Towards practical AI-enhanced portfolio construction: Integrating NLP models with goal programming under real-world constraints

Conventional portfolio optimization models mainly focus on the compromise between risk and potential gain and often overlook qualitative and forward-looking information about firms, even though such information contribute significantly to investment managerial decision process under uncertainty. This limitation reduces the ability of traditional models to capture insights related to companies’ performance, managerial expectations, and future prospects as reflected in financial disclosures. To address this gap, this study introduces a novel three-objective portfolio optimization model that extends beyond the conventional dual-focus framework centered solely on risk and return. In addition to these two traditional criteria, the proposed approach incorporates a third dimension: The evaluation of companies’ past performance and forward-looking perspective . To achieve this, we leverage advanced Natural Language Processing (NLP) techniques. Specifically, we use the Financial Bidirectional Encoder Representations from Transformers (FinBERT) structure, which is fine-tuned for Financial Statement Analysis (FSA), together with the Notebook LM tool to systematically extract and quantify insights derived from firms’ quarterly financial statements. These reports provide valuable qualitative and quantitative information about current operations, historical trends, and management’s expectations for the future. The sentiment scores derived from this process are then used to quantify each company’s performance and outlook and to construct the third objective function in the proposed model. To address this issue , we employ the Weighted Goal Programming (WGP) method, which offers a structured and flexible approach for handling multiple, potentially conflicting objectives. The findings show that portfolios incorporating the third criterion of firm outlook often achieve improved outcomes and better returns normalized for risk . The outcomes also indicate that the proposed model is capable of identifying companies that are not only financially sound but also strategically positioned for future growth. Overall, these findings highlight the value of integrating qualitative financial insights and NLP-driven sentiment analysis into quantitative portfolio optimization and provide a more realistic and adaptable structure for investment managerial decision process.

کلیدواژه ها:

portfolio optimization ، multi-objective ، Goal Programming ، Financial bidirectional encoder representations from transformers ، NotebookLM ، Analytic Hierarchy Process ، Dow Jones industrial average