# Changelog

<!--

Changelog follow the https://keepachangelog.com/ standard (at least the headers)

This allow to:

* auto-parsing release notes during the automated releases from github-action:
  https://github.com/marketplace/actions/pypi-github-auto-release
* Have clickable headers in the rendered markdown

To release a new version (e.g. from `1.0.0` -> `2.0.0`):

* Create a new `# [2.0.0] - YYYY-MM-DD` header and add the current
  `[Unreleased]` notes.
* At the end of the file:
  * Define the new link url:
  `[2.0.0]: https://github.com/google/meridian/compare/v1.0.0...v2.0.0`
* Update the `[Unreleased]` url: `v1.0.0...HEAD` -> `v2.0.0...HEAD`

-->

## [Unreleased]

*   Allocate spend that is aggregated over the geo and time dimensions once,
    when the analysis data tensors are built, instead of imputing it separately
    in `Analyzer.get_aggregated_spend`. `summary_metrics` now supports spend
    provided with dimensions `(n_channels,)`.
*   Fix calibrated ROI priors on the JAX backend so they match TensorFlow.

## [2.1.0] - 2026-09-17

*   Add `get_selected_dates_str` to TimeCoordinates.
*   Update `mmm-proto-schema` dependency to >= 1.4.0.
*   Add a declarative vocabulary to `ModelSpec`, so that calibration windows,
    holdouts and population scaling can be expressed with channel names, geo
    names and dates rather than raw NumPy index arrays. New public types in
    `meridian.model.spec`: `DateRange`, `CalibrationSpec`,
    `ChannelCalibrationSpec`, `HoldoutSpec`, `GeoHoldoutSpec` and
    `RandomHoldoutSpec`. A `DateRange` is inclusive of both bounds, and each
    bound must name one of the input data's time coordinates exactly.
*   Deprecate the five array-valued `ModelSpec` attributes in favor of
    declarative equivalents: `roi_calibration_period` → `roi_calibration`,
    `rf_roi_calibration_period` → `rf_roi_calibration`, `holdout_id` →
    `holdout`, `control_population_scaling_id` → `population_scaled_controls`,
    and `non_media_population_scaling_id` →
    `population_scaled_non_media_channels`. The deprecated attributes keep
    working and still take precedence when both members of a pair are set.
    `non_media_baseline_values` additionally accepts a channel-name mapping.
*   Add compiled properties to `ModelContext` that resolve the declarative
    `ModelSpec` attributes against the `InputData` coordinates:
    `compiled_roi_calibration_period`, `compiled_rf_roi_calibration_period`,
    `compiled_holdout_id`, `compiled_control_population_scaling_id`,
    `compiled_non_media_population_scaling_id`,
    `compiled_non_media_baseline_values` and `resolved_random_holdout`. The
    model and analysis layers read these instead of the raw `ModelSpec`
    attributes.
*   Persist the declarative `ModelSpec` attributes through `save_meridian` and
    `load_meridian`. A `RandomHoldoutSpec` is drawn stratified by geo — each
    geo holds out the same number of time periods — and the drawn holdout is
    recorded alongside the spec when the model is saved, so reloading restores
    the holdout that was actually fitted rather than redrawing it.
*   Add `TimeCoordinates.period_ends`.
*   Replace direct TensorFlow calls in `meridian.model.calibration` with
    `meridian.backend` to avoid unintended GPU memory pre-allocation when using
    the JAX backend.
*   Add `batch_size` to `Meridian.sample_prior` and `PriorDistributionSampler`
    to reduce peak memory usage.
*   Add an `include_knots` argument to `ModelFit.plot_model_fit()`, which draws
    vertical reference lines at the time periods where the model's knots are
    located.

## [2.0.0] - 2026-09-02

*   **Breaking change**: Refactor `NotFittedModelError` to `common.errors.NotFittedModelError`.
*   **Breaking change**: Toggle default backend from TensorFlow to JAX.
*   Add prior calibration through incrementality experiments such as Meridian GeoX.
*   Add channel calibration recommendations.
*   **Breaking change**: Standardized time selection arguments
    (`selected_times`, `media_selected_times`) across `Analyzer` and
    `BudgetOptimizer` to strictly require date string coordinates
    (`Sequence[str]`), removing support for positional boolean masks and
    normalizing `DataTensors.time` to `tuple[str, ...]`.
*   Relax `jax` and `jaxlib` dependency pins to `>= 0.7.2, < 1.0.0` and add
    `jax[cuda12]` to `[and-cuda]` optional dependencies to support Python 3.13
    and prevent PJRT accelerator plugin version mismatches.
*   **Breaking change**: Rename `max_rhat` to `max_r_hat` in
    `ConvergenceCheckResult` and update R-hat summary metric constants
    (`AVG_R_HAT`, `MAX_R_HAT`, `PERCENT_BAD_R_HAT`, `ROW_IDX_BAD_R_HAT`,
    `COL_IDX_BAD_R_HAT`) to use the `_r_hat` convention.
*   **Breaking change**: Removed `Meridian.populate_cached_properties()`. Use
    `ModelContext.populate_cached_properties()` directly.
*   **Breaking change**: Change `EDASeverity` outcome statuses from
    `INFO`/`ATTENTION`/`ERROR` to `INFO`/`REVIEW`/`FAIL` for consistency across
    the library.
*   Update `BayesianPPPCheck` calculation to use the posterior predictive
    distribution with `sigma` rather than the posterior expected outcome.
*   Make 64-bit precision the default for the JAX backend. Users can opt out and
    use 32-bit precision by setting `MERIDIAN_ENABLE_JAX_X64=false`.
*   Attach the `use_kpi` attribute to the summary metrics dataset.
*   Add channel-level inspection methods (`get_media_scaling_factor`,
    `get_channel_parameters`, `get_channel_parameter_tensor`) and
    `ChannelParameters` dataclass to `ModelContext`.
*   Add public methods to `Analyzer`: `incremental_outcome_xr` for calculating
    incremental outcomes as labeled `xarray.DataArray` objects,
    `get_incremental_kpi`, `yield_batched_distribution_tensors`, and
    `get_kpi_means`.
*   Add `WeeklyOptimizationGrid` to pre-compute weekly incremental outcomes over
    spend multiplier grids for budget optimization.
*   Add `currency_code` as an optional ISO 4217 model specification attribute
    on `InputData` and `InputDataBuilder`, with automatic symbol resolution for
    reporting and optimization visualizations.
*   Fix single-element `knots` deserialization ambiguity by introducing
    `knots_spec` (`n_knots` and `knot_locations`) in the `Hyperparameters` proto
    and SerDe, deprecating `knots` with backwards-compatible rehydration
    support.

## [1.8.0] - 2026-08-14

*   Implement `reconstruction_batch_size` to chunk posterior reconstruction evaluations, fixing peak resource exhausted memory crashes on wide modeling pipelines.
*   Update `tensorflow`, `tf-keras`, and `tensorflow[and-cuda]` dependencies to `>= 2.21.0, < 2.22` to address CVE-2026-2492.
*   Fix `AttributeError: 'Dataset' object has no attribute 'roi_m'` when
    running `ModelReviewer` on models configured with non-ROI priors
    (`media_prior_type='coefficient'` or `'contribution'`).
*   Fix AKS fallback behavior to correctly generate a single common intercept for non-national models when no internal knots are selected.
*   Add backward-compatible deserialization support in serde (`legacy_aks_v1_7_1.py`) for models saved with Automatic Knot Selection (`enable_aks=True`) under v1.7.1.
*   Add validation for prior distribution dtypes to match 64-bit JAX backend precision.
*   Fix spurious `UserWarning: Paid media prior type is unspecified` during
    model deserialization by suppressing warnings in `ProtoEnumConverter` when
    `default_when_unspecified` is `None`.
*   Fix OLS residual variance calculation (use `mse_resid` rather than `mse_resid ** 2`) in Automatic Knot Selection.
*   Fix docstring for `get_round_factor` to accurately describe budget rounding by powers of ten ($10^n$).

## [1.7.1] - 2026-07-20

*   Enable `MeridianEDA._generate_prior_specifications_card` and
    `MeridianEDA.plot_prior_mean` for national models.
*   Fix `ValueError: Interval length between selected times must be consistent`
    when serializing models with calendar-monthly or quarterly time coordinates.
*   Fix a silent attribution misalignment bug in `InputData` and `InputDataBuilder` by enforcing strict exact coordinate match ordering across channels.
*   Fix backward-compatibility deserialization error for older legacy models
    saved with Automatic Knot Selection (`enable_aks=True`).

## [1.7.0] - 2026-06-17

*   Fix serialization error in serde when saving models with
    `IndependentMultivariateDistribution` priors.
*   Move DataTensors and DistributionTensors out of analyzer.py into tensors.py.
*   Pin `matplotlib` dependency to `< 3.11.0` to fix upstream `arviz` import
    error.
*   Fix R&F optimization converters to use optimized channels.
*   Fix a JAX compatibility issue in `Summarizer` when calling `.idxmax()` on
    data backed by JAX.
*   Add support for posterior downsampling for faster, near approximate
    inference.

## [1.6.2] - 2026-05-15

*   Filter linear channels from Hill curves and refactor R-hat in visualizer.
*   Fix TypeError by wrapping KPI outcome in backend.to_tensor() in
    BayesianPPPCheck.
*   Expose `inverse_outcome` as a public method in `Analyzer`.

## [1.6.1] - 2026-04-30

*   Update `mmm-proto-schema` dependency to >= 1.2.1.

## [1.6.0] - 2026-04-29

*   Add MeridianEDA for EDA visualizations and two-pager generation.
*   Add support for user-configured EDA Specs.
*   Fix `MediaTransformer` median calculation when tensor equality is disabled
    in TensorFlow.
*   Add EDA check for data-to-parameter ratio (DATA_ADEQUACY).
*   JAX support is now available.
*   Add JAX 64-bit precision opt-in configuration.
*   Ensure consistent float precision across tensors, NumPy arrays, and prior
    distributions.
*   Deprecate model.NotFittedModelError and migrate it to common/errors.py.
*   Deprecate importing `model` module from `analyzer` and `review` modules.

## [1.5.3] - 2026-03-04

*   Pin `arviz` dependency to `< 0.20.0` to fix upstream `InferenceData`
    deprecation.
*   Dropped support for Python 3.10. The minimum required Python version is now
    3.11.
*   Upgraded JAX dependencies to version `0.7.2`.
*   Refactor model fit plot to include interactive tooltips and hover effects.
*   Fix a bug in tagging in `mmm_ui_proto_generator.py`.
*   Update MarketingData serialization to support float population values.
*   Add Model Health Summary Card.

## [1.5.2] - 2026-02-10

*   Move `schema` package as `meridian.schema`.
    *   Introduce a temporary shim `schema.py` module for backwards
        compatibility. This shim will be removed in the next cut release.

## [1.5.1] - 2026-02-04

*   Fix serialization for binary and text files.

## [1.5.0] - 2026-01-27

*   Remove dependency on the unmaintained `patsy` library.
*   Add interactive zooming to prior-posterior distribution plots.
*   Fix plots exceeding the width of the HTML 2-pagers.
*   Raise exception when paid media channels have zero total spend.
*   Include ArviZ version in model serialization.
*   Add more non-negative checks for model input data.
*   Add EDA check for treatment/control geo and time collinearity.
*   Refactored `model.Meridian` with stateful `ModelContext` and stateless
    `ModelEquations` helper classes.
*   Refactored samplers classes for direct injection of `ModelContext` and
    `ModelEquations` classes.

## [1.4.0] - 2025-12-08

*   Introduce modules needed for Meridian Scenario Planner and add
    `scenarioplanner` extra.

## [1.3.2] - 2025-11-26

*   Fixed an out-of-bounds bug in EDA's VIF check.
*   Added cost per media unit checks to EDA.
*   Add support for holdout set in `GoodnessOfFitCheck`.
*   Add more helpful error message for AKS min/max knot selection
*   Add support for python 3.13 and tensorflow 2.20.

## [1.3.1] - 2025-11-12

*   Fix `schema` dependency issues.

## [1.3.0] - 2025-11-10

*   Add `EDAEngine` for exploratory data analysis.
*   Add model fitting guardrail using EDA to `Meridian`.
*   Introduce serde package: a serialization and deserialization library for
    Meridian model with a protocol buffer schema.
*   Add model quality checks in the `analysis.review` module.
*   Add currency support to optimization summary and visualizer.
*   Expose new hyperparameters in AKS public api.
*   Refactor the TensorFlow RNG handler to use stateless seed generation.
*   Add `selected_geos` arg to the optimizer.
*   Add `selected_geos` arg to `get_aggregated_spend`.
*   Fix bug in `optimize()` when using `new_data` with `start_date` and
    `end_date` matching the first and last dates in the new data.
*   Move `use_kpi` to `Summarizer` and `Visualizer` class initialization.
*   Make KPI analysis the default when revenue data is unavailable.

## [1.2.1] - 2025-09-22

*   Add `use_kpi` arg to `output_model_results_summary`.
*   Add `lognormal_dist_from_mean_std` and `lognormal_dist_from_ci` helper
    functions.
*   Add support for forecasted data in the optimization 2-pager visualizations.
*   Change AKS algorithm to use AIC instead of EBIC.
*   Fix dtype issue when scaling integer kpi/population.

## [1.2.0] - 2025-09-04

*   Fix channel data misalignment in `Analyzer.hill_curves` when input channels
    are not in alphabetical order.
*   Add `negative_baseline_probability` method to `Analyzer` class.
*   Add per-channel adstock decay function definition.
*   Methods in the `analyzer` module now return backend-agnostic tensors.
*   Validate distribution support ranges for custom priors.
*   Add `IndependentMultivariateDistribution` for per-channel distribution
    definition.
*   Add automatic knot selection (AKS) to modeling.
*   Fix numerical stability of Adstock computation around `alpha = 1`.
*   Add `binomial` decay option to Adstock.
*   Make `trim_grids()` a public method of `OptimizationGrid` and update it to
    remove rows of NaNs.
*   Add organic RF support for adstock decay in analyzer.
*   Add organic RF support for Hill curves in analyzer.
*   Set the `random_seed` in `sample_prior()` to match the seed parameter.
*   Add organic RF support for `plot_hill_curves` in visualizer.
*   Raise a `ValueError` if any media channel have all zeros or all `NaN`
    impressions.
*   Add validation for constant KPI with contribution prior types.

## [1.1.7] - 2025-07-16

*   Fix `rhat_summary()` to work with a vector sigma dim.

## [1.1.6] - 2025-07-14

*   Convert stateful seeds into stateless seeds in `sample_posterior()` to
    ensure the sampling is deterministic.

## [1.1.5] - 2025-07-10

*   Remove `sigma_dims` pseudo-dimension from inference data.
*   Sets builder-wide default column names in `DataFrameInputDataBuilder`.

## [1.1.4] - 2025-06-30

*   `XrDatasetDataLoader` to use new `InputDataBuilder` API under the hood.
    These changes are backwards compatible.
*   Fix mishandling of an empty `controls` column list in the data loader
    params.
*   Maintain user-given channel ordering in `InputData`'s channels' coordinates.

## [1.1.3] - 2025-06-23

*   Add MLflow autologging support.
*   Fix bug where channels were being mapped to the wrong column name.
*   Add the ability to set `max_frequency` to `optimal_freq()` and update the
    `new_data` argument to take in `rf_impressions` rather than `reach` and
    `frequency` separately.
*   Add a helper method to create new data for optimizations with just spend or
    impressions and CPM. This includes geo allocation based on population.

## [1.1.2] - 2025-06-11

*   Add new `InputDataBuilder` APIs.
*   `DataFrameDataLoader` to use new `DataFrameInputDataBuilder` API under the
    hood. These changes are backwards compatible.
*   Keep rounded spend as int64 in optimizer.

## [1.1.1] - 2025-05-28

*   Rename the directory of unit testing datasets from `sample` to
    `unit_testing_data` and add a README.
*   Make `controls` data optional in the model.
*   Add a time variation error message for national models.

## [1.1.0] - 2025-05-20

*   Add `media_prior_type`, `rf_prior_type`, `organic_media_prior_type`,
    `organic_rf_prior_type`, `non_media_prior_type` parameters to `ModelSpec`.
*   Add `'contribution'` prior type option.
*   Add a data simulation demo notebook and update/add simulated datasets.
*   Change `VEGALITE_FACET_EXTRA_LARGE_WIDTH` from 900 to 700.
*   Prevent negative media effect priors when using lognormal distribution upon
    model init.
*   Add an optional `optimization_grid` arg to the optimizer.
*   Fix `incremental_outcome` to accept unscaled
    `non_media_treatments_baseline`.
*   Add spend allocation per geo and time if per-channel spend is provided.
*   Validate no time variation for non-media treatments, organic media, and
    organic reach.
*   Add optimizer parameters `start_date` and `end_date` to replace
    `selected_times`.

## [1.0.9] - 2025-04-17

*   Add support for optimization with forecasted data.
*   Deprecate `get_historical_spend` for `get_aggregated_spend` with `new_data`
    support.
*   Raise an error when `kpi_scaled` is all zero and `paid_media_prior_type` is
    anything other than "coefficient".
*   Prevent negative media effect priors when using lognormal distribution.
*   Add support for weekly time granularity to the contribution area and bump
    charts.
*   Update channel contribution over time charts in Summarizer to toggle weekly
    vs quarterly granularity based on the selected time period length.
*   Adjust the optimization summary to display dates in the same format as
    "Marketing Mix Modeling Report".
*   Increase width of model fit and channel contribution over time charts in the
    Visualizer.

## [1.0.8] - 2025-04-08

*   Update contribution calculation methods in `MediaSummary` with
    `aggregate_times` parameter to support granular time.
*   Add a `new_data` argument to `analyzer.optimal_freq()`.
*   Refactor args in `create_optimization_grid` to be consistent with
    `optimize(...)`.
*   Fix response curves for KPI-based optimization.
*   Add `plot_channel_contribution_area_chart` method to `MediaSummary` in the
    visualizer.
*   Add `plot_channel_contribution_bump_chart` method to `MediaSummary` in the
    visualizer.
*   Add organic media support for adstock decay in analyzer.
*   Add channel contribution area chart and channel contribution bump chart to
    model results summary report in the summarizer.
*   Add an extra check for zeros or negative values in `revenue_per_kpi`.
*   Add per-channel constraints parameters to `OptimizationGrid.optimize(...)`.
*   Add organic media support for hill curves in analyzer.
*   Add organic media support for `plot_hill_curves` in visualizer.

## [1.0.7] - 2025-03-19

*   Bump tensorflow to 2.18.
*   Bump tensorflow-probability to 0.25.
*   Bump numpy to 2.0.2.
*   Bump pandas to 2.2.2.
*   Bump scipy to 1.13.1.

## [1.0.6] - 2025-03-18

*   Fix issue #548: Make time coordinate regularity check less stringent.
*   Force `DataTensors` to have all tensors with `dtype=tf.float32`.
*   Refactor new data validation and data filling into the `DataTensors` class.
*   Fix bug in marginal ROI calculation in `summary_metrics` when new spend data
    is passed in.
*   Add `by_reach` param to `incremental_outcome()` to allow scaling `reach` or
    `frequency`.
*   Refactor `marginal_roi()` and the mROI calculation in `summary_metrics` to
    use the scaling factors in `incremental_outcome()`, removing duplicate code.
*   Allow `new_data` to have any number of time dimensions for `roi()`,
    `marginal_roi()`, `cpik()`, and `summary_metrics()`. This allows the use of
    forecasted data for analysis.
*   Add `optimize()` method to the `OptimizationGrid` dataclass.
*   Add a warning when the target constraint of flexible budget optimization is
    not met.

## [1.0.5] - 2025-03-06

*   Add technical support for python 3.10.
*   Align `NaNs` in `spend_grid` and `incremental_outcome_grid` in the
    optimizer.
*   Fix the stopping criteria of target total ROI in flexible budget
    optimization.
*   Separate creation of the grid data and the optimization.

## [1.0.4] - 2025-02-28

*   Fix validation on injected inference data when `unique_sigma_for_each_geo`
    is used in the model initialization.
*   Fix a divide-by-zero error in spend ratio calculation when historical spend
    is zero, preventing a `ValueError` in `output_optimization_summary`.
*   Add `non_media_baseline_values` argument to `MediaSummary` visualizations.
*   Refactor prior and posterior sampling logic into separate modules,
    simplifying `model` module.
*   Create a helper argument builder construct for API parameters that require
    an ordered list/array of values. See, e.g.,
    `InputData.get_paid_channels_argument_builder()`.

## [1.0.3] - 2025-02-07

*   Temporarily downgrade `tensorflow` version to 2.16 until the convergence
    issues on L4 and A100 GPU runtimes are resolved.

## [1.0.2] - 2025-02-06

*   Bump minimum `tensorflow` version to 2.18.
*   Add `[and-cuda]` optional dependencies to install `tensorflow` dependency
    with GPU support.
*   Add `non_media_baseline_values` argument to `summary_metrics`,
    `baseline_summary_metrics` and `expected_vs_actual` methods.
*   Update `compute_incremental_outcome_aggregate` docstring to match
    `incremental_outcome`.

## [1.0.1] - 2025-02-04

*   Bump minimum `pandas` version to 2.2.
*   Make `compute_incremental_outcome_aggregate` public.
*   Add `new_data` argument to `Analyzer.summary_metrics` method.
*   Add `use_kpi` argument to the `optimize()` method.

## [1.0.0] - 2025-01-24

*   Bump `tensorflow` version to 2.16 to support Python 3.12.

## [0.17.0] - 2025-01-23

*   Define constants for channel constraints in the optimizer.
*   Remove `aggregate_times` from `roi`, `marginal_roi`, and `cpik` methods in
    `Analyzer` and do not report these metrics in the `summary_metrics` method
    when `aggregate_times=False` as these metrics do not have a clear
    interpretation by time period.

## [0.16.0] - 2025-01-08

*   Organize tensor arguments of `roi`, `mroi`, and `cpik` methods of Analyzer
    into a `DataTensors` container.
*   Add warning message when user sets custom priors that will be ignored by the
    `paid_media_prior_type` argument.

## [0.15.0] - 2025-01-07

*   Convert `InputData` geo coordinates to strings upon initialization to avoid
    type mismatches with `GeoInfo` proto which expects strings.
*   Add `get_historical_spend` method to `Analyzer` class.
*   Split up `roi_*` and `mroi_*` parameters.

## [0.14.0] - 2024-12-17

*   Remove deprecated `use_roi_prior` attribute from `ModelSpec`.

## [0.13.0] - 2024-12-11

*   Add support for marginal ROI priors in Meridian.

## [0.12.0] - 2024-12-09

*   Rename `incremental_impact` to `incremental_outcome`.
*   Rename `plot_incremental_impact_delta` to `plot_incremental_outcome_delta`.

## [0.11.2] - 2024-11-27

*   Remove deprecated `all_channel_names` property from `Meridian` class.

## [0.11.1] - 2024-11-22

*   Remove unneeded argument `include_non_paid_channels` from
    `expected_outcome()`.
*   Fix a bug in the custom RF prior validation.

## [0.11.0] - 2024-11-19

*   Consistent naming for "rhat" methods.

## [0.10.0] - 2024-11-18

*   Add support for organic media, organic reach and frequency, and non-media
    treatment variables.
*   Rename `Analyzer.media_summary_metrics` method to `Analyzer.summary_metrics`
    with `include_non_paid_channels` argument.

## [0.9.0] - 2024-11-15

*   Organize arguments of `incremental_impact` and `expected_outcome` methods
    into a `DataTensors` container.

## [0.8.0] - 2024-11-05

*   Expand media summary metrics to return ROI, mROI, and CPIK in all scenarios
    with the addition of the `use_kpi` argument.
*   Optimal frequency now calculates the frequency that maximizes the mean ROI
    in all cases such that it is consistent when used in the budget optimization
    that optimizes revenue.
*   Fix an error in the data loader that occurs when the geo column is an
    integer.
*   Add a `_check_if_no_time_variation` method to Meridian to raise an error if
    a variable has no time variation.
*   Make the performance breakdown section of the model summary report display
    both ROI and CPIK charts for all scenarios.
*   Set default ROI priors for non-revenue, no revenue-per-KPI models.
*   Do not specify significant digits in the y-axis labels in plot_spend_delta,
    trim insignificant trailing zeros in all charts.
*   Rename `ControlsTransformer` to `CenteringAndScalingTransformer`.

## [0.7.0] - 2024-09-20

*   Make `get_r_hat` public.
*   Add `media_selected_times` parameter to `Analyzer.incremental_impact()`
    method. This allows, among other things, to project impact for future media
    values.
*   For `"All Channels"` media summary metrics: `effectiveness` and `mroi` data
    variables are now masked out (`math.nan`).
*   Introduce a `data.TimeCoordinates` construct.
*   Pin numpy dependency to ">= 1.26, < 2".
*   `InputData` now has `[media_]*time_coordinates` properties.
*   `InputData` now explicitly checks that time coordinate values are evenly
    spaced.

## [0.6.0] - 2024-08-20

*   Add `Analyzer.baseline_summary_metrics()` method.
*   Fix a bug where custom priors were sometimes not able to be detected.
*   Fix a bug in the controls transformer with mean and stddev computations.

## [0.5.0] - 2024-08-15

*   Include `pct_of_contribution` and `effectiveness` data to
    `OptimizationResults` datasets.
*   Add `Analyzer.get_aggregated_impressions()` method.
*   Add `spend_step_size` to `OptimizationResults.optimization_grid`.
*   Add `use_posterior` argument to the budget optimizer.
*   Rename `expected_impact` to `expected_outcome`.

## [0.4.0] - 2024-07-19

*   Refactor `BudgetOptimizer.optimize()` API: it now returns an
    `OptimizationResults` dataclass.

## [0.3.0] - 2024-07-19

*   Rename `tau_t` to `mu_t` throughout.

## [0.2.0] - 2024-07-16

## 0.1.0 - 2022-01-01

*   Initial release

<!-- mdlint off(LINK_UNUSED_ID) -->

[0.2.0]: https://github.com/google/meridian/releases/tag/v0.2.0
[0.3.0]: https://github.com/google/meridian/releases/tag/v0.3.0
[0.4.0]: https://github.com/google/meridian/releases/tag/v0.4.0
[0.5.0]: https://github.com/google/meridian/releases/tag/v0.5.0
[0.6.0]: https://github.com/google/meridian/releases/tag/v0.6.0
[0.7.0]: https://github.com/google/meridian/releases/tag/v0.7.0
[0.8.0]: https://github.com/google/meridian/releases/tag/v0.8.0
[0.9.0]: https://github.com/google/meridian/releases/tag/v0.9.0
[0.10.0]: https://github.com/google/meridian/releases/tag/v0.10.0
[0.11.0]: https://github.com/google/meridian/releases/tag/v0.11.0
[0.11.1]: https://github.com/google/meridian/releases/tag/v0.11.1
[0.11.2]: https://github.com/google/meridian/releases/tag/v0.11.2
[0.12.0]: https://github.com/google/meridian/releases/tag/v0.12.0
[0.13.0]: https://github.com/google/meridian/releases/tag/v0.13.0
[0.14.0]: https://github.com/google/meridian/releases/tag/v0.14.0
[0.15.0]: https://github.com/google/meridian/releases/tag/v0.15.0
[0.16.0]: https://github.com/google/meridian/releases/tag/v0.16.0
[0.17.0]: https://github.com/google/meridian/releases/tag/v0.17.0
[1.0.0]: https://github.com/google/meridian/releases/tag/v1.0.0
[1.0.1]: https://github.com/google/meridian/releases/tag/v1.0.1
[1.0.2]: https://github.com/google/meridian/releases/tag/v1.0.2
[1.0.3]: https://github.com/google/meridian/releases/tag/v1.0.3
[1.0.4]: https://github.com/google/meridian/releases/tag/v1.0.4
[1.0.5]: https://github.com/google/meridian/releases/tag/v1.0.5
[1.0.6]: https://github.com/google/meridian/releases/tag/v1.0.6
[1.0.7]: https://github.com/google/meridian/releases/tag/v1.0.7
[1.0.8]: https://github.com/google/meridian/releases/tag/v1.0.8
[1.0.9]: https://github.com/google/meridian/releases/tag/v1.0.9
[1.1.0]: https://github.com/google/meridian/releases/tag/v1.1.0
[1.1.1]: https://github.com/google/meridian/releases/tag/v1.1.1
[1.1.2]: https://github.com/google/meridian/releases/tag/v1.1.2
[1.1.3]: https://github.com/google/meridian/releases/tag/v1.1.3
[1.1.4]: https://github.com/google/meridian/releases/tag/v1.1.4
[1.1.5]: https://github.com/google/meridian/releases/tag/v1.1.5
[1.1.6]: https://github.com/google/meridian/releases/tag/v1.1.6
[1.1.7]: https://github.com/google/meridian/releases/tag/v1.1.7
[1.2.0]: https://github.com/google/meridian/releases/tag/v1.2.0
[1.2.1]: https://github.com/google/meridian/releases/tag/v1.2.1
[1.3.0]: https://github.com/google/meridian/releases/tag/v1.3.0
[1.3.1]: https://github.com/google/meridian/releases/tag/v1.3.1
[1.3.2]: https://github.com/google/meridian/releases/tag/v1.3.2
[1.4.0]: https://github.com/google/meridian/releases/tag/v1.4.0
[1.5.0]: https://github.com/google/meridian/releases/tag/v1.5.0
[1.5.1]: https://github.com/google/meridian/releases/tag/v1.5.1
[1.5.2]: https://github.com/google/meridian/releases/tag/v1.5.2
[1.5.3]: https://github.com/google/meridian/releases/tag/v1.5.3
[1.6.0]: https://github.com/google/meridian/releases/tag/v1.6.0
[1.6.1]: https://github.com/google/meridian/releases/tag/v1.6.1
[1.6.2]: https://github.com/google/meridian/releases/tag/v1.6.2
[1.7.0]: https://github.com/google/meridian/releases/tag/v1.7.0
[1.7.1]: https://github.com/google/meridian/releases/tag/v1.7.1
[1.8.0]: https://github.com/google/meridian/releases/tag/v1.8.0
[2.0.0]: https://github.com/google/meridian/releases/tag/v2.0.0
[2.1.0]: https://github.com/google/meridian/releases/tag/v2.1.0
[Unreleased]: https://github.com/google/meridian/compare/v2.1.0...HEAD