Ten business-analysis seeds

These small examples turn analytical requirements into checkable contracts. They use MatrixOne for facts and SQL transformations, then Python libraries for scientific operations where appropriate.

Install the basic dependencies, plus SciPy 1.17.0, scikit-learn 1.8.0, NetworkX 3.6.1 and Shapely 2.1.2. Run business_seeds.py with common.py:

python -m pip install -r requirements-scientific.txt
python business_seeds.py

The script creates seven small synthetic tables, compares selected source facts with explicit references, checks all ten results and removes its tables. It is a set of starter contracts, not the full complexity or data scale of the larger campaign.

Requirement

Executed seed and expected result

Extension needed for a full workflow

Strict point-in-time features

Visible entity totals 30 and 60; a later arrival is excluded.

Versioned dimensions, latest visible revisions, label-ready time and future-only invariance.

Incremental corrections

Replacing one contribution changes total 30 to 35; rollback restores every source value.

Durable delta ledger, delivery-key replay and full-rebuild comparison.

Event analysis

Two bounded touches form sessions 1 and 2 under a one-minute gap rule.

Continuous session merging, watermarks, late data and recovery.

Experiments and causal analysis

Synthetic randomized arms have a mean difference of 1; Welch statistic matches the explicit standard-error formula.

Assignment and estimand checks, CUPED/DiD/IPW assumptions, clustered errors and multiple comparisons.

Attribution and journeys

Last-touch channel B receives 100 exactly once.

Identity-time validity, lookback windows, Shapley/Markov assumptions and unconverted journeys.

Hierarchical forecasts

Last-known leaf forecasts 20 and 40 add to parent 60.

Rolling-origin backtests, seasonal baselines, validation-only intervals and cold-start controls.

Graph and fraud analysis

Visible static graph has the cycle {1,2,3}.

Increasing-time path constraints, cycle chronology, changing edges and fraud-label evaluation.

Risk and simulation

Historical linear-interpolated VaR95=9; mean of observed losses >= VaR is 10.

Explicit CVaR estimator, portfolio assumptions, Monte Carlo, stress scenarios and uncertainty.

Spatial trajectories

Strict interior count 1; boundary-inclusive count 2.

CRS, invalid geometry, geodesic distance, antimeridian and DST rules.

Model monitoring

After labels mature, tied-score AUC=0.875 and Brier=0.135; the earlier cutoff has only three labels.

Label-delay cohorts, PSI/KS, segment tests, calibration and threshold governance.

Interpret the outputs at their declared scope

The experiment fixture is synthetic and randomized; a difference in observational means alone is not a causal conclusion. Last-touch credit is a rule, not estimated incremental advertising impact.

The graph seed finds a static cycle within a time-visible snapshot. It does not establish that a path traversed edges in increasing event time. The spatial seed uses planar coordinates, not longitude/latitude distance.

The risk seed defines a finite-sample historical quantile and tail-average convention. Different quantile interpolation and fractional-tail conventions can produce a different CVaR result. The forecast seed demonstrates leakage-free input selection and coherent sums, not forecasting quality.

For predictive workflows, use strict time visibility and train-only preprocessing. For longer SQL graphs, retain the fact-grain checks at each join. For streaming claims, use the connector scope, rather than extrapolating from bounded SQL results.