What you will do
You will prepare a compatible scenario, confirm its basket and parameter assumptions, and know which household effects the run measures.
Before you start
- An owned project with retained historical inputs and a reviewed input-driven household population.
- Compatible world scope, currencies, base date, model versions, and the resources and consumption shares needed by the budget model.
- Working API, queue worker, Python simulation CLI, PostgreSQL, Redis, and private object storage.
Step / 01
Prepare a compatible world
Choose Household and FX from the project workflow overview. Select the source world and retained population, then choose a reviewed baseline FX mapping or a separately confirmed assumption. Base-date observations, units, ownership, and evidence availability are checked before the scenario is admitted.
An FX rate names quote-currency units per one base-currency unit. For a hypothetical 10,000 UGX/USD intervention, the base currency is USD and the quote currency is UGX. This is a scenario input, not an estimated future rate.
Preparing the scenario creates a new world version and freezes the reviewed manifest. Existing worlds and historical snapshots are not silently upgraded. If a model or source changes, review the new configuration and its compatibility rather than reusing an old validation claim.
Step / 02
Declare both branches of the experiment
Choose the time-step frequency, horizon, paired replicas, seed, and checkpoint interval. Resource time bases must be comparable; calendar-daily allocation uses actual Gregorian days rather than assuming every month has 30 days.
The baseline is either an explicitly constant FX continuation or a complete supplied path, including the reviewed anchor and a rate for every tick. The intervention can be no shock or a declared step/linear ramp. Record its start, ramp, and good-specific lag. Pre-base lag history needs a separately confirmed hold-anchor assumption.
Within a world pair, both arms share their conditional random draws. Separate replicas describe model variability. Changing the seed does not represent omitted structural mechanisms or uncertainty in population representativeness.
Step / 03
Bind the basket, prices, and resources
Map every household consumption category to a named good. Declare the quantity unit, base price, import share, pass-through, and lag. Each required number is a confirmed assumption with rationale and uncertainty, or an exact reviewed source reference. Package size and currency are not converted silently.
The declared FX price channel multiplies the base price by one plus import share × pass-through × the proportional lagged FX change. Import share and pass-through are explicit model choices; a run does not establish that they are observed or empirically valid.
The budget model constructs a baseline bundle from supplied consumption shares and the declared expenditure share. Confirm proportional rationing when resources cannot buy the full bundle, and the period-flow accounting model. It does not invent savings carry-over, loans, or market-clearing behavior.
Step / 04
Enable only channels you can support
Optional labor/wage transitions require supplied sector and income attributes plus declared loss, reemployment, and income parameters. Foreign receipts require explicit household exposure and currency choices. Poverty needs a comparable real-resource line and declared equivalence weights. Unselected or unsupported effects remain unavailable.
Parameter uncertainty is separate from stochastic transitions. Fixed treatment uses declared points. Uniform sampling requires confirmation of its interpretation and correlation rule; interval bounds alone are not a probability distribution. Calibration-derived joint draws have their own analysis treatment.
Step / 05
Review, run, and inspect the evidence
Review the scenario and resolve blocking readiness diagnostics. Confirm the trace-retention choice, then run the exact reviewed configuration. The worker pins input/runtime identity and publishes results only after a complete verified attempt. Cancellation, failed uploads, and unavailable sources remain explicit states.
Use Aggregate outcomes for weighted baseline/intervention values, paired effects through time, and group coverage. Choose Individuals and events for synthetic household or person-level evidence, then select the evidence kind, arm, replica, and household/member filters. Inputs and evidence contains the source and replay context. Prices, ledgers, checkpoints, and causal event references answer different questions. Apply the selected evidence filters before reading a page, and use verified downloads for complete precision.
Exact replay checks the retained inputs, identities, calculations, and artifacts. A positive-effect frequency is the fraction of these simulated pairs, not a real-world forecast probability. Source expiry or erasure can invalidate dependent evidence even if an old result was once completed.