Scholarlink

<p>We present an end-to-end multi agent AI system for personal finance that integrates three components: (i) a multi-stage synthetic transaction data generator (Bayesian demographic priors, rule-based simulation, CTGAN refinement, and LLM-based validation), (ii) a budget analysis stack that combines a robust transaction classifier with a constrained optimization agent for actionable budgeting recommendations, and (iii) an LLM-driven investment advisor that fuses news sentiment, fundamental ratios, and technical signals to produce risk-aware suggestions. The novelty of this work lies in the system-level composition and several pragmatic modeling choices: explicit handling of class imbalance (via oversampling or class-weighted objectives), lightweight feature scaling/upweighting for rare categories, and a risk-profiling layer that steers the LLM’s generation toward transparent, auditable outputs. In addition, we standardize training and evaluation across modules and release all implementation details for reproducibility. On our internal synthetic-plus-realistic evaluation and a 20-stock daily backtest, the budgeting classifier matches strong baselines while the advisor’s price-movement model attains approximately 60% accuracy. We outline statistical testing protocols and report aggregate metrics with accompanying uncertainty where applicable. The complete code and evaluation scripts are archived to support long-term reproducibility (Zenodo DOI to be issued).</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An end-to-end multi agent AI system for personal finance: synthetic data generation, budget optimization, and investment advisory

  • 期刊论文

  • 发表日期：

    2026年1月16日

查看全文
  • Smit Snehal Pancholi,
  • Aman Jaglan,
  • Nemi Makadia,
  • Yash Doshi,
  • Amir Jafari

摘要

We present an end-to-end multi agent AI system for personal finance that integrates three components: (i) a multi-stage synthetic transaction data generator (Bayesian demographic priors, rule-based simulation, CTGAN refinement, and LLM-based validation), (ii) a budget analysis stack that combines a robust transaction classifier with a constrained optimization agent for actionable budgeting recommendations, and (iii) an LLM-driven investment advisor that fuses news sentiment, fundamental ratios, and technical signals to produce risk-aware suggestions. The novelty of this work lies in the system-level composition and several pragmatic modeling choices: explicit handling of class imbalance (via oversampling or class-weighted objectives), lightweight feature scaling/upweighting for rare categories, and a risk-profiling layer that steers the LLM’s generation toward transparent, auditable outputs. In addition, we standardize training and evaluation across modules and release all implementation details for reproducibility. On our internal synthetic-plus-realistic evaluation and a 20-stock daily backtest, the budgeting classifier matches strong baselines while the advisor’s price-movement model attains approximately 60% accuracy. We outline statistical testing protocols and report aggregate metrics with accompanying uncertainty where applicable. The complete code and evaluation scripts are archived to support long-term reproducibility (Zenodo DOI to be issued).