Consumer Confidence — when the longitudinal axis is information, not time

studies/consumer_confidence/ · reference: true · Path C (wrapped ConsumerSim pipeline)

Given a fixed population pool and the news available at a point in time, can you forecast that month's US consumer confidence — and how does the forecast update as information accrues within the same month? That second clause is what makes this case a pattern worth copying: the panel axis is not calendar time but the as_of information cutoff. Every step re-forecasts the same target month from a strictly larger, point-in-time-safe information set. It is the general trick for building an honest longitudinal story out of a simulator that is natively single-shot.

P / E / B

	what it is	where it comes from
P	12 synthetic stratified personas, us-demo-0001 … us-demo-0012, each a distinct cell of region × age group × income group × education group × location group. interaction: none, propagation independent — respondents do not talk to each other, which is right for a survey instrument.	population/population.json (materialized_count: 12)
E	Two layers only, and no scheduled events, no broadcasts — see below. Provider consumersim.env with region: US, three target_months (all 2026-06) and three as_of_dates.	environment/environment.json
B	Decision consumersim.decision: sampled core agents emit an Action of kind survey_response answering five UMich-style questions positive / neutral / negative. Non-core agents emit not_sampled, which the Bayesian expansion later rewrites to expanded_response.	adapter/decision.py, adapter/providers.py

layer	modality	scope	dynamics	content
consumer_information	information	macro	scheduled	point-in-time news + indicator snapshot, refreshed each step as as_of advances
consumer_demographics	physical	local	static	the fixed pool's demographics as ConsumerSim consumes them

An environment with no events is still a dynamic environment

scheduled_events and broadcasts are both empty. E still changes every step — because the information layer is rebuilt against a new as_of cutoff. Dynamism does not require an intervention.

The core/expansion split is the other half of the design. core_ratio: 0.5 means 6 of the 12 agents are predicted directly by the LLM each step; the other 6 are filled in by a hierarchical Bayesian small-area expansion from the core's group-level posterior. Over three steps that is 18 simulation-layer LLM calls for full coverage of the pool.

One step, and the horizon

One step is one as_of cutoff inside the same target month: the same 12 agents are re-surveyed against everything published on or before that date. n_steps: 3, seed: 42, interaction_rounds: 1.

step	as_of	target month
1	2026-06-13	2026-06
2	2026-06-20	2026-06
3	2026-07-01	2026-06

Step 0 is a null warm-up row (no as_of, no score) — the pool exists but nothing has been forecast yet.

Results

step	as_of	raw_score	corrected_score	news_count
0	—	—	—	0
1	2026-06-13	128.33	129.83	4
2	2026-06-20	138.33	139.83	5
3	2026-07-01	121.67	123.17	10

population_size stays 12 and core_size stays 6 throughout — the panel is genuinely fixed, only the information changes.

The forecast traces an inverted V that tracks the information environment rather than drifting with time: combined_score runs 0.0563 → 0.0900 → 0.0675 and the corrected score follows the same shape. Note that news volume rises monotonically (4 → 5 → 10) while net sentiment peaks at the middle cutoff and is then diluted by later, more mixed headlines. More news is not better news — which is exactly the reading a point-in-time-safe information set is designed to make visible.

Two honest caveats the report states itself

• indicator_count is 0 at every step: the bundled indicator observations are dated after the as_of cutoffs and were correctly filtered out by the point-in-time guard. combined_score is therefore carried by news sentiment alone.
• The absolute level (~120–140) is not calibrated. The grounded real-world readings for the same month are Conference Board CCI 91.2 and UMich ICS 44.8. Read this template for direction under accruing information and for the component structure, not for the point level.

What the wrapped engine can do at full scale is a different picture. This is ConsumerSim's own 2020–2026 historical backtest (from the external pipeline, not this 12-persona study): monthly CCI forecasts against ground truth and classical baselines, through COVID, the inflation shock, and the 2025 tariff shock:

Grounding

Five sourced facts, three declared assumptions, two implementation references.

entry	basis	role
Conference Board CCI, Jun 2026 = 91.2 (1985=100)	sourced	the level the study's own output is explicitly not calibrated to
its components: Present Situation 116.4, Expectations 74.4	sourced	the empirical "weak now / stronger later" split
UMich ICS Jun 2026 = 44.8 (FRED, via the SocioVerse Event service)	sourced	second index scale, for direction
UMich Jul 2026 = 54.4, all five components improved	sourced	the out-of-sample continuation to test against
macro backdrop: CPI +3.9% YoY, UNRATE 4.2%, gasoline $3.831/gal	sourced	the drivers behind the June news snapshot
the showcase core_ratio raised above the pipeline default	assumed	at the default, a 12-cell pool collapses to a single-person core
reading the longitudinal axis as as_of accrual	assumed	a defensible use of the data actually on hand, not invented future months
treating the output as a Conference-Board-style index scale	assumed	inferred from the pipeline's own correction anchor, not documented upstream

Implementation references: a Statistics Canada (Survey Methodology, 2002) hierarchical Bayes nonignorable-nonresponse model for multinomial small-area data — the basis for the core → expansion design; and the University of Michigan Surveys of Consumers methodology — the five-question skeleton the decision model answers.

That last assumption is worth reading closely: the scale mapping was inferred from data files, so it is declared rather than cited. See Grounding & Provenance for what sourced / proxy / assumed commit you to.

What you can fork

param	the question it opens
region	does the same information set move a different regional pool the same way?
target_months	add 2026-07 once July news exists and test the grounded rebound to 54.4.
as_of_dates	denser cutoffs turn the inverted V into a real intraday-style information curve.
news / indicator / history inputs	drop the favourable gas-price stories to get a counterfactual — how much of June's forecast was the gasoline narrative? Or feed indicators dated before the cutoff to make indicator_count non-zero.
core_ratio	trade LLM budget against expansion variance; at n=12 a single core respondent's flip swings the headline.
prediction provider	swap the model behind the survey answers and diff the trajectories.

The natural first fork is scale: a larger pool with a lower core_ratio smooths the headline while keeping the LLM budget flat. Reuse is a fork, never an in-place edit — see Iterate, Versions & Reports.