Skip to content

1.0.5

Latest

Choose a tag to compare

@winedarksea winedarksea released this 09 Nov 05:17
6c6ca10

Thanks to AI, you actually get proper release notes this time:

1.0.5 ๐Ÿ‡บ๐Ÿ‡ฆ ๐Ÿ‡บ๐Ÿ‡ฆ ๐Ÿ‡บ๐Ÿ‡ฆ

  • Saved-template compatibility: fixes and new defaults below can change forecasts when refitting templates from 1.0.4 or earlier.
  • Linear models: replaced inverse/Gram least-squares coefficient fits with direct solves and objective-preserving augmented least-squares fallback. BasicLinearModel now fits finite observations per series and computes normal observation intervals from the training fit, including short histories and trend damping. Cassandra skips redundant condition-number checks.
  • Feature detection and trend forecasting: Added joint detection of slope changes and level shifts; improved changepoint handling around missing data; added adaptive seasonality fitting, trend damping, slope shrinkage, and wider intervals for possible future trend changes.
  • TVA: Added forecast origin anchoring and a configurable internal fit_horizon for factor and torch-free modes. With fit_horizon=None, the internal horizon is now capped at max(28, T // 4), changing long forecasts in factor and 'none' modes. Expanded factor forecasting, covariance estimation, and reconciliation behavior.
  • Cassandra: Fixed ridge penalties to exempt the intercept instead of the first feature, changing explicit lambda and automatic regularization fits; x_scaler=True now fits an intercept. Revised level autoregression changes forecasts from 1.0.4 templates. Fixed autoregressive forecasting on detrended residuals. Fixed max_colinearity to use maximum absolute pairwise correlation instead of a triangular minimum that could never exceed a positive threshold. The filter now drops aliased Fourier columns when enabled (default 0.998; None disables it), changing refits of existing templates. Changed the separate max_multicolinearity eigenvalue filter default from 0.001 to None (off); 1.0.4 get_params omitted this key, so saved Cassandra templates without it now refit with the filter off. General template 45 explicitly retains 0.001.
  • Anomaly detection: Added calibrated Gaussian mixture anomaly scores and improved handling of univariate and multivariate anomaly outputs. (why do LLMs like calibration so much?) med_diff now scales by the median absolute difference rather than the median signed difference, changing scores. For nonparametric with output='univariate' across multiple input columns, corrected flattened-index mapping changes which rows are flagged; scores now report the fraction of anomalous columns (higher means more anomalous) instead of its complement, and return a one-column DataFrame instead of a Series.
  • Changepoints: CUSUM now uses a robust noise scale when normalized, changing threshold/drift semantics and detections (especially on trends); get_new_params still samples thresholds of 5โ€“30, so existing settings may need retuning. PELT and L0 now honor min_segment_length. L1 applies a spacing filter in the batch path used by aggregate_method='individual', but the mean/median path only collapses adjacent detections and does not enforce minimum spacing; segment_length controls the scoring window in noise mode; bottom_up treats max_changepoints as a cap instead of a floor; ED-PELT removed its ร—5 penalty calibration and can find fewer changepoints. L1/L0 trend extraction now uses offset ceil(order / 2) instead of order, reporting order 2+ changepoints floor(order / 2) steps earlier; adjacent detections collapse to their strongest position for all orders in the default 'relative' mode. EWMA now starts at lambda * x[0] instead of x[0], changing detections near the start of a series. half_yr_spacing now uses elapsed rather than calendar years, changing default spacing in BasicLinearModel and transformers.
  • Time-series utilities and transformations: fixed hourly Fourier features to include hours, changing hourly forecasts; FastICA now passes integer column labels (0..k-1) to subsequent transformers instead of series names. infer_frequency now returns None for fewer than two points; exactly two timestamps infer a frequency from their interval, preserving calendar month-start/month-end spacing for midnight pairs. Improved probabilistic outputs and short-history handling. Fixed the HistoricValues transformation key in the AlignLastValue + HistoricValues horizontal post-processor to match its parameters, changing its application. Halved FFTDecomposition's transformer search weight from 0.01 to 0.005.
  • PWA dependencies: Updated the Rust dependency lockfile.
  • LevelShiftMagic bug fixes that change outputs at the defaults: per-column selection (shift ids were shared across columns, a row-0 group stalled selection, one extra shift was kept), one flag per step (exact ties on integer data flagged both sides), detection windows need max(3, window_size // 3) points so series edges are no longer flagged and the last row can't leak an offset into forecasts, univariate mode combines per-series squared scores so opposite shifts no longer cancel and only individually significant series are corrected. Also: diff_overlap sizing (was ~0), remove_at_shift now sizes from raw data and fills only the removed points.
  • LevelShiftMagic: opt-in window_method="linear_gap" (detrended multiscale MOSUM with robust z, bump test spike rejection), plus scales (default [14, 30, 60]), deseasonalize, anomaly_window, tail_scan (default on, linear_gap only), alignment "median"/"linear_robust" and correction_mode="auto" (safe ratio for non-negative series). Its alpha is a robust z threshold (3-6), not the legacy std multiplier. alignment defaults per method (None: "average" legacy, "linear_robust" linear_gap); linear_gap accepts only "linear_robust", "median" or "gap" and raises otherwise. No slower than "overlap" at the defaults; benchmark in examples/level_shift_benchmark.py.
    • Also: tail_scan only fills free max_level_shifts slots; univariate shared scores average over series with a valid score at each date (missing series no longer dilute a shared shift).
  • TimeSeriesFeatureDetector (and so TVA): default level shift params changed to linear_gap (alpha 3.5, scales [14, 30, 60], linear_robust, tail_scan). The old default (window 364, alpha 1.8, rolling_diff) relied on the LevelShiftMagic bugs and gave ~2.1 validated false shifts per 2-year noise series once they were fixed; the new default measured recall 1.00 at 0.10-0.15.
  • TimeSeriesFeatureDetector: new prelim_level_shift (default 'auto' = on for linear_gap params) runs level shift detection before the final seasonality fit, which otherwise absorbs steps near year boundaries.
  • TimeSeriesFeatureDetector compatibility: older templates inherit trend_damping='auto', rather than undamped trends. New detector settings are not serialized into saved templates, so refits use their defaults. general_transformer_params={} now disables the transform instead of selecting the default. Explicit level_shift_validation relative_threshold and absolute_threshold values are now used as given; 1.0.4 capped them at per-series adaptive thresholds. Omitted or None thresholds retain adaptive behavior.
  • TimeSeriesFeatureDetector seasonality: adaptive_fourier now uses recurring rather than simple_3 date features. components['seasonality'] is now mean-zero, with its level/intercept moved into the residual for trend fitting, changing component attribution and downstream fits. seasonality_strength retains its existing formula but now uses seasonal fits that exclude a trend prior, with trend added back to the residual before scoring; values can fall on trending series. The score compares residual/seasonal variance with original-series variance and includes squared correlation with the original series, so it remains sensitive to trend share rather than measuring seasonality on detrended data.
  • Holiday attribution: corrected lunar calendar horizon/start-range dependence; added opt-in Easter rules (holiday_params={'use_easter_holidays': True}) and per-series suppression (global_holiday_anomaly_suppression=None) with annual holiday peers and an exceptional-outage escape hatch. True retains batch suppression; False disables it. Detector defaults (including those used by TVA) are unchanged: attribution improved, but forecast error and synthetic precision regressed in validation. See examples/holiday_attribution_report.md.
    • Fixed holiday name lookup keys dom_10_31 and dom_11_11 so Halloween and ArmisticeDay receive their intended names.
  • New get_new_params search options and their draw rates: BasicLinearModel linear_model='huber' (2%, not in fast; huber_iterations fixed at 50), ChangepointDetector l1_fused_lasso/l1_total_variation threshold_mode='noise' (20%) with threshold_multiplier, LevelShiftMagic window_method='linear_gap' (~13%), TVAModel origin_anchor (last_value 55%, deseasonalized 20%) and continuation_force='constant' (50%).
  • RadiusNeighbors regression: features are standardized and each query's radius is a multiple of its own nearest-neighbor distance (sampled 1.1-5, capped at 50 neighbors), so it is never empty and adapts to series of different scale. At 1.0.4 the absolute radius returned NaN for most predictions on load_daily/load_monthly (RollingRegression failed outright), and the RadiusNeighbors-specific pchip fills in Window/DatepartRegression are removed; the general non-finite guards in the recursive models stay. Saved templates change meaning: radius < 1 now means the single nearest row, and large radii (5, 10) reach the 50-neighbor cap.
  • DatepartRegression transformer search: corrected RadiusRegressor to RadiusNeighbors, enabling radius-specific parameter sampling for about 6.25% of draws in the default regressor pool (excluding fast/all/deep modes).
  • General templates 29 and 37: replaced sokalmichener with rogerstanimoto for SciPy compatibility (sokalmichener was deprecated in SciPy 1.15 and removed in 1.17; see the SciPy release notes).
  • TVAModel/TVA season_length: None (also older templates without the key) now infers the season for inner folds and seasonal-naive baselines instead of always using 7. The origin-anchor window stays 7. Safety-layer folds run at the full horizon when history allows and fall back to fit_horizon only when it does not.
  • sklearn models no longer install a process-wide warnings filter at import; the Liblinear convergence warning is silenced only inside LinearSVR fits.
  • MLEnsemble has a new 'pbc' method for estimation
  • MLflow: load_model (now also autots.mlflow_load_model) restores AutoTS runs fully. It brings back the logged constructor settings (newly saved as init_params in model_as_code.json) and the best_model template, so fit_data(df) + predict() refits only the winner; keyword overrides such as forecast_length= replace logged values. Previously only 7 settings came back (losing no_negatives, constraint, model_list, etc.) and best_model stayed empty, breaking export/per-series helpers. ModelObject payloads now record frequency and prediction_interval. Runs logged by older versions still load. Tests use a sqlite tracking store, since MLflow 3.18 rejects the file store by default.
  • MLflow template registry: mlflow_log_template registers a fitted model's top templates as a model version (optionally moving an alias such as rolling), and import_template/import_best_model/mlflow_load_model accept models:/name@alias, so production runs can keep templates in the registry instead of CSV files and roll back by moving the alias. Registered templates are pyfunc models whose predict refits the winner on given history (no future regressors). Autologged fits also record best score per generation, per-series error, horizontal/mosaic routing tables, ensemble model families, and seed drift; log_insights=False disables this. AutoTS.predict runs log the long-form forecast. See extended_tutorial.md.