Explainable Multi-Agent Financial Advisory: Design and Comparative Evaluation of Collaborative Reasoning Models for Investment Recommendation
Generative-AI agents are increasingly proposed for investment analysis, yet two questions remain under-examined: how should multiple specialist agents combine their opinions, and what does the combination cost in explainability? This paper designs a financial multi-agent advisory framework in which four tool-grounded analyst agents—momentum, mean-reversion, risk, and macro-regime—emit bounded conviction scores, confidences, and structured evidence. On top of a single agent interface we formalize six collaborative reasoning models as coalition-evaluable operators: a best single agent, uniform voting, confidence-weighted fusion, iterative debate (DeGroot-style consensus formation), a hierarchical coordinator with risk-veto authority, and a learned ridge-regression fusion. Coali-tion evaluability makes exact Shapley attributions over agents computable in closed form (24 coalitions), yielding a unified explanation object—per-agent attributions, evidence rationale, disagreement-based uncertainty, and a natural-language brief—for every recommendation, regardless of the reasoning model. In a 4.6-year out-of-sample evaluation (2022–2026) on ten large-cap U.S. equities with transaction costs, confidence-weighted fusion and debate achieve Sharpe ratios of 1.03 and 1.02 versus 0.78 for a passive S&P 500 benchmark and 0.74 for the best single agent, while the hierarchical veto committee is over-conservative (0.52); block bootstrap shows this protocol ordering is robust (stable in 99.7% of resamples) even where absolute market outperformance is not statistically separable at this horizon. Explanation faithfulness degrades monotonically with reasoning interactivity (from 1.00 for additive committees to 0.78 for debate) even though exact Shapley attributions remain stable under input perturbation (≥0.995 cosine similarity), quantifying an explainability cost of collaboration that simple deletion-based explanations would misreport. All data, code, figures, and tables are reproducible from a released Python artifact.
Authors
- Inderdeep Kaur (ORCID: https://orcid.org/0009-0005-8345-0676)
- Aman Goyal
- Varun Gupta
- Sahil Rauniyar
Institutions
- Chandigarh University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11780690/v1
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- preprint