We built a custom capital‑gains calculation engine and paired it with a retrieval‑augmented generation (RAG) vector store of market advisory reports to supply context for a multi‑agent trade recommendation system. The study used a 2x2 repeated‑measures design, toggling the tax‑optimization engine and the RAG module on or off, and measured recommendation quality by relative capital gains incurred during portfolio liquidation.
A repeated‑measures ANOVA revealed a significant main effect of the tax‑optimization engine ($F(1,29) = 9.17$, $p = .005$, $\\eta^2_p = .240$): enabling the engine reduced tax savings by roughly 55 percentage points compared to the no‑engine condition. The RAG main effect was not significant ($p = .841$), nor was the interaction ($p = .553$).
Descriptively, the RAG‑only condition achieved the highest mean tax savings (47.7%), with the baseline condition second (30.6%). This suggests that the pre‑trained language model’s internalized financial knowledge may be sufficient for competent tax‑loss harvesting recommendations without explicit tooling. The findings indicate that augmenting LLM agents with domain‑specific computation engines does not guarantee performance gains and may introduce conflicting optimization signals.
Blogger's Review: The work cautions against the assumption that more components automatically improve AI‑driven financial advice. Careful evaluation of each module’s contribution is essential, and leveraging the inherent knowledge of large language models can sometimes be more effective than brute‑force integration of external engines.