Demand Forecasting System with Automated Retraining and Cloud-Based MLOps

---

Overview

This project implements a production-ready demand forecasting system using Facebook Prophet and MLflow, designed to follow real-world MLOps best practices.

The system supports:

• Automated retraining
• Model evaluation and versioning
• Controlled model promotion
• Cloud-based artifact storage
• Remote inference
• A user-facing web application for interactive forecasting

The primary objective of this project is to demonstrate how a time-series forecasting solution can be built, deployed, governed, and consumed in a scalable and maintainable manner.

---

Problem Statement

Accurate demand forecasting is critical for inventory planning, logistics, and operational efficiency.

This system forecasts daily demand for multiple store–item combinations, ensuring:

• Independent modeling per store and item
• Controlled model evolution over time
• Safe deployment of improved models
• Reproducibility and traceability

---

Key Features

• Time-series forecasting using Facebook Prophet
• Separate forecasting models per Store × Item
• Automated retraining triggered by new data
• Model evaluation using RMSE
• Automatic promotion of superior models
• MLflow-based experiment tracking and model registry
• AWS S3-backed artifact storage
• Remote inference through MLflow Tracking Server
• Interactive forecasting interface using Streamlit

---

Data Description

• Time period: 2013 – 2017
• Forecast horizon: Up to 2018
• Stores: 1 to 3
• Items: 1 to 50
• Granularity: Daily demand

Each store–item pair is treated as an independent time series to preserve demand patterns and seasonality.

---

Model Training & Validation

• Model: Facebook Prophet
• Validation strategy: Last 30-day holdout
• Metric: Root Mean Squared Error (RMSE)

---

Model Promotion Logic

Promotion Rules

• If no existing model exists → Promote new model
• If new model RMSE is lower → Promote new model
• Otherwise → Retain current model

The promoted model is assigned the champion alias in MLflow.

---

Inference Design

• Remote inference via MLflow Tracking Server
• Explicit user-defined forecast dates
• Forecasting restricted to 2013–2018
• No reliance on system date

---

Streamlit Web Application

User Capabilities

• Select Store ID (1–3)
• Select Item ID (1–50)
• Choose date range within 2013–2018
• Visualize forecasted demand
• View forecast data table

---

Deployment Architecture

• Training & retraining: EC2
• Tracking server: MLflow on EC2
• Artifact storage: AWS S3
• Inference: Remote via MLflow
• UI: Streamlit

---

Engineering Principles Applied

• End-to-end MLOps lifecycle
• Automated retraining & evaluation
• Version-controlled deployment
• Cloud-native artifacts
• Reproducible workflows