This project investigates how different outlier treatment techniques affect the performance of machine learning models.

We perform a comparative study using real-world weather data and evaluate how preprocessing decisions influence predictive accuracy.

The dataset used in this project was obtained from NASA's POWER (Prediction Of Worldwide Energy Resources) Data Access Viewer:

👉 https://power.larc.nasa.gov/data-access-viewer/

NASA POWER provides satellite-based meteorological data, widely used in climate, agriculture, and energy research.

• 📍 Location: Hyderabad, India
• 📅 Time Period: 1984 – 2025 (Daily data)
• 📁 Format: CSV
• 📡 Source: NASA Langley Research Center

• T2M_MAX → Maximum Temperature
• T2M_MIN → Minimum Temperature
• RH2M → Humidity
• PRECTOTCORR → Rainfall
• WS2M → Wind Speed
• ALLSKY_SFC_SW_DWN → Solar Radiation

• Analyze the impact of outliers on model performance
• Compare different outlier treatment techniques
• Evaluate model sensitivity to outliers
• Provide practical insights for preprocessing decisions

• Interquartile Range (IQR)

• Trimming (Removal of outliers)
• Winsorization (Capping extreme values)
• Log Transformation
• Box-Cox Transformation

• Multiple Linear Regression
• XGBoost (Extreme Gradient Boosting)

👉 Best Methods: Winsorization & Box-Cox

👉 Best Method: Log Transformation

• Linear Regression is highly sensitive to outliers
• Winsorization and Box-Cox improve linear model performance
• XGBoost is robust to outliers
• Tree-based models require less preprocessing
• Outlier treatment should be applied based on model type

Outlier handling is not a one-size-fits-all approach.

• For Linear Models → Proper treatment improves accuracy
• For Tree-Based Models → Minimal preprocessing is sufficient

Careful selection of preprocessing techniques leads to better model performance.

• Python / R
• Pandas, NumPy
• Scikit-learn
• XGBoost
• Matplotlib, Seaborn

Aditya Charan Eranki

M.Sc Mathematics (Data Science) Vellore Institute of Technology