Uncovering Salience-Driven Dynamics in Consumer Confidence with Generative Social Simulation

ConsumerSim's core contribution isn't just better CCI prediction; it's the claim that heterogeneous agent simulation can recover causal structure from aggregate economic time series. Most forecasting work treats confidence as a black-box statistical target. This paper inverts that: it uses LLM-adjacent reasoning to construct individual decision models, then aggregates them to match macro outcomes. The question is whether this interpretability gain is real or whether the model is simply fitting noise with more parameters.

This connects directly to the June arXiv paper on situation perception, which argued that LLMs lack the ability to build internal simulations of possible worlds and reason about counterfactuals. ConsumerSim operationalizes exactly that capability: it constructs synthetic households, feeds them timestamped signals, and simulates how they update beliefs about economic conditions. Where the situation perception paper identified a missing primitive, ConsumerSim attempts to build it for a bounded domain (consumer confidence). The architecture assumes that if you can simulate individual reasoning correctly, aggregate behavior should follow. That's a testable hypothesis about whether world modeling is actually the bottleneck.

If ConsumerSim's performance holds when tested on out-of-sample policy shocks (e.g., a new stimulus announcement or rate hike not in the training window), that validates the causal reasoning claim. If it degrades significantly, the model is likely pattern-matching on historical correlations rather than learning decision logic. Within six months, check whether central banks or policy institutions cite this framework in their own forecasting workflows; adoption by actual forecasters would signal practical credibility beyond benchmark performance.