Rainfall Occurrence Prediction (Binary Classification)

This project develops an end-to-end machine learning pipeline to predict the probability of rainfall occurrence using tabular environmental data.
The focus is on building a clean, interpretable baseline model and evaluating probability quality rather than maximizing leaderboard performance.

---

Project Overview

• Task: Binary classification — predict whether rainfall occurs
• Output: Probabilistic predictions (likelihood of rainfall)
• Approach: Baseline Logistic Regression with a full preprocessing pipeline
• Evaluation: ROC-AUC, calibration analysis, and threshold-based error trade-offs

This project was conducted as a self-directed machine learning workflow and emphasizes reproducibility, interpretability, and decision-oriented evaluation.

---

Dataset

• Tabular environmental features (e.g., temperature, humidity, pressure, wind)
• Target variable: rainfall (binary)
• Dataset size: 2,190 observations
• No missing values in training data; missing values in test data handled via imputation in the pipeline

---

Workflow Summary

1) Data Inspection

• Verified dataset schema and feature types
• Identified non-predictive identifier column (id)
• Checked class balance and data quality

2) Exploratory Data Analysis (Minimal)

• Visualized target distribution
• Inspected numeric feature distributions
• Confirmed no obvious anomalies prior to modeling

3) Baseline Model

Logistic Regression with a preprocessing pipeline:

• Median imputation for numeric features
• Most-frequent imputation for categorical features
• Standard scaling and one-hot encoding
• Stratified train/validation split

4) Model Evaluation

• Baseline performance: ROC-AUC ≈ 0.88 (validation split)
• ROC curve used to visualize class separation

5) Probability Calibration

• Examined a calibration curve to assess the reliability of predicted probabilities
• Observed mild under-confidence in mid-range probabilities

6) Threshold Analysis

• Evaluated confusion matrices across multiple decision thresholds
• Analyzed trade-offs between false positives and false negatives
• Demonstrated how threshold choice affects outcomes

7) Prediction on Unseen Data

Generated probabilistic rainfall predictions for unseen samples. Output format:

id,rainfall_probability
2190,0.9449
2191,0.9519
2192,0.7798
...

Predictions were saved as a reusable CSV artifact.

---

Key Results

• Strong baseline classification performance (ROC-AUC ≈ 0.88)
• Probability outputs are reasonably calibrated
• Clear threshold-dependent error trade-offs
• End-to-end pipeline from data loading to prediction generation

---

Why Logistic Regression?

Logistic Regression was intentionally chosen as a baseline model to:

• Provide interpretable coefficients
• Produce well-defined probability estimates
• Support calibration and threshold-based analysis

This establishes a reliable reference point for future model extensions.

---

Outputs

• outputs_rainfall/predictions.csv
  Probabilistic rainfall predictions on unseen data.

• images/roc_curve.png (optional)
  ROC curve exported from the notebook.

• images/calibration_curve.png (optional)
  Calibration curve exported from the notebook.

If you don’t commit images, the README will still render fine—those sections will simply show empty image placeholders.

---

Reproducibility Notes

• A fixed random seed is used for the train/validation split.
• Missing values in unseen data are handled using imputers inside the preprocessing pipeline.

---

Takeaways

This project demonstrates a complete and reproducible ML workflow:

• From tabular data to calibrated probabilistic predictions
• With explicit evaluation and threshold-based decision considerations

It serves as a solid baseline for future model comparisons, feature engineering, or domain-specific tuning.