Changelog

0.1.15 (2026-05-16)

New features

  • Power analysis — causaltensor.analysis.power_analysis: CLI and programmatic API for A/A null draws, empirical |τ| critical values, Monte Carlo power over a relative-effect grid; CSV + Matplotlib figures under analysis/results/power_analysis/<dataset>/.

  • Empirical power helpers — causaltensor.semi_synthetic.empirical_power for thresholds and power grids (used by the power-analysis pipeline).

  • Real-data report — run_real_data_report / save_report: one get_fit_result_from_method fit per estimator for both the summary table and optional counterfactual PNGs (Plotly static export); writes <root>/<dataset>/real_data_report_<dataset>.csv; CLI --plots and --plot-unit-row.

  • Semi-synthetic analysis script — optional --plots saves relative-error box plot PNGs alongside the detailed/aggregated CSVs.

Improvements & fixes

  • OLS synthetic control and robust synthetic control: staggered adoption (unit-specific treatment starts; donor pre-periods and ATT aligned with assignment geometry).

  • Load tests (causaltensor.analysis.load_tests) — Windows-oriented grid harness over synthetic (N, T): subprocess timing, rss_fit_peak_mb as incremental ΔRSS during get_tau_from_method_with_error (poll sampling), ATT relative error vs generate truth; optional --timeout (wall-clock kill) and --memory-mb enforced on that same ΔRSS definition (not summed child-tree RSS).

Documentation & packaging

  • Tutorial guide notebooks refreshed; README updates; optional kaleido dependency for Plotly PNG export.

  • Removed causaltensor.analysis.aa_tests in favor of the semi-synthetic A/A workflow.

0.1.14 (2026-05-15)

New features

  • Rich result objects. Every solver.fit() call now returns a fully self-contained result object.

    • result.O and result.Z are attached automatically, enabling all diagnostics without re-running the model.

    • New computed properties: residuals, effect_matrix, z_pattern, untreated_r2, control_rmse, pre_exposure_rmse, rmspe_ratio.

    • ``result.summary()`` prints a formatted table covering panel info, ATT estimate (with SE when available), fit diagnostics, and estimator-specific model internals (weights, rank, fixed effects, factor matrix shape). Returns self for method chaining.

    • ``result.plot_actual_vs_counterfactual(unit)`` renders an interactive Plotly chart — actual vs. counterfactual with green treatment-period shading, T0 marker, and a unit-level annotation box.

  • Estimator-specific result attributes (all new, accessed directly on the returned object):

    • DID / SDID / MC-NNM: row_fixed_effects, column_fixed_effects

    • SDID: unit_weights (donor simplex weights), time_weights

    • MC-NNM: M (low-rank component separate from fixed effects), beta

    • DC-PR: std / std_tau (sandwich SE), inference_method

    • OLS SC: beta (per-unit donor weight vectors), individual_te, control_units, treatment_units

    • CovPCA: U (left factor matrix, N × r)

  • Large recommendation / retail panels enabled. load_dataset now accepts two keyword arguments forwarded to the four large-panel loaders (retailrocket, dunnhumby, truus, movielens):

    • n_units (default 2500) — retain only the top-N items by event count.

    • time_freq ('W' / 'M' / 'D', default weekly) — aggregate raw daily data before pivoting, controlling panel width.

  • Tutorial Guide 04 (tutorials/guides/04_inspecting_results.ipynb) — end-to-end walkthrough of the result object API using synthetic data with a known ground-truth ATT.

Performance improvements

  • SDID: Replaced np.eye(N) @ w >= 0 / np.ones(N).T @ w == 1 constraint matrices with native CVXPY w >= 0 / cp.sum(w) == 1 forms; solver pinned to CLARABEL for faster, more reliable convergence.

  • OLS SC: Inner donor-weight optimisation replaced with a CLARABEL-based QP (cvxpy), dropping the sklearn dependency. The outer predictor-importance loop retains fmin_slsqp as it operates in a low-dimensional covariate space.

  • RSC: Per-treated-unit projection vectorised — pseudoinverse of the pre-period donor matrix is computed once and applied to all treated rows in a single matrix multiply instead of a Python loop.

Documentation

  • docs/api.rst extended with a “Result Objects” section: common-attribute table and per-estimator attribute reference with quick-access code examples.

  • fit() docstrings for SDID, MC-NNM, and OLS SC updated with complete Returns sections.

  • Cross-reference from Guide 01 to Guide 04 added.

0.1.13 (2026-05-12)

Breaking changes

  • Solver constructor API changed. All seven PanelSolver classes now require (O, Z) at construction time and expose a single fit() method. The old pattern of constructing a solver with no arguments and calling solve_with_cross_validation / solve_with_suggested_rank is removed. Migration:

    # Before
solver = DCPanelSolver()
res = solver.solve_with_suggested_rank(O, Z, suggest_r=3)

# After
res = DCPanelSolver(O, Z).fit(suggest_r=3)

  • ``MC_NNM`` cross-validation is now an argument to ``fit()``. Pass cross_validation=True (default) instead of calling the old solve_with_cross_validation entry-point.

  • ``causaltensor.matlib.util`` removed. All linear-algebra helpers (SVD, SVD_soft, transform_to_3D, etc.) have moved to causaltensor.utils.linalg. Update any direct imports:

    # Before
from causaltensor.matlib.util import SVD_soft

# After
from causaltensor.utils.linalg import SVD_soft

  • ``causaltensor.matlib.generation`` and ``causaltensor.matlib.generation_treatment_pattern`` removed. Use the replacement modules instead:

    • generate_low_rank_M, add_noise → causaltensor.synthetic.utils

    • Z_iid, Z_block, Z_stagger, Z_adaptive → causaltensor.utils.treatment_patterns

New features

  • Three tutorial notebooks: real observed panels, synthetic DGP study, semi-synthetic benchmarks.

  • Sphinx API documentation with full NumPy-style docstrings for all solvers, PanelDataset, generate, run_experiment, and run_aa_test.

  • causaltensor.utils.linalg — consolidated linear-algebra utilities with seeded rng= API.

0.1.12 (2025-03-12)

  • Added CVXPY package for SDID method

0.1.11 (2025-03-12)

  • Fix a bug in the DC method: suggest_r was ignored due to the priority of auto_rank and now it will be prioritized over auto_rank

0.1.10 (2025-02-08)

  • Added Covariate support for SDID method

0.1.9 (2025-02-07)

  • Added Panel Solver Interface

  • Added more test cases

  • Added covariate support for synthetic control

0.1.8 (2023-11-05)

  • Enhanced MC-NNM functionality with covariate integration and improved handling of missing data.

0.1.7 (2023-08-24)

  • Introduced support for synthetic control methodology.

0.1.5 (2023-05-16)

  • Expanded capabilities to address multiple-treatment problems using panel regression methods with debiasing features.