rollingcv 0.1.0

Fixed-window cross-validation for time series (scikit-learn compatible)

RollingWindowSplit

A smarter, production-ready alternative to scikit-learn's TimeSeriesSplit.

Perform rolling window cross-validation with fixed-size training sets and forecast horizons. Fully compatible with scikit-learn pipelines and supports percent-based sizes, configurable gaps, and clean fold visualization — even without matplotlib.

---

🚀 Why RollingWindowSplit?

When you use TimeSeriesSplit, each fold has a larger training set than the previous. That means:

• You're training different models on different data sizes in each fold
• You're not properly evaluating the model you plan to deploy
• You're not reproducing the same model k times

With RollingWindowSplit, you:

• ✅ Fix the training window size (just like in production)
• ✅ Fix the forecast horizon (true out-of-sample evaluation)
• ✅ Optionally add a gap to avoid data leakage
• ✅ Run folds that actually replicate your deployment logic

---

📦 Installation

Install via pip:

pip install rollingcv

Or install locally for development:

pip install -e .

---

💡 Example

import numpy as np
from rollingcv import RollingWindowSplit

data = np.arange(1000)

rws = RollingWindowSplit(n_splits=5, window_size=0.6, horizon=0.1, gap=5)

# Show a text preview of the folds
rws.preview(data, style='default')

# Show a bar-style fold preview in console
rws.preview(data, style='bar')

---

🔍 Console Preview (Bar Style)

RollingWindowSplit Visual Preview (width=80):

Fold  1: ====================-----                    
Fold  2:   ====================-----                  
Fold  3:     ====================-----                
Fold  4:       ====================-----              
Fold  5:         ====================-----            

---

🧠 Key Features

• ✅ Fixed or percent-based window and horizon sizes
• ✅ Optional gap to simulate production delays
• ✅ Clean __repr__ for logging/debugging
• ✅ Compatible with scikit-learn pipelines
• ✅ Console previews — no matplotlib required

---

🛡️ Error Handling

RollingWindowSplit protects against common pitfalls:

• n_splits must be at least 2
• window_size, horizon, and gap must be non-negative
• Float sizes must be between 0 and 1
• Raises a friendly message if there's not enough data to split

---

🧪 Testing

Run the built-in unit tests with:

pytest tests/

---

📄 License

MIT License. Use it, modify it, love it.

---

✨ Contribute

Pull requests welcome! Fork it, play with it, and if you improve it — share it back!

---

🙌 Credits

Created with care by marianotir