Global Temperature Forecasting

Machine Learning & Time-Series Techniques

A complete end-to-end data science project that analyses 175 years of global surface temperature data (1850–2026) and builds multiple forecasting models to predict future climate trends.

Project Highlights

• Zero data preprocessing headaches — uses the NOAA GCAG dataset with no missing values
• 5 models compared — Linear Regression, Random Forest, SVR, ARIMA, and Facebook Prophet
• Hyperparameter tuning via GridSearchCV (not arbitrary defaults)
• Proper time-series methodology — no-shuffle train/test split, lag features, stationarity testing
• Future forecast up to 2030 with a correctly implemented rolling prediction loop

Dataset

Property	Details
Source	NOAA GCAG via datahub.io
Coverage	January 1850 — present (updated monthly)
Rows	~2,100+ months
Missing values	None
Target variable	Temperature anomaly (°C deviation from 20th-century average)

No download required. The dataset is fetched automatically inside the notebook via a single URL.

Models Used

Model	Type	Notes
Linear Regression	ML Baseline	Interpretable, fast
Random Forest	ML Ensemble	Best accuracy; tuned with GridSearchCV
SVR	ML Kernel	Robust to outliers
ARIMA(2,1,2)	Classical Time-Series	Principled sequential model
Facebook Prophet	Time-Series	Handles trend + seasonality automatically

Project Structure

global-temperature-forecasting/
│
├── Weather_Forecasting_NOAA_Portfolio.ipynb   
├── requirements.txt                           
├── README.md                                  
└── .gitignore                                 

📈 Key Findings

• Global surface temperatures have risen +1.3°C above the 20th-century baseline
• Warming rate since 1980 is approximately +0.2°C per decade
• Random Forest achieved the best predictive accuracy on test data
• The 12-month lag feature was the most important predictor, capturing annual seasonality

🛠️ Tech Stack

• Python 3.10+
• pandas, numpy — data manipulation
• matplotlib, seaborn — visualisation
• scikit-learn — ML models and evaluation
• statsmodels — ARIMA, ADF test, seasonal decomposition
• prophet — Facebook Prophet forecasting