AI Financial Advice: Supply, Demand, and Lifecycle Implications

Battle Count: 50

The paper focuses on household-level financial advice (spending, saving, investing) rather than quantitative trading strategies, market microstructure, or algorithmic execution. While it discusses portfolio allocation (equity shares, diversified vs individual stocks, crypto), the context is personal financial planning over a life cycle, not trading signals or market prediction. The life cycle model and SMM estimation techniques have some methodological overlap with quantitative finance, but the application domain is fundamentally different from quantitative trading. The findings about passive portfolio drift and rebalancing behavior could tangentially inform understanding of retail investor behavior relevant to market dynamics.

High complexity due to multiple interconnected components: (1) Survey design and administration via Prolific with demographic balancing; (2) Quantitative life cycle model with 7 state variables, 4 asset classes, stochastic income/employment/returns, calibrated to SIPP/SSA/CRSP data; (3) Prompt bucketing and variable insertion system with complex household scaling rules; (4) Two-step LLM pipeline (advice generation + quantitative translation) with deterministic extraction rules; (5) Full life cycle simulation iterating ages 22-90 with random prompt draws and shock realizations; (6) SMM estimation on a 201x121 grid; (7) Randomized label experiments for demand/supply decomposition; (8) Dictionary-based textual analysis with 27 categories. Requires expertise in computational economics, survey methodology, LLM APIs, and dynamic programming.

The paper uses closed-source models (GPT-5.2, GPT-5 Mini, Gemini 3 Flash, GPT-5.6 Terra) which limits full reproducibility. The survey instrument, prompt construction rules, translation prompt, academic prompt, and life cycle model calibration are extensively documented in the appendix. However, the exact API versions, reasoning effort settings, and stochasticity of LLM outputs make exact replication difficult. The authors acknowledge this trade-off, prioritizing ecological validity (using models households actually use) over reproducibility. The methodology framework itself is reproducible with any LLM.