeline

The training pipeline performs the following steps:

1. Data Loading: Imports historical sales data from CSV files
2. Data Preprocessing:
   • Transforms data into Prophet-compatible format
   • Creates separate time series for each store-item combination
   • Performs train/test splitting based on configurable ratio
3. Model Training:
   • Trains multiple Facebook Prophet models simultaneously, one for each store-item combination
   • Configures seasonality parameters based on domain knowledge
   • Handles price changes as regressors when available
4. Model Evaluation:
   • Calculates MAPE, RMSE, and MAE metrics on test data
   • Generates visual diagnostics for model performance
5. Forecasting:
   • Produces forecasts with uncertainty intervals
   • Creates interactive HTML visualizations

ZenML visualization of the training pipeline DAG

2. Inference Pipeline

The inference pipeline enables fast forecasting with pre-trained models:

1. Data Loading: Imports the most recent sales data
2. Data Preprocessing: Transforms data into Prophet format
3. Forecasting: Generates predictions using production models
4. Visualization: Creates interactive dashboards with forecasts

ZenML visualization of the inference pipeline DAG

📈 Model Details

The forecasting solution uses Facebook Prophet, chosen specifically for its combination of accuracy and simplicity in retail forecasting scenarios:

• Multiple Models Approach: Rather than a one-size-fits-all model, we generate individual Prophet models for each store-item combination, allowing forecasts that capture the unique patterns of each product in each location
• Components: Prophet automatically decomposes time series into trend, seasonality, and holidays
• Seasonality: Captures weekly, monthly, and yearly patterns in sales data
• Special Events: Handles holidays and promotions as custom seasonality effects
• Uncertainty Estimation: Provides prediction intervals for better inventory planning
• Extensibility: Supports additional regressors like price and marketing spend

Prophet was selected for this solution because it excels at:

• Handling missing data and outliers common in retail sales data
• Automatically detecting seasonal patterns without extensive feature engineering
• Providing intuitive parameters that business users can understand
• Scaling to thousands of individual time series efficiently

💻 Technical Implementation

The project leverages ZenML's MLOps framework to provide:

• Model Versioning: Track all model versions and their performance metrics
• Reproducibility: All experiments are fully reproducible with tracked parameters
• Pipeline Caching: Speed up experimentation with intelligent caching of pipeline steps
• Artifact Tracking: All data and models are properly versioned and stored
• Deployment Ready: Models can be directly deployed to production environments

A key innovation in this project is the custom ProphetMaterializer that enables serialization/deserialization of Prophet models for ZenML artifact storage.

ZenML model registry tracking model versions and performance

🛠️ Getting Started

Prerequisites

• Python 3.9+
• ZenML installed and configured

Installation

# Clone the repository
git clone https://github.com/zenml-io/zenml-projects.git
cd zenml-projects/retail-forecast

# Install dependencies
pip install -r requirements.txt

# Initialize ZenML (if needed)
zenml init

Running the Pipelines

To train models and generate forecasts:

# Run the training pipeline (default)
python run.py

# Run with custom parameters
python run.py --forecast-periods 60 --test-size 0.3 --weekly-seasonality True

To make predictions using existing models:

# Run the inference pipeline
python run.py --inference

Viewing Results

Start the ZenML dashboard:

zenml login

Navigate to the dashboard to explore:

• Pipeline runs and their status
• Model performance metrics
• Interactive forecast visualizations
• Version history of all models

🔄 Integration with Retail Systems

This solution can be integrated with existing retail systems:

• Inventory Management: Connect forecasts to automatic reordering systems
• ERP Systems: Feed forecasts into financial planning modules
• BI Dashboards: Export forecasts to Tableau, Power BI, or similar tools
• Supply Chain: Share forecasts with suppliers via API endpoints

📊 Example Use Case: Store-Level Demand Planning

A retail chain with 50 stores and 500 products uses this pipeline to:

1. Train models on 2 years of historical sales data
2. Generate daily forecasts for the next 30 days for each store-item combination
3. Aggregate forecasts to support central purchasing decisions
4. Update models weekly with new sales data

The result: 15% reduction in stockouts and 20% decrease in excess inventory.