Time Series Forecasting Experiments

An experimental project exploring and comparing different time series forecasting techniques on real-world air passenger traffic data. This project serves as a hands-on laboratory for understanding the strengths and limitations of classical statistical methods versus modern machine learning approaches.

Overview

This project is a practical exploration of time series analysis, where I experiment with three distinct forecasting methodologies:

• Classical Decomposition (Trend + Seasonality)
• ARIMA time series modeling
• Machine Learning algorithms (Linear Regression, Random Forest, SVR, XGBoost)

The goal is not just to predict, but to understand how different techniques handle temporal patterns, trends, and seasonality in real data.

Dataset

• Source: US Air Passengers Monthly.csv (original source unknown, but available via Google Drive)
• Training Period: 2016-2018 (monthly aggregated data)
• Test Period: 2019 (first 6 months)
• Target Variable: Total passenger count (Sum_PASSENGERS)

Note: I couldn't locate the original data source, but the dataset is available for download via the Google Drive link above.

Experimental Methodology

1. Data Preprocessing

• Conversion of date columns to datetime index
• Monthly resampling and aggregation
• Train/test split (2016-2018 vs 2019)

2. Stationarity Analysis

• Augmented Dickey-Fuller (ADF) Test: Tests for stationarity
• Seasonal Decomposition: Separates the series into:
  • Trend component (linear)
  • Seasonal component
  • Residual component

3. Forecasting Approaches

Experiment 1: Trend + Seasonality

• Linear regression to model and extrapolate trend
• Addition of historical seasonal patterns
• Simple, interpretable approach

Experiment 2: ARIMA Model

• Auto ARIMA for parameter selection
• ARIMA(2,0,2) fitted on detrended series
• Trend component added back to final predictions

Experiment 3: Machine Learning

Models tested:

• Linear Regression
• Random Forest Regressor
• Support Vector Regressor (SVR)
• XGBoost Regressor

Project Structure

traffic_prediction/
├── main.ipynb      # Main analysis notebook
├── .gitignore      # Git ignore file
└── README.md       # This file

Note: The dataset (US Air Passengers Monthly.csv) is not included in this repository. Please download it from the Google Drive link and place it in the project root directory.

Results

The notebook produces:

• Stationarity test results (ADF test statistics)
• Decomposition plots showing trend, seasonality, and residuals
• ACF/PACF plots for autocorrelation analysis
• RMSE comparison across all prediction methods
• Visual comparison of predictions vs actual values for 2019

Evaluation Metrics

All models are evaluated using Root Mean Squared Error (RMSE) on the 2019 test set.

Key Findings

The results interpretation section in the notebook provides:

• Quantitative RMSE comparison between methods
• Analysis of strengths and weaknesses of each approach
• Insights on model performance and suitability

Future Experiments

Potential directions to explore:

• Include external variables (economic indicators, fuel prices)
• Test additional ML models (LSTM, Prophet)
• Implement cross-validation for time series
• Extend forecasting horizon
• Add confidence intervals to predictions

---

Learn More

For a detailed walkthrough of this project, including insights and lessons learned, check out my blog post:

Time Series Forecasting Experiments: Classical vs ML Approaches