⚠️ This is an imbalanced classification problem (~78% No Rain, ~22% Rain). F1 and ROC-AUC are more meaningful than accuracy alone. The model uses class_weight="balanced" to handle this imbalance.

Confusion Matrix

🗂️ Dataset

Source: weatherAUS.csv
Rows: ~145,000 daily weather observations
Coverage: 49 weather stations across Australia (2007–2017)
Target: RainTomorrow — Yes (1) / No (0)
Processed data: amarshiv86/weather-aus-rain-dataset

Features Used (20 total)

Numeric (16)

Feature	Description
`MinTemp`	Minimum temperature (°C)
`MaxTemp`	Maximum temperature (°C)
`Rainfall`	Rainfall today (mm)
`Evaporation`	Evaporation (mm)
`Sunshine`	Hours of bright sunshine
`WindGustSpeed`	Max wind gust speed (km/h)
`WindSpeed9am`	Wind speed at 9am (km/h)
`WindSpeed3pm`	Wind speed at 3pm (km/h)
`Humidity9am`	Relative humidity at 9am (%)
`Humidity3pm`	Relative humidity at 3pm (%)
`Pressure9am`	Atmospheric pressure at 9am (hPa)
`Pressure3pm`	Atmospheric pressure at 3pm (hPa)
`Cloud9am`	Cloud cover at 9am (oktas 0–8)
`Cloud3pm`	Cloud cover at 3pm (oktas 0–8)
`Temp9am`	Temperature at 9am (°C)
`Temp3pm`	Temperature at 3pm (°C)

Categorical — label-encoded to integer (3)

Feature	Description	Encoding
`WindGustDir`	Direction of strongest wind gust	N=0, NNE=1, … NNW=15
`WindDir9am`	Wind direction at 9am	N=0, NNE=1, … NNW=15
`WindDir3pm`	Wind direction at 3pm	N=0, NNE=1, … NNW=15

Binary (1)

Feature	Description	Encoding
`RainToday`	Did it rain today?	Yes=1, No=0

🤖 Model Details

Property	Value
Model Type	`Pipeline(StandardScaler → RandomForestClassifier)`
Library	scikit-learn
Class Weight	`balanced` (handles rain/no-rain imbalance)
Training Framework	MLflow
Pipeline	GitHub Actions (automated retraining)
Artifact Storage	Hugging Face Hub

The model is wrapped in a scikit-learn Pipeline — raw feature values are passed directly with no manual scaling required.

🚀 How to Use

python

import pickle
import numpy as np
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="amarshiv86/weather-aus-rain-model",
    filename="models/model.pkl",
    repo_type="model"
)
model = pickle.load(open(model_path, "rb"))

# Wind direction encoding
WIND_DIRS = ['N','NNE','NE','ENE','E','ESE','SE','SSE',
             'S','SSW','SW','WSW','W','WNW','NW','NNW']

def wind_to_int(d): return WIND_DIRS.index(d.upper())

# Feature order must match exactly (20 features):
# Numeric (16) + Categorical encoded (3) + Binary (1)
sample = np.array([[
    13.4,  # MinTemp
    22.9,  # MaxTemp
    0.6,   # Rainfall
    5.0,   # Evaporation
    7.5,   # Sunshine
    44.0,  # WindGustSpeed
    20.0,  # WindSpeed9am
    24.0,  # WindSpeed3pm
    71.0,  # Humidity9am
    22.0,  # Humidity3pm
    1007.7,# Pressure9am
    1007.1,# Pressure3pm
    4,     # Cloud9am
    5,     # Cloud3pm
    16.9,  # Temp9am
    21.8,  # Temp3pm
    wind_to_int('S'),    # WindGustDir
    wind_to_int('W'),    # WindDir9am
    wind_to_int('SW'),   # WindDir3pm
    0,     # RainToday (0=No, 1=Yes)
]])

pred = model.predict(sample)[0]
prob = model.predict_proba(sample)[0][pred]
print(f"Rain Tomorrow: {'Yes' if pred == 1 else 'No'} (confidence: {prob:.1%})")

🔁 MLOps Pipeline

This model is automatically retrained via GitHub Actions whenever:

Raw data changes (data/raw/weatherAUS.csv)
Source code changes (src/)
Hyperparameters change (params.yaml)

GitHub Push
    ↓
GitHub Actions
    ↓
prepare.py → train.py → evaluate.py
    ↓                        ↓
model.pkl              metrics.json
                       confusion_matrix.png
    ↓
Hugging Face Hub (this repo)
    ↓
HF Spaces — FastAPI live demo

📁 Repository Structure

amarshiv86/weather-aus-rain-model
├── models/
│   └── model.pkl              # sklearn Pipeline (scaler + RandomForest)
├── artifacts/
│   └── confusion_matrix.png   # latest evaluation plot
└── metrics.json               # latest evaluation metrics

🌐 Live Demo

Try the model live via the FastAPI app hosted on HF Spaces: 👉 amarshiv86/weather-aus-rain-api

📄 License

MIT — free to use, modify, and distribute.