Global Weather Repository

Advanced Data Science Assessment

PM Accelerator | AI Engineer Internship Technical Assessment

---

PM Accelerator Mission

PM Accelerator empowers aspiring and experienced product managers with the tools, community, and real-world experience needed to accelerate their careers. Through hands-on projects, mentorship, and a global network, PM Accelerator bridges the gap between ambition and achievement.

---

Project Overview

This project builds a complete machine learning pipeline to forecast daily temperatures across 191 countries using the Global Weather Repository dataset — 131,365 daily weather snapshots with 41 features.

Best model: Gradient Boosting — MAE = 2.61°C, R² = 0.886 on held-out test data (Nov 2025 → Mar 2026)

---

Repository Structure

├── Global_Weather_Repo.ipynb   # Full analysis notebook
├── README.md                   # This file
├── requirements.txt            # Python dependencies

Note: The dataset is not included. Download GlobalWeatherRepository.csv from Kaggle and place it at /content/GlobalWeatherRepository.csv (or update the path in Cell 3).

---

What's Inside the Notebook

1. Data Cleaning & Preprocessing

• Parsed last_updated datetime and extracted time features
• Removed unit duplicates (°F columns, kph columns, imperial pressure/precipitation)
• Dropped high-correlation redundants confirmed by EDA (r > 0.90)
• Removed 3 physically impossible sensor readings (wind > world record, pressure outside valid range) — 0.002% of data
• Fixed 25+ country name typos discovered during EDA (multilingual API aggregation artifacts)
• Leakage-safe: all statistical operations (scaling, imputation) fitted on train set only

2. Exploratory Data Analysis

• Target distribution: Left-skewed (skewness = -0.88) — explained by geographic concentration of cities in warm latitudes
• Seasonality: Clear 12-month cycle; Northern/Southern hemisphere seasons cancel globally, justifying cyclical encoding
• Correlation analysis: UV index (r=0.489) and pressure (r=-0.421) are strongest predictors; year dropped after discovering a sampling artifact (uneven city coverage across years)
• Geography: Latitude vs mean temperature per city — desert/altitude effects explain why latitude alone gives only r=-0.38
• Anomaly detection: Three methods compared — Z-score (7%), IQR (37.5%), Isolation Forest (3%); IF used as primary method for multivariate anomaly detection

3. Feature Engineering

• Cyclical time encoding: sin/cos transforms for month, hour, and day-of-year — preserves adjacency of Dec/Jan, 23:00/00:00
• Anomaly flag: is_anomaly composite feature — model learns from extreme weather events rather than ignoring them

4. Forecasting Models

Model	MAE (°C)	RMSE (°C)	R²
Baseline (mean)	9.47	12.85	-0.326
Ridge Regression	6.01	8.00	0.487
Random Forest	3.29	4.75	0.819
Gradient Boosting	2.61	3.76	0.886
Stacking Ensemble	3.00	4.21	0.858

Time-based 80/20 split: train on past (May 2024 → Nov 2025), test on future (Nov 2025 → Mar 2026).

5. Advanced Analyses

• Feature importance: Random Forest impurity + SHAP values — latitude and day-of-year cyclical encoding are dominant features
• Climate analysis: Seasonal patterns by hemisphere, temperature range by climate zone (tropical/subtropical/temperate/polar)
• Air quality: PM2.5 vs visibility, ozone vs temperature, pollutant correlations with weather parameters
• Spatial analysis: Interactive world map of mean temperature and precipitation per city

---

How to Run

Option A — Google Colab (recommended)

1. Upload Global_Weather_Repo.ipynb to colab.research.google.com
2. Upload GlobalWeatherRepository.csv to /content/
3. Runtime → Run all

Option B — Local

git clone <repo-url>
cd <repo-folder>
pip install -r requirements.txt
jupyter notebook Global_Weather_Repo.ipynb

Update the CSV path in Cell 3 from /content/GlobalWeatherRepository.csv to your local path.

---

Key Findings

1. Latitude alone is insufficient — desert cities at 25°N (Riyadh, Kuwait) are as hot as tropical cities despite higher latitude. Longitude and climate zone context matter.
2. Month has near-zero global correlation with temperature — Northern and Southern hemisphere seasons cancel out. Cyclical encoding combined with latitude captures the real signal.
3. Year was dropped — appeared to negatively correlate with temperature, but this was a sampling artifact: 2024 had 47 more cities than 2025, artificially inflating 2024 means.
4. Isolation Forest outperforms univariate methods — catches unusual combinations (high temperature + high humidity) invisible to Z-score and IQR applied column-by-column.
5. 2.61°C MAE is strong for global forecasting — the dataset spans a 62°C temperature range across 191 countries; 2.61°C represents ~4.2% of that range.

---

Requirements

See requirements.txt for full list. Core dependencies:

• pandas, numpy, matplotlib, seaborn, plotly
• scikit-learn, shap

---

Author

Nayra Saadawy

Built as part of the PM Accelerator AI Engineer Internship Technical Assessment.