📊 End-to-End Time Series Forecasting System with API

🚀 Overview

This project implements a complete end-to-end time series forecasting system that predicts the next 8 weeks of sales for each state using historical data.

The system covers the full pipeline:

• Data preprocessing
• Feature engineering
• Model training & comparison
• Automatic model selection
• Deployment via a REST API

---

🎯 Objective

Build a production-ready forecasting system that:

• Trains multiple forecasting algorithms
• Compares and selects the best model
• Exposes predictions via a REST API
• Follows a real backend service structure

---

📊 Dataset

• Source: Provided Excel dataset
• Contains historical state-wise sales data

---

🧹 Data Preprocessing

• Handled missing values using forward fill (ffill)
• Resampled data to ensure weekly consistency
• Ensured proper datetime indexing

---

⚙️ Feature Engineering

Created time-based features to capture temporal patterns:

• Lag features: t-1, t-7, t-30
• Rolling statistics: mean & standard deviation
• Month extraction
• Time-series aware train-validation split (no leakage)

---

🤖 Models Implemented

The system trains and compares:

• SARIMA (Statistical Model)
• Facebook Prophet (Trend + Seasonality)
• XGBoost (Feature-based ML)
• LSTM (Deep Learning sequence model)

---

📈 Model Selection

• Evaluation metric: RMSE (Root Mean Squared Error)
• Best model automatically selected based on performance
• SARIMA achieved the lowest RMSE and was deployed

---

💾 Model Saving

• Models saved using pickle format (.pkl)
• Due to size constraints, a subset of states is included for demo
• System is designed to scale to all states

---

⚡ API Deployment

Built using FastAPI

Endpoint:

GET /predict?state=&steps=

Example:

Output:

{
  "state": "California",
  "forecast": [values...]
}