One Stop Analytics 1.13.25

Cross-platform analytics toolkit for Python >=3.10,<3.14 (Python 3.10–3.13 CI lanes are configured; Python 3.14 is not yet certified).

Highlights

• Pandas / NumPy based analytics
• Lazy imports for faster startup
• Automatic EDA and data-quality profiling
• CSV streaming/chunk processing for large files
• Parallel batch analysis with bounded workers
• Persistent file-aware cache with automatic invalidation
• SciPy statistical tests
• scikit-learn regression/classification helpers
• Matplotlib/Seaborn visualization
• OpenPyXL Excel reports
• Windows, macOS and Linux

Install

python -m pip install one-stop-analytics

Examples

from one_stop_analytics import Analytics

a = Analytics("sales.csv")
report = a.eda(cached=True)
a.clean()
a.export_excel("report.xlsx")

Large CSV

from one_stop_analytics.io import iter_data_chunks

for chunk in iter_data_chunks("huge.csv", chunksize=200_000):
    # process one chunk at a time
    print(len(chunk))

Parallel batch

from one_stop_analytics import batch_analyze

results = batch_analyze(["a.csv", "b.csv", "c.csv"], workers=4)

CLI

osa eda sales.csv
osa eda sales.csv --no-cache
osa batch data/*.csv --workers 4
osa cache-clear

Cache location can be overridden with OSA_CACHE_DIR. The default is a per-user cache directory; avoid shared cache directories. Cache keys include the package/cache schema version.

Output safety: XLSX exports neutralize spreadsheet-formula-looking text. CSV, JSON and Parquet preserve literal values; if CSV files will be opened in spreadsheet software, apply an application-specific safe-export policy before sharing. Writing legacy .xls is intentionally unsupported; use .xlsx.

License and commercial licensing

The public distribution is intended to be licensed under GNU Affero General Public License v3.0 only (AGPL-3.0-only); see LICENSE. The project may offer a separate commercial license for organizations that need different permissions, subject to a separate written agreement and confirmation that the licensor owns or is authorized to license all relevant code.

Commercial licensing contact: Srijan Das. Contact email: otanalytics.contact@gmail.com. The commercial license document in this repository is a draft for legal review only; it is not a customer-ready offer or contract. Third-party dependencies retain their own licenses. See LICENSING_AUDIT.md and THIRD_PARTY_LICENSE_AUDIT.md for the current, incomplete release audit.

Release blockers: confirm copyright ownership and dual-licensing authority for all contributions, finish the exact-version dependency audit, and obtain legal review of the commercial agreement before publication. The contact address above should be monitored before release.

End-to-end workflow

from one_stop_analytics import Analytics

result = (
    Analytics("sales.csv")
    .analyze(
        clean=True,
        include_statistics=True,
        include_ml=False,
        include_charts=True,
        output_excel="analytics.xlsx",
        output_html="analytics.html",
    )
)

For supervised ML:

result = Analytics("sales.csv").analyze(
    target="Sales",
    include_ml=True,
    output_excel="analytics.xlsx",
    output_html="analytics.html",
)

The high-level workflow covers import → cleaning → automated EDA → statistics → optional ML → visualization → Excel/HTML reporting.

Excel / Power BI-style transformations

from one_stop_analytics import Analytics

t = Analytics("sales.csv").transform()
summary = t.group("Region", {"Revenue": "sum", "Cost": "sum"})
filtered = t.filter(Region="West")

CLI:

osa analyze sales.csv --output analytics.xlsx --html analytics.html
osa ml sales.csv --target Sales --task regression

1.8.8 performance & visualization hardening

• Optimized EDA cardinality reuse to avoid repeated full-column scans.

• Reduced visualization overhead for common bar charts and bounded automatic chart generation.

• Fixed release metadata/cache namespace alignment for 1.8.8.

• Fixed tab-delimited CSV auto-detection.

• Bounded CSV dialect/header sampling without loading the whole source file.

• Large CSV eda(), profile(), correlations() and kpi() now use bounded-memory streaming paths instead of dereferencing a missing full DataFrame.

• Decision-report CLI preserves streaming for large CSV inputs.

• Cache serialization is pandas-2.x/3.x compatible and preserves common DataFrame/Series metadata. Non-finite cache scalars are normalized safely.

• Formula-injection sanitization is applied only to XLSX exports; CSV, JSON and Parquet preserve literal text values.

• CLI converts common runtime errors into clean non-zero errors instead of raw tracebacks.

1.8.8 reliability baseline

• Raw-vs-analyzed report scopes are explicit and consistent across output formats.

1.8.8 transaction analytics

• Returns/cancellations-aware gross, return and net revenue metrics.

v0.6.1 hardening

• Automatic identifier-column detection for ML (e.g. customerID, order_id, UUID-like columns).
• High-cardinality categorical protection in automatic chart selection.
• Automatic target/churn KPIs when a target is supplied to analyze().
• Numeric scaling in classification preprocessing for more stable Logistic Regression.
• ROC-AUC included when binary classification probabilities are available.

v0.7.1 Time-Series & Forecasting

Adds lightweight time-series detection, resampling, validation, baseline forecasting (naive, seasonal-naive, moving average, exponential smoothing), forecast metrics, future forecast output, and Analytics.time_series() / Analytics.forecast(). The workflow aggregates large raw event data before forecasting, making million-row retail datasets practical without requiring statsmodels.

Time-series forecasting (v0.7.1)

Install optional forecasting models with:

pip install one-stop-analytics[forecasting]

The forecasting layer supports automatic date detection, resampling, leakage-safe missing-period handling, naive/seasonal-naive baselines, moving average, Holt, Holt-Winters, and metrics including MAE, RMSE, MAPE, RMSLE, sMAPE and MASE.

from one_stop_analytics import Analytics

result = Analytics("sales.csv").forecast(
    date_column="date",
    target="sales",
    forecast_periods=30,
    method="auto"
)

1.8.8 Performance

• Vectorized EDA outlier statistics: shared quantile/IQR computation across numeric columns instead of repeated per-column scans.
• Target-aware categorical analysis converts the target once instead of repeating the same conversion for every categorical column.
• Business cohort indexing uses native pandas period ordinals instead of Python-level apply().
• Existing pandas/NumPy/scikit-learn optimized implementations remain the computational backend; the package avoids reimplementing their core algorithms.
• Release artifacts exclude build caches, test caches, generated reports and previous wheels.

v0.8.0 Advanced Analytics

• Advanced statistics: Spearman/Kendall, Mann-Whitney, Wilcoxon, Kruskal-Wallis, effect sizes, bootstrap confidence intervals, OLS diagnostics and VIF.
• Advanced ML: Extra Trees, GradientBoosting, SVM and robust model wrappers.
• AutoML: cross-validated model comparison and automatic winner selection.
• Business analytics: RFM, cohort, retention, funnel, CLV and ABC/Pareto.
• Integrated through Analytics.auto_ml(), Analytics.advanced_statistics(), Analytics.business_analytics() and analyze() flags.

Decision-ready reports

Analytics.analyze() now adds a schema-aware decision_report containing Executive Summary, KPI, Data Quality, Trends, Top/Bottom Products, Customer Segments, Anomalies, Forecast, Key Findings, and Business Recommendations. The report is deterministic and only emits sections when the dataset contains sufficient evidence; unavailable analyses are marked unavailable rather than inferred.

Decision-ready multi-format reports

Generate one analysis result and export it to HTML, PDF, XLSX, and JSON without re-running the analytics engine:

osa decision-report data.csv --output-dir reports

Or select formats:

osa decision-report data.csv --output-dir reports --formats html,pdf,xlsx,json

Outputs:

• HTML — self-contained client-facing report with embedded charts.
• PDF — client-ready static report.
• XLSX — detailed analyst workbook with report sections as sheets.
• JSON — machine-readable report payload for APIs/automation.

The report layer includes Executive Summary, KPI, Data Quality, Trends, Top/Bottom Products, Customer Segments, Anomalies, Forecast availability/results, Key Findings, and Business Recommendations. It is schema-aware and does not fabricate unavailable metrics.

1.9.0 reliability and release improvements

• Batch scheduling bounds queued futures to reduce memory overhead for very large input lists.
• Batch inputs reject duplicate paths and invalid worker/timeout/retry settings early.
• Cache override warnings are emitted once per process rather than on every cache access.
• Holt/Holt-Winters optimization avoids brute-force starts and uses a finite iteration budget to reduce worst-case forecast latency.
• Release tooling runs test, compile, wheel build/install, import, CLI and archive-integrity checks; the verification JSON records each executed command.
• Cache schema version is incremented to invalidate incompatible cached values.

1.9.2 reliability and release improvements

• Batch scheduling bounds queued futures to reduce memory overhead for very large input lists.
• Batch inputs reject duplicate paths and invalid worker/timeout/retry settings early.
• Cache override warnings are emitted once per process rather than on every cache access.
• Holt/Holt-Winters optimization avoids brute-force starts and uses a finite iteration budget to reduce worst-case forecast latency.
• Release tooling runs test, compile, wheel build/install, import, CLI and archive-integrity checks; the verification JSON records each executed command.
• Cache schema version is incremented to invalidate incompatible cached values.

1.9.3 output accuracy and presentation hardening

• Missingness KPIs are recalculated from canonical raw counts and source dimensions when available.
• JSON reports normalize missing/non-finite values to null and serialize dates and chart metadata consistently.
• HTML tables handle source columns that collide with index labels.

1.9.3 reliability and release improvements

• Batch scheduling bounds queued futures to reduce memory overhead for very large input lists.
• Batch inputs reject duplicate paths and invalid worker/timeout/retry settings early.
• Cache override warnings are emitted once per process rather than on every cache access.
• Holt/Holt-Winters optimization avoids brute-force starts and uses a finite iteration budget to reduce worst-case forecast latency.
• Release tooling runs test, compile, wheel build/install, import, CLI and archive-integrity checks; the verification JSON records each executed command.
• Cache schema version is incremented to invalidate incompatible cached values.

1.9.4 streaming report correctness

• Large CSV decision reports compute exact weekly and monthly sums in bounded-memory chunks.
• Aggregate trend charts are built from period summaries, not raw rows.
• Partial calendar periods are flagged and excluded from headline trend comparisons.
• Streaming duplicate counts stay unknown unless explicitly computed.
• HTML, JSON, PDF and XLSX share canonical KPI/trend data; XLSX includes a KPI Summary sheet.

1.9.4 reliability and release improvements

• Batch scheduling bounds queued futures to reduce memory overhead for very large input lists.
• Batch inputs reject duplicate paths and invalid worker/timeout/retry settings early.
• Cache override warnings are emitted once per process rather than on every cache access.
• Holt/Holt-Winters optimization avoids brute-force starts and uses a finite iteration budget to reduce worst-case forecast latency.
• Release tooling runs test, compile, wheel build/install, import, CLI and archive-integrity checks; the verification JSON records each executed command.
• Cache schema version is incremented to invalidate incompatible cached values.

1.9.6 forecast validation engine

• Chronological development validation and a separately reserved final test window.
• Expanding rolling-origin folds for temporal robustness checks.
• Seasonal-naive baseline always included in model selection and final-test comparisons.
• Forecast metrics include MAE, RMSE, WAPE, RMSLE and MASE (plus MAPE/sMAPE for continuity).
• Metric notes flag zero-actual limitations: WAPE is undefined when total absolute actuals are zero; MAPE excludes zero-actual rows. RMSLE clips negative values and flags that interpretation caveat.
• Hyperparameter/model selection uses development validation only; final test scores are reported separately and are not used for selection.

Reporting and connector validation (v1.11.0)

from one_stop_analytics.report import html_report, decision_report_xlsx, decision_report_json, export_dashboard
from one_stop_analytics.validation import validate_report_consistency, validate_charts, audit_recommendations
from one_stop_analytics.connectors import extract_sql, extract_crm_pages

# After constructing a canonical `report_result`:
html_report(report_result, "report.html")
decision_report_xlsx(report_result, "report.xlsx")
decision_report_json(report_result, "report.json")
print(validate_report_consistency(report_result, "report.html", "report.xlsx", "report.json"))
export_dashboard(report_result, "dashboard.html")  # offline snapshot

SQL extraction yields bounded chunks and permits only a single SELECT statement. Install one-stop-analytics[connectors] to use SQLAlchemy engines. extract_crm_pages is a generic paginated JSON REST adapter; check your CRM vendor's API contract and scopes before use. Recommendation audit flags missing evidence or limitations instead of manufacturing them.

1.11.0 reliability and release improvements

• Batch scheduling bounds queued futures to reduce memory overhead for very large input lists.
• Batch inputs reject duplicate paths and invalid worker/timeout/retry settings early.
• Cache override warnings are emitted once per process rather than on every cache access.
• Holt/Holt-Winters optimization avoids brute-force starts and uses a finite iteration budget to reduce worst-case forecast latency.
• Release tooling runs test, compile, wheel build/install, import, CLI and archive-integrity checks; the verification JSON records each executed command.
• Cache schema version is incremented to invalidate incompatible cached values.

v1.12.0 — Unified Python analytics workspace

Workspace provides a single entry point for common spreadsheet, SQL, pandas/NumPy, visualization and report workflows. It is a Python-first toolkit, not a pixel-for-pixel replacement for the Excel, Tableau or Power BI desktop applications.

from one_stop_analytics import Workspace

ws = Workspace("sales.xlsx", name="sales")       # CSV, TSV, Excel, JSON, Parquet, Feather
print(ws.profile())                               # quality + descriptive statistics
by_region = ws.groupby("Region", {"Revenue": "sum", "Cost": "sum"})
wide = ws.pivot(index="Region", columns="Month", values="Revenue", aggfunc="sum")
query = ws.sql("SELECT Region, SUM(Revenue) AS revenue FROM sales GROUP BY Region")
ws.chart(x="Region", y="Revenue", kind="bar", aggregation="sum", output="revenue.png")
ws.export("report.xlsx", include_profile=True)
ws.export("report.html")
ws.export("report.pdf")
ws.export("data.json", include_profile=True)
ws.dashboard("dashboard.html", charts=[{"x": "Region", "y": "Revenue", "kind": "bar"}])

For richer analytical SQL, install pip install one-stop-analytics[duckdb] and call ws.duckdb("SELECT Region, SUM(Revenue) FROM sales GROUP BY Region"). For a browser-based interactive application, install pip install one-stop-analytics[dashboard]; use Streamlit/Plotly to build a front end around the same Python workspace and data contracts. Excel workbook formula recalculation, Tableau Server publishing, Power BI semantic-model/DAX compatibility, and a full spreadsheet cell editor are not provided by this library.

Unified BI workspace (v1.13.0)

from one_stop_analytics import Workspace

ws = Workspace("sales.csv", name="sales")
# Interactive browser dashboard (filters, KPI cards, chart drill-down)
ws.interactive_dashboard("dashboard.html", max_rows=50_000)
# Excel-compatible multi-sheet report with optional pivot summaries
ws.export_excel_report("sales_report.xlsx", pivots={
    "Regional Revenue": {"group": "Region", "metric": "Revenue", "agg": "sum"}
})
# Same canonical KPI contract feeds every export format
ws.export_report_bundle("reports", formats=("json", "html", "pdf", "xlsx"), pivots={
    "Regional Revenue": {"group": "Region", "metric": "Revenue", "agg": "sum"}
})

# Out-of-core query on CSV/Parquet files (source data is not loaded into pandas)
engine = Workspace.large_data_engine()  # pip install "one-stop-analytics[large-data]"
result = engine.query_csv("large_sales.csv", "SELECT Region, SUM(Revenue) AS revenue FROM source GROUP BY Region")
profile = engine.profile_arrow("large_sales.csv")  # requires PyArrow

# Database connectors (drivers depend on backend)
with Workspace.database("sqlite:///analytics.db") as db:
    result = db.query("SELECT Region, SUM(Revenue) AS revenue FROM sales GROUP BY Region")
# PostgreSQL: pip install "one-stop-analytics[postgres]"
# MySQL:      pip install "one-stop-analytics[mysql]"

The standalone dashboard embeds a bounded row sample to remain portable; it is not a server-side BI service. For larger-than-RAM files, run DuckDB SQL against source files first, then dashboard the bounded result. PostgreSQL/MySQL support depends on the matching SQLAlchemy driver being installed. SQL query helpers accept read-only SELECT/CTE statements; write_table is a separate explicit database write API.

Optional dependency matrix and database integration (v1.13.2)

Install only the feature group needed:

Extra	Installs / enables	CI validation
forecasting	statsmodels forecasting	Clean install + import smoke
advanced	statsmodels advanced analytics	Clean install + import smoke
connectors	SQLAlchemy database APIs	Clean install + import smoke
duckdb	DuckDB SQL over local files	Clean install + import smoke
large-data	DuckDB + PyArrow batch scanning	Clean install + import smoke
postgres	SQLAlchemy + psycopg 3	Live PostgreSQL service integration
mysql	SQLAlchemy + PyMySQL	Live MySQL service integration
visual	Plotly	Clean install + import smoke
dashboard	Streamlit + Plotly	Clean install + import smoke
all	All optional feature dependencies	Clean install + import smoke

Example installs:

python -m pip install "one-stop-analytics[postgres]"
python -m pip install "one-stop-analytics[mysql]"
python -m pip install "one-stop-analytics[large-data]"
python -m pip install "one-stop-analytics[all]"

The repository CI has separate Optional dependency matrix, Clean install, and Database integration workflows. PostgreSQL/MySQL integration tests require disposable test databases and run in CI service containers; locally they skip unless OSA_TEST_DATABASE_URL is set. Never point these integration tests at a production database: they create and drop uniquely named test tables.

For maintainers, run python scripts/check_optional_dependencies.py postgres (replace postgres with an extra name), python scripts/clean_install_smoke.py after building a wheel into dist/, and python -m pytest -q tests/test_database_integrations.py with the test database environment variables configured.

1.13.2 optional dependency and database integration

• Adds a CI matrix for each documented optional dependency extra, with import checks in clean runner environments.
• Adds fresh-virtual-environment wheel installation smoke tests across Linux, Windows, macOS, and supported Python lanes.
• Adds PostgreSQL and MySQL service-container integration tests for write/read round trips, parameterized queries, and read-only query enforcement.
• Uses SQLAlchemy dialect-aware identifier quoting for portable table reads, including MySQL without ANSI_QUOTES.
• Expands the all extra to include the optional Streamlit dashboard dependency.

1.13.2 optional dependency and database integration

• Adds a CI matrix for each documented optional dependency extra, with import checks in clean runner environments.
• Adds fresh-virtual-environment wheel installation smoke tests across Linux, Windows, macOS, and supported Python lanes.
• Adds PostgreSQL and MySQL service-container integration tests for write/read round trips, parameterized queries, and read-only query enforcement.
• Uses SQLAlchemy dialect-aware identifier quoting for portable table reads, including MySQL without ANSI_QUOTES.
• Expands the all extra to include the optional Streamlit dashboard dependency.

1.13.2 optional dependency and database integration

• Adds a CI matrix for each documented optional dependency extra, with import checks in clean runner environments.
• Adds fresh-virtual-environment wheel installation smoke tests across Linux, Windows, macOS, and supported Python lanes.
• Adds PostgreSQL and MySQL service-container integration tests for write/read round trips, parameterized queries, and read-only query enforcement.
• Uses SQLAlchemy dialect-aware identifier quoting for portable table reads, including MySQL without ANSI_QUOTES.
• Expands the all extra to include the optional Streamlit dashboard dependency.

Credit-note reconciliation (diagnostic only)

from one_stop_analytics import reconcile_credit_note_lines

result = reconcile_credit_note_lines("retail_export.csv")
print(result["report"]["ledger_readiness"])
result["detail"].to_csv("credit_note_line_audit.csv", index=False)

The credit-note matcher is a provisional diagnostic, not an accounting posting engine. It uses customer, SKU, unit price and chronology to identify candidate prior-sale quantities. It does not prove the original invoice relationship. Statutory-ledger use remains blocked unless source-system references, stable line IDs, tax/currency fields, transaction type, duplicate policy and independent finance controls are supplied and validated.

Data quality and validation engine (v1.13.12)

Validation is non-mutating by default. Automatic cleaning is opt-in and each requested change is recorded in the result.

from one_stop_analytics import DataQualityEngine

schema = {
    "required_columns": ["order_id", "amount", "status"],
    "unique": ["order_id"],
    "business_keys": [["order_id", "line_id"]],
    "columns": {
        "amount": {"type": "number", "nullable": False, "min": 0},
        "status": {"type": "string", "allowed": ["open", "closed"]},
        "order_date": {"type": "date", "format": "%Y-%m-%d"},
    },
}
engine = DataQualityEngine(schema, row_id_column="order_id")
result = engine.validate_file("orders.csv")  # or engine.validate(dataframe)
print(result["summary"]["counts"])
print(result["exceptions"])
from one_stop_analytics import write_exception_report
write_exception_report(result, "exceptions.xlsx")

# Explicit cleaning only; no action occurs unless clean=True and configured.
cleaned = engine.validate(dataframe, clean=True, cleaning={
    "strip_text": ["status"],
    "fill_missing": {"status": "unknown"},
    "drop_exact_duplicates": True,
})

Schema keys: required_columns, columns (per-column type, nullable, min, max, allowed/domain, format), categorical_domains, unique, business_keys, and row_id_column. Supported declared types include any, string, category, boolean, integer, number, and date/datetime. Numeric/date parsing during validation does not mutate the input. Failed numeric/date parsing is reported as malformed; schema violations as invalid; null/blank cells as missing; exact duplicate rows and repeated business keys as duplicate; and IQR outliers or padded text as suspicious warnings.

The result includes summary, exceptions, cleaning, data, and provenance. Row-count definitions are stage snapshots: raw_input_rows = parsed_rows + rejected_rows; filtered_rows counts rows excluded by an explicit caller-provided filter; cleaned_rows is the output row count after only configured cleaning; and analyzed_rows is the supplied analyzed count or cleaned row count. A single row may have multiple exception findings. File parse failures (including empty or malformed CSV files) raise DataQualityError with the original cause rather than pretending to have row-level counts for an unparsed file.

When a result is passed to the reporting layer under data_quality_validation, canonical reports retain its deterministic summary without embedding the potentially large exception table. Export exception details separately from result["exceptions"] when needed.

Business rules and KPI calculation engine (v1.13.14)

See BUSINESS_RULES_ENGINE.md for configurable, versioned KPI definitions, Decimal-based financial calculations, transaction-evidence policies, explicit unavailable states, source lineage, and invoice/customer/product reconciliation. Public APIs include BusinessRulesConfig, BusinessRulesEngine, evaluate_business_rules, and save_business_rules. The engine is opt-in and does not change existing Analytics.analyze() behavior.

Large CSV workloads

See PERFORMANCE_OPTIMIZATION.md for measured profiling results, exact/approximate streaming guarantees, cache behavior, and reproducible 1M/5M/10M-row benchmark commands. The optional chunked group aggregation API is aggregate_csv().

Evidence-based Automated Business Summary (v1.13.17)

Canonical report exports now include a deterministic business_summary with validated KPI highlights, data coverage, limitations, rule-backed alerts, descriptive changes, and recommendations. It uses no LLM and is emitted consistently to HTML, PDF, XLSX, and JSON.

from one_stop_analytics import BusinessSummaryConfig, generate_business_summary

summary = generate_business_summary(canonical_report, BusinessSummaryConfig(
    profile_name="generic",
    max_findings=5,
    metric_thresholds={"conversion_rate": {"operator": "lt", "threshold": 0.03}},
))

See BUSINESS_SUMMARY_GENERATOR.md for finding evidence, suppression rules, configurable profiles, exports, and optional AI-claim validation.

Lightweight local Reporting Studio (v1.13.18)

Run the built-in UI with osa-web (or python -m one_stop_analytics.webapp). By default it binds to 127.0.0.1:8765. Upload a CSV or Excel file, review the sample, configure cleaning, run the existing Python engine, inspect the canonical report, filter permitted records, and download HTML/PDF/XLSX/JSON.

See WEB_UI.md for installation, API contracts, limits, cleanup behavior, and security assumptions. The UI is static HTML/CSS/JavaScript; all calculations remain in the Python package.

Explainable anomaly detection and business alerts (v1.13.19)

The package exposes deterministic AnomalyConfig, detect_anomalies(data, config), and attach_anomaly_detection(canonical_report, data, config) APIs. It distinguishes data-quality violations, configured business-rule violations, numeric/categorical statistical anomalies, and optional time-series deviations. Each alert carries the observed value, baseline/expected range, deviation, method, threshold, reason, source fields, and stable record identifier.

See ANOMALY_DETECTION.md and examples/anomaly_detection.py for configuration, severity policy, evaluation examples, and limitations. Detection is descriptive; it does not infer fraud, errors, losses, or causes.

Forecasting validation hardening (v1.13.20)

Forecasting now uses disjoint chronological train/validation/test blocks and expanding rolling-origin folds contained within the training block. Model selection uses rolling-origin RMSE; validation reports selected-model and baseline metrics without selecting the model; the final test is reserved for post-selection scoring. Seasonal-naive is the baseline only when at least two training seasonal cycles are available; otherwise the baseline is naive. Metric notes explicitly mark undefined WAPE, RMSLE, and MASE cases rather than applying silent clipping or arbitrary denominators. Non-converged statsmodels folds are excluded from model scoring; a failed final refit falls back to the baseline with an explicit reason. Missing-period handling is configurable with missing_period_policy="zero" (backward-compatible, assumes absent periods mean zero activity) or "ffill" (causal forward fill, leading gaps set to zero and counted). See FORECASTING_VALIDATION.md for split boundaries, metric definitions, uncertainty interpretation, limitations, and reproducible test commands.

Security and privacy hardening (v1.13.21)

Commercial-distribution hardening details, dependency constraints, SBOM scope, deployment assumptions, residual risks, and reproducible security checks are documented in SECURITY_HARDENING.md. Run python -m pytest -q tests/test_security_privacy_hardening.py for malicious-input, cache, export, and local-web regression tests.

Unified end-to-end reporting pipeline (v1.13.25)

This release adds independent expected-value regressions for cleaning, EDA, KPI calculations, refund/cancellation accounting, invalid/empty inputs, CLI errors, and HTML/JSON/XLSX/PDF canonical-value consistency. Development installs include Hypothesis-based property tests. CI builds and tests wheel and source distributions in clean environments. Release readiness is not synonymous with a successful local build: inspect RELEASE_CHECKLIST.md, RELEASE_VERIFICATION.json, dependency audit results, and the CI matrix before distribution.

Reproducible local commands:

python -m pip install -e ".[dev]"
python -m pytest -q tests/test_release_regression_v11322.py
python -m pytest -q tests/test_security_privacy_hardening.py
python -m compileall -q src
python -m build
python -m twine check dist/*

Unified end-to-end reporting pipeline (v1.13.25)

Use the orchestrator to pass one validated DataFrame through existing One-Stop-Analytics components without duplicating KPI or export calculations:

from one_stop_analytics import run_pipeline

result = run_pipeline(
    "sales.csv",
    {
        "schema": {"required_columns": ["Date", "Sales"]},
        "cleaning": {
            "duplicates": True,
            "missing": "median",
            "standardize_text": True,
        },
        "business_rules": {
            "rule_version": "retail-v1",
            "columns": {
                "invoice_id": "Invoice",
                "quantity": "Quantity",
                "unit_price": "UnitPrice",
                "status": "Status",
            },
        },
        "trend_analysis": {
            "date_column": "Date",
            "revenue_column": "Sales",
        },
        "include_charts": True,
        "exports": ["html", "json", "xlsx", "pdf"],
        "output_dir": "reports",
        "report_stem": "sales",
    },
)

if not result["complete"]:
    raise RuntimeError(result["error"]["message"])

print(result["status"])
print(result["report"]["kpis"])
print(result["exports"])

No cleaning is performed unless the cleaning configuration is supplied. Row-level validation issues are recorded; set fail_on_validation_errors=True for strict runs. Missing required columns fail by default when a schema is supplied. Any failed stage returns report=None. Requested exports are all-or-nothing for the current run.

See PIPELINE_INTEGRATION.md for stage contracts, configuration versioning, security assumptions, CLI and API usage, and reproducible checks.

Automated customer reporting service (local/manual MVP)

One-Stop-Analytics includes reusable retail_sales, ecommerce_orders, inventory, and generic profiles for a repeatable customer reporting workflow. It supports preflight header mapping, validation, opt-in cleaning, canonical reports, Excel/HTML/PDF/JSON delivery bundles, and a data-coverage manifest. See CUSTOMER_REPORTING_SERVICE.md.

osa customer-report sales.csv --profile retail_sales --validate-only
osa customer-report sales.csv --profile retail_sales --output-dir customer_reports --report-stem customer_delivery

Python API: validate_customer_input() and run_customer_report().