Fix simulate_data.py reproducibility + remove synthetic CSVs (#794)

* fix: make all data generators reproducible via seed parameter

- Add `seed` parameter to all generator functions that lacked it
- Replace scipy.stats RNG calls (norm().rvs, uniform.rvs, dirichlet().rvs)
  with numpy Generator methods (rng.normal, rng.uniform, rng.dirichlet)
- Replace np.random.* global state calls with local rng instances
- Fix create_series() bug: hardcoded length_scale=2 now uses parameter
- Rename create_series, generate_seasonality, periodic_kernel to private
  (_create_series, _generate_seasonality, _periodic_kernel) since they
  are internal helpers that require an rng instance from their caller
- Fix deprecated pandas freq="M" → freq="ME"
- Remove unused scipy.stats imports (norm, uniform, dirichlet)
- Remove module-level global rng; keep RANDOM_SEED constant for use
  by load_data() and test fixtures
- Standardize generate_staggered_did_data to use same rng pattern

Closes #545

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor: generate synthetic datasets on-the-fly in load_data()

Split DATASETS dict into SYNTHETIC_DATASETS (generated via seeded
functions) and REAL_WORLD_DATASETS (loaded from CSV). This removes
the dependency on synthetic CSV files while keeping real-world data
as shipped CSVs.

- Replace _get_data_home() with simple _DATA_DIR = Path(__file__).parent
- Remove circular `import causalpy as cp` dependency
- All 8 synthetic datasets now call their generator with RANDOM_SEED

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test: add session-scoped fixtures for all synthetic datasets

Eight fixtures (did_data, its_data, its_simple_data, rd_data, sc_data,
anova1_data, geolift1_data, geolift_multi_cell_data) generate data once
per test session using RANDOM_SEED, avoiding redundant calls to
load_data() or generators in individual tests.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* test: replace cp.load_data() with fixtures for synthetic datasets

Update 8 test files to use session-scoped fixtures (did_data, its_data,
rd_data, sc_data, anova1_data, geolift1_data) instead of calling
cp.load_data() for synthetic datasets. Real-world dataset loading
(banks, brexit, drinking, risk, nhefs) remains unchanged.

Also rewrite test_data_loading.py to parametrize over all datasets
(synthetic + real-world) and add reproducibility + unknown-key tests.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore: delete synthetic CSVs and unused gt_social_media_data.csv

Remove 8 synthetic CSV files (~172K) now generated programmatically:
  did.csv, regression_discontinuity.csv, synthetic_control.csv,
  its.csv, its_simple.csv, ancova_generated.csv, geolift1.csv,
  geolift_multi_cell.csv

Also remove gt_social_media_data.csv which was never referenced
in the DATASETS dict or any code.

Real-world CSVs (12 files, ~3.2MB) remain in the repo.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* fix(datasets): use Callable type hint instead of callable for mypy

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(simulate_data): remove unused generate_time_series_data function

Superseded by generate_time_series_data_seasonal and
generate_time_series_data_simple. Not imported or called anywhere.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Benjamin T. Vincent <inferencelab@gmail.com>

Author: 3 people

causalpy/data/ancova_generated.csv

@@ -15,39 +15,62 @@
 Functions to load example datasets
 """
 
-import pathlib
+from collections.abc import Callable
+from pathlib import Path
 
 import pandas as pd
 
-import causalpy as cp
-
-DATASETS = {
-    "banks": {"filename": "banks.csv"},
-    "brexit": {"filename": "GDP_in_dollars_billions.csv"},
-    "covid": {"filename": "deaths_and_temps_england_wales.csv"},
-    "did": {"filename": "did.csv"},
-    "drinking": {"filename": "drinking.csv"},
-    "its": {"filename": "its.csv"},
-    "its simple": {"filename": "its_simple.csv"},
-    "rd": {"filename": "regression_discontinuity.csv"},
-    "sc": {"filename": "synthetic_control.csv"},
-    "anova1": {"filename": "ancova_generated.csv"},
-    "geolift1": {"filename": "geolift1.csv"},
-    "geolift_multi_cell": {"filename": "geolift_multi_cell.csv"},
-    "risk": {"filename": "AJR2001.csv"},
-    "nhefs": {"filename": "nhefs.csv"},
-    "schoolReturns": {"filename": "schoolingReturns.csv"},
-    "pisa18": {"filename": "PISA18sampleScale.csv"},
-    "nets": {"filename": "nets_df.csv"},
-    "lalonde": {"filename": "lalonde.csv"},
-    "zipcodes": {"filename": "zipcodes_data.csv"},
-    "nevo": {"filename": "data_nevo.csv"},
+from .simulate_data import (
+    RANDOM_SEED,
+    generate_ancova_data,
+    generate_did,
+    generate_geolift_data,
+    generate_multicell_geolift_data,
+    generate_regression_discontinuity_data,
+    generate_synthetic_control_data,
+    generate_time_series_data_seasonal,
+    generate_time_series_data_simple,
+)
+
+_DATA_DIR = Path(__file__).parent
+
+# Synthetic datasets are generated programmatically for reproducibility.
+# .reset_index() on ITS functions because generators set date as the index,
+# but the old CSV-based load_data returned date as a column.
+SYNTHETIC_DATASETS: dict[str, Callable[[], pd.DataFrame]] = {
+    "did": lambda: generate_did(seed=RANDOM_SEED),
+    "rd": lambda: generate_regression_discontinuity_data(
+        true_treatment_threshold=0.5, seed=RANDOM_SEED
+    ),
+    "sc": lambda: generate_synthetic_control_data(seed=RANDOM_SEED)[0],
+    "its": lambda: generate_time_series_data_seasonal(
+        treatment_time=pd.to_datetime("2017-01-01"), seed=RANDOM_SEED
+    ).reset_index(),
+    "its simple": lambda: generate_time_series_data_simple(
+        treatment_time=pd.to_datetime("2015-01-01"), seed=RANDOM_SEED
+    ).reset_index(),
+    "anova1": lambda: generate_ancova_data(seed=RANDOM_SEED),
+    "geolift1": lambda: generate_geolift_data(seed=RANDOM_SEED).reset_index(),
+    "geolift_multi_cell": lambda: generate_multicell_geolift_data(
+        seed=RANDOM_SEED
+    ).reset_index(),
 }
 
-
-def _get_data_home() -> pathlib.Path:
-    """Return the path of the data directory"""
-    return pathlib.Path(cp.__file__).parents[1] / "causalpy" / "data"
+# Real-world datasets remain as CSV files shipped with the package.
+REAL_WORLD_DATASETS: dict[str, str] = {
+    "banks": "banks.csv",
+    "brexit": "GDP_in_dollars_billions.csv",
+    "covid": "deaths_and_temps_england_wales.csv",
+    "drinking": "drinking.csv",
+    "risk": "AJR2001.csv",
+    "nhefs": "nhefs.csv",
+    "schoolReturns": "schoolingReturns.csv",
+    "pisa18": "PISA18sampleScale.csv",
+    "nets": "nets_df.csv",
+    "lalonde": "lalonde.csv",
+    "zipcodes": "zipcodes_data.csv",
+    "nevo": "data_nevo.csv",
+}
 
 
 def load_data(dataset: str) -> pd.DataFrame:
@@ -84,6 +107,7 @@ def load_data(dataset: str) -> pd.DataFrame:
         - ``"zipcodes"`` - Geo-experimentation zipcode data for comparative interrupted time
           series analysis. Based on synthetic data from Juan Orduz's blog post on
           `time-based regression for geo-experiments <https://juanitorduz.github.io/time_based_regression_pymc/>`_.
+        - ``"nevo"`` - Berry, Levinsohn, and Pakes (1995) cereal data for BLP estimation
 
     Returns
     -------
@@ -106,11 +130,9 @@ def load_data(dataset: str) -> pd.DataFrame:
 
     >>> df = cp.load_data("rd")
     """
-
-    if dataset in DATASETS:
-        data_dir = _get_data_home()
-        datafile = DATASETS[dataset]
-        file_path = data_dir / datafile["filename"]
-        return pd.read_csv(file_path)
+    if dataset in SYNTHETIC_DATASETS:
+        return SYNTHETIC_DATASETS[dataset]()
+    elif dataset in REAL_WORLD_DATASETS:
+        return pd.read_csv(_DATA_DIR / REAL_WORLD_DATASETS[dataset])
     else:
-        raise ValueError(f"Dataset {dataset} not found!")
+        raise ValueError(f"Dataset {dataset!r} not found!")