Global Weather Analysis & Forecasting

Overview

The objective of this assessment was to analyze a global weather dataset, build forecasting models, generate advanced insights and demonstrate practical data science skills through end-to-end project execution.

The project includes data cleaning, exploratory data analysis, anomaly detection, machine learning forecasting, ensemble modeling and geographical climate insights.

---

PM Accelerator Mission

PM Accelerator’s mission is to empower future product, AI and technology leaders through hands-on learning experiences, real-world projects and career acceleration opportunities that help candidates build impactful solutions and grow professionally.

---

How to Run

1. Clone the repository
2. Install required libraries:

pip install -r requirements.txt

3. Open the notebook:

weather.ipynb

or run in Google Colab

---

Project Structure

1. Exploratory Data Analysis (EDA)

• Dataset loading and initial inspection (shape, info, null values, summary statistics)
• Domain-based data cleaning : rows with physically impossible values (e.g. temperature outside −50°C to 60°C, wind speed > 200 kph) are removed; only 7 rows were filtered, confirming high dataset quality
• Distribution plots for temperature and precipitation (histograms, KDE curves, violin plots)
• Feature correlation heatmap across all numeric variables
• Top 15 hottest countries by average temperature
• Time-series trends for temperature and precipitation over time
• Scatter plot of temperature vs. precipitation

2. Baseline Time Series Forecasting

Three simple baseline models are built on daily average temperature (80/20 train/test split):

Model	Description
Naive (Last Value)	Repeats the last training value for all future steps
Seasonal Naive	Repeats the last 7-day cycle
Rolling Mean	Forecasts using the 7-day trailing average

All three produced negative R² scores, confirming that simple persistence-based methods are insufficient for complex global weather data.

3. Anomaly Detection

Three methods are applied to identify unusual weather observations:

• Z-Score - flags univariate temperature outliers beyond ±3 standard deviations
• IQR (Interquartile Range) - flags values outside 1.5×IQR from Q1/Q3
• Isolation Forest - multivariate anomaly detection across temperature, wind speed, pressure, humidity and precipitation (contamination = 1%)

Each method's anomalies are visualized as scatter plots of Temperature vs. Humidity. Isolation Forest is noted as the most suitable for complex, multi-dimensional weather data.

4. Advanced Forecasting

Five models are trained and compared using lag features (lag-1, lag-7, 7-day rolling mean):

Model	Notes
ARIMA(5,1,0)	Classical statistical model
Prophet	Facebook's time series model with seasonality
Random Forest	Ensemble of decision trees
Gradient Boosting	Sequential boosting; best individual performer
XGBoost	Optimized gradient boosting

Key finding: Tree-based ML models (Gradient Boosting, XGBoost, Random Forest) significantly outperform ARIMA and Prophet. Gradient Boosting achieves the lowest MAE and RMSE with the highest R². ARIMA and Prophet produce negative R² values due to the dataset's complexity and globally aggregated, multi-region nature.

An Ensemble model averaging Random Forest, Gradient Boosting, and XGBoost predictions is also evaluated. It provides stable performance but does not meaningfully exceed Gradient Boosting alone.

5. Unique Analyses

Climate Analysis

• Monthly global average temperature trend over time
• Top 10 rainiest countries by average precipitation
• Top 10 hottest countries by average temperature
• Global temperature distribution histogram

Environmental Impact

• Correlation heatmap between air quality indicators (PM2.5, PM10, CO, Ozone, NO₂) and weather variables
• Scatter plots of PM2.5 vs. humidity, wind speed, and temperature
• Finding: higher wind speeds generally reduce pollution via dispersion; humidity shows a moderate positive association with particulate accumulation

Feature Importance

• Tree-based importance from Gradient Boosting
• Permutation importance (10 repeats)
• Correlation-based feature analysis
• Lag features and weather variables (humidity, pressure) are consistently the most predictive

Spatial Analysis

• Latitude vs. temperature - clear decrease in temperature toward higher latitudes
• Longitude vs. temperature distribution
• Top 15 rainiest countries
• Geographic heatmaps of temperature and PM2.5 pollution plotted on longitude/latitude axes

Geographical Patterns (Continental)

• Countries mapped to continents (Africa, Asia, Europe, North America, South America, Oceania)
• Average temperature, precipitation, and PM2.5 by continent
• Tropical/equatorial regions (Africa, South America, Asia) show higher temperatures; industrialized regions show elevated PM2.5

---

Requirements

numpy
pandas
matplotlib
seaborn
scikit-learn
scipy
statsmodels
prophet
xgboost

---

Dataset

GlobalWeatherRepository.csv — a multi-country weather dataset with fields including:

• temperature_celsius, humidity, wind_kph, pressure_mb, precip_mm
• air_quality_PM2.5, air_quality_PM10, air_quality_Carbon_Monoxide, air_quality_Ozone, air_quality_Nitrogen_dioxide
• latitude, longitude, country, last_updated

---

Key Takeaways

• The dataset is high quality with very few invalid records.
• Machine learning ensemble methods substantially outperform classical statistical forecasting for complex, globally aggregated weather data.
• Gradient Boosting is the single best-performing model.
• Air quality has a meaningful relationship with weather conditions, particularly wind speed and humidity.
• Global temperature patterns follow expected geographical distributions driven by latitude.