What you will do
You will have reviewed historical mappings and a retained household population with declared scope, weights, and generation assumptions.
Before you start
- An owned project with working Python review CLIs and private S3-compatible storage.
- Permission to use the source data, with declared location, units, frequency, coverage, sensitivity, and retention.
- Whole household donor records without real personal identifiers, including the attributes required by your selected model.
Step / 01
Review the observations you will actually use
Open Historical data from your project. Upload UTF-8 CSV or enter a series manually, then declare the source, usage rights, geography, units, frequency, date coverage, missing codes, and retention. Original bytes are preserved and checksum-verified; the source declaration alone does not make a series a usable model input.
Choose the date and value columns. Mapping choices are explicit: identity, trim, or positive scaling with a rationale. The review shows original rows and diagnostics for invalid dates, missing values, duplicates, gaps, and coverage. It does not silently interpolate, deduplicate, convert currency, or extend a series beyond its coverage.
Resolve blocking diagnostics and save the reviewed mapping version. A blocked review may remain available for inspection, but it cannot make a world data-ready. Required base-date inputs need a unique usable observation with compatible units and model bounds.
Step / 02
Enter complete household records
Open Population and households. Declare the geographic boundary, represented population scope, household definition, currency, resource period, reference date, source rights, uncertainty, and limitations. These declarations must match the world in which the population will be used.
Enter donor households or import a JSON array. Each household supplies a weight, location, members, baseline resources, and consumption shares. Members include children and other dependents; their ages, economic status, and any supplied sector, education, or income remain part of the complete household record. Use null for unknown optional JSON fields, or leave the corresponding form fields blank.
The generator copies complete household records and gives them synthetic IDs. It does not infer occupations, income, product preferences, or relationships from a place name. A contextual dataset link records provenance context; it does not prove that your donor records were derived from that dataset.
Step / 03
Choose generation and review the fit
Weighted resampling uses your supplied donor weights and a seed to sample whole households. Optional marginal fitting adjusts donor weights to your declared category shares using the supplied households as joint support. It never invents an unobserved household combination to satisfy a target.
Set the scale
Declare the synthetic household sample size, represented household total, seed, maximum share error, and evidence-retention period.
Declare fitting assumptions
For marginal fitting, supply category shares, bounds, tolerance, and an iteration limit, and confirm the donor-support/joint-distribution assumption. Shares within each dimension must sum to one.
Inspect the review
Compare declared and realized shares, diagnostics, household membership, donor/member references, and expansion weights. Missing support, inconsistent targets, or unmet tolerance blocks saving.
Save a version
Save the exact reviewed inputs. Any substantive revision requires a new review and appends an immutable version instead of changing previous evidence.
Step / 04
Keep weights, ties, and retention in context
Synthetic counts and expanded represented totals are different quantities. Every member inherits its household's expansion weight. Good marginal fit and matching weighted totals do not establish empirical representativeness.
A household-tie graph is optional. If selected, it connects members within the supplied household under an explicit edge-weight assumption. Choosing no graph leaves those relationships unavailable rather than inventing a generic social network.
Uploading a dataset and saving a population creates project resources, not an executable bound scenario. Select the retained population and required reviewed mappings in world setup, then review the compatibility of geography, scope, currency, and dates. Private downloads provide the original specification, population artifact, and generation report. Expiry, erasure, or corrupt evidence makes dependent worlds and runs unavailable; metadata is not a replacement for missing source bytes.