tsbootstrap roadmap 2025-2026

[astrogilda]

This is an updated roadmap for tsbootstrap development, replacing the 2024-2025 roadmap in #144.

Immediate Priorities (Q3 2025)

• Complete test coverage: Achieve >90% test coverage for all core modules
• Test coverage for model_selection module: Currently has no tests
• Documentation: Complete API documentation and user guide
• User Guide/Cookbook: Add practical examples (e.g., "Bootstrapping for financial returns", "Validating a forecasting model")
• Bug fixes: Address known issues ([BUG] index set returned by some bootstraps is list of numpy arrays, not single numpy array #81 index sets, bootstrap classes: running methods multiple times has unexpected side effects #71 side effects, ARIMA LU decomposition issue #41 LU decomposition)
• Sparse array support: Add support for scipy sparse arrays for memory-efficient large datasets
• DataFrame support via Narwhals: Use Narwhals to support both pandas and Polars DataFrames with a unified API
• Replace TSFit with StatsForecast: Migrate from custom TSFit implementation to StatsForecast for better maintained and more robust time series models
• Standardize X and y conventions: Ensure consistency with scikit-learn across all estimators
• Tighten type hints: Use numpy.typing (npt.NDArray) instead of generic np.ndarray
• Add coverage.py to CI: Enforce minimum test coverage and prevent regressions
• Enhanced input validation: Add robust validation with clear error messages

Model Enhancement & Optimization (Q4 2025)

• Multivariate enhancements: Better support for multivariate time series with new data structures
• Optimal block length selection (Research and implement optimal block length selection methods #106): Implement data-driven methods for automatic block length selection
• Adaptive block length: Basic implementation based on autocorrelation structure
• Consolidate bootstrap implementations: Reduce duplication between bootstrap.py, base_bootstrap.py, async_bootstrap.py
• Profile and optimize core loops: Focus on performance bottlenecks in block_bootstrap
• Data preprocessing helpers: Add utilities for differencing, detrending, seasonal adjustment

New Bootstrap Methods (Q4 2025 - Q1 2026)

• Frequency-domain bootstrap (Implement frequency-domain bootstrapping #107): Implement spectral density based bootstrapping
• Local block bootstrap (Implement local block bootstrap #105): Add locally weighted block bootstrap
• Generalized block bootstrap (Implement generalized block bootstrap #104): Implement the generalized block bootstrap method
• Tapered block bootstrap: Complete implementation using existing windowing functions

Performance & API Improvements (Q1 2026)

• Memory optimization: Implement chunked processing for large datasets
• Parallel processing: Add multiprocessing support for multiple bootstrap samples
• Numba acceleration: Use numba for performance-critical loops
• Streaming updates: Basic support for updating bootstrap with new data
• Serialization: Add proper model serialization/deserialization
• Async improvements: Enhance async bootstrap implementations
• Integration tests: Add complex multi-step scenarios (data simulation → model fitting → bootstrapping)
• Unify fit, predict, bootstrap methods: Consistent signatures across all estimators
• Documentation for model_selection: Complete module documentation with examples

Integration & Ecosystem (Q2 2026)

• scikit-learn compatibility: Ensure full compatibility with sklearn pipelines
• Basic sktime integration: Add adapters for common sktime forecasters
• Evaluation framework: Add statistical tests and evaluation metrics
• Cross-validation for block length: Add CV methods for block length tuning
• Visualization utilities: Bootstrap distributions, time series plots, model diagnostics
• Plugin architecture: Replace factory pattern with extensible plugin system
• Automated performance benchmarking: CI step for tracking performance regressions
• Refactor bootstrap_factory.py: Move to more extensible registration pattern

Future Considerations (2026+)

• Advanced adaptive methods: Variance/skewness-based adaptive resampling
• Probabilistic models: Bootstrap-based uncertainty quantification
• Time series augmentation: Data augmentation techniques
• GPU acceleration: CUDA/JAX implementations for large-scale applications
• Quarterly dependency review: Regular updates and security checks

This roadmap reflects that we're already at the end of Q2 2025 and focuses on achievable goals for the remainder of 2025 and early 2026.