Agentic AI-Powered Investment Portfolio Management

Solution Overview

In today’s capital markets, effective risk management is rapidly evolving as data becomes central to strategic decision-making. Portfolio managers must navigate massive volumes of diverse data—from structured financial feeds to unstructured content like news articles and social media—posing challenges that traditional systems aren’t built to handle.

AI-driven solutions are redefining portfolio management by moving beyond rule-based strategies toward adaptive, context-aware decision-making.

Our intelligent investment portfolio manager addresses the complexity of integrating and analyzing varied data sources. By combining real-time analysis with generative AI models, it offers an intuitive, end-to-end experience through an AI-powered Portfolio Assistant—transforming how portfolio managers explore insights and take action.

Powered by large language models (LLMs), retrieval-augmented generation (RAG), and MongoDB Atlas, including Vector Search, Time Series collections, and Atlas Charts—this solution empowers AI agents to:

• Analyze vast financial datasets from multiple sources, and process it into meaningful analytics—like macroeconomic correlations, moving averages, and portfolio-level metrics.

• Provide dynamic insights and recommendations based on portfolio performance.

This advanced approach enhances decision-making, improves risk management, and enables more accurate performance and impact analysis—helping portfolio managers make smarter, faster, and more informed investment decisions through “human in the loop” (HITL) implementation.

Reference Architecture

The solution is composed of three agents and two applications:

• Market analysis agent

• Market new agent

• Market assistant agent

• Data ingestion application

• Market insight generation application

Figure 1. High-level Architecture

As shown in the diagram, MongoDB serves as the backbone for storing the data ingestion application data. Market data (time-series), financial news (unstructured) and macroeconomic indicators (structured) data is extracted, transformed and loaded into MongoDB Atlas collections.

After the data ingestion is carried out, the first two agents: market analysis agent and market news agent are scheduled to analyze real-time market data and financial news, process news sentiment analysis, and power portfolio-specific insights based on current portfolio allocation. Finally, a third agent known as the market assistant agent, provides insights and recommendations for capital markets using a LangGraph ReAct agent, leveraging MongoDB Atlas for data storage, vector search, and agent state management.

Building the Solution

Let’s now understand how the solution was built, the code is split into three microservices:

Part 1. Loaders Service

Figure 2. Loaders Service Architecture

This part focuses on the data ingestion application, which handles the extraction of data from three sources:

• Yahoo Finance Market Data ETL: Extracts, transforms, and loads market data for various asset types using the yfinance Python package.

• FRED API Macroeconomic Data ETL: Extracts, transforms, and loads macroeconomic data using the pyfredapi Python package.

• Financial News Processing: Scrapes financial news, generates embeddings using voyage-finance-2 model from Voyage AI, and calculates sentiment scores using the FinBERT model from Hugging Face.

• Generating portfolio performance (emulation).

Afterwards, financial news articles are processed to calculate their sentiment score using FinBERT, a pre-trained LLM model. Additionally, using Voyage AI’s embedding model voyage-finance-2, news articles are vectorized and stored into the MongoDB collection together with their vector embeddings.

{
"_id": {
    "$oid": "67c09436bd1cbed4d0185880"
},
    "headline": "QQQ Leads Inflows as VGIT, HYG Jump: ETF Flows as of Feb. 27",
    "source": "etf.com ·  via Yahoo Finance",
    "posted": "1 year ago",
    "description": "Top 10 Creations (All ETFs) Ticker Name Net Flows ($, mm) AUM ($, mm) AUM % Change QQQ Invesco QQQ...",
    "link": "https://finance.yahoo.com/news/qqq-leads-inflows-vgit-hyg-005429051.html?fr=sycsrp_catchall",
    "synced": false,
    "extraction_timestamp_utc": {
        "$date": "2025-02-27T16:35:02.823Z"
},
    "ticker": "HYG",
    "article_string": "Headline: QQQ Leads Inflows as VGIT, HYG Jump: ETF Flows as of Feb. 27\n/n Description: Top 10 Creations (All ETFs) Ticker Name Net Flows ($, mm) AUM ($, mm) AUM % Change QQQ Invesco QQQ...\n/n Source: etf.com ·  via Yahoo Finance\n/n Ticker: HYG\n/n Link: https://finance.yahoo.com/news/qqq-leads-inflows-vgit-hyg-005429051.html?fr=sycsrp_catchall",
    "sentiment_score": {
        "positive": 0.03415600210428238,
        "negative": 0.04247178137302399,
        "neutral": 0.9233722686767578
},
    "article_embedding": [
        0.025981411337852478,
        0.03783617168664932,
        0.01893029175698757,
        0.019744139164686203,
        -0.009678225964307785,
        0.0296250581741333,
        -0.06560207903385162,
        -0.03349149599671364,
        0.005457616411149502,
        -0.004056802950799465,
        -0.07076755166053772,
        -0.04305347055196762,
        ...]
}

Finally, all collections are stored into MongoDB Atlas.

Part 2. Agents Service

Figure 3. Agents Service Architecture

Once data is ingested and stored into the appropriate MongoDB collections, two scheduled agents analyze it to generate portfolio-specific insights.

Market Analysis Agent

Evaluates macroeconomic indicators (via FRED), asset price trends (using 50-day moving averages), and market volatility (via VIX), combining these inputs with LLM-based reasoning (via AWS Bedrock/Anthropic) to produce tailored portfolio recommendations.

Market News Agent

Uses MongoDB Atlas Vector Search to retrieve semantically relevant financial news per asset, followed by LLM-based sentiment analysis and summarization to highlight key signals.

As shown in the architecture diagram, the agents are LLM-powered, tool-augmented workflows, orchestrated using LangGraph. Each follows a structured, deterministic process where tools are invoked in a fixed sequence, ensuring reliable and explainable execution.

Agent Tools

Each agent uses a tailored set of tools to retrieve, analyze, and synthesize insights:

• check_portfolio_allocation_tool: Retrieves the current asset allocation from MongoDB.

• calculate_asset_trends_tool: Compares current prices vs. 50-day moving averages to detect trends.

• assess_macro_indicators_tool: Fetches and interprets FRED economic data (e.g., GDP, rates, unemployment).

• assess_vix_tool: Analyzes market volatility using VIX data.

• generate_overall_diagnosis_tool: Synthesizes all inputs and delivers LLM-based recommendations.

• fetch_market_news_tool: Uses vector search to find semantically relevant financial news.

• generate_news_summaries_tool: Applies sentiment analysis and summarization with an LLM.

Each tool includes a natural-language docstring, which guides the agent in understanding when and how to use the tool by describing its purpose, expected input, and output behavior.

Semantic Vector Storage

Finally, all generated insights—market analysis and news sentiment reports—are embedded using the Voyage-Finance-2 model and stored in MongoDB Atlas, enabling semantic retrieval and historical traceability across both workflows.

Part 3. Market Assistant Agent Service

Figure 4. Market Assistant Agent Service Architecture

Once the data is ingested and embedded into MongoDB Atlas collections, the Market Assistant Agent uses the ReAct reasoning pattern, implemented through LangGraph, to autonomously interpret user queries, access relevant tools, and generate actionable financial insights.

This agent functions as a structured LLM-powered financial advisor, capable of step-by-step reasoning, action execution, and memory retention. Built with LangGraph’s ReAct agent framework, it follows a cycle of:

• Reasoning about the query

• Acting (e.g., calling a tool)

• Observing the result

• Iteratively continuing until a final response is reached.

Agent behavior is defined via a customizable agent profile stored in MongoDB, while long-term memory and state persistence are managed using the MongoDB checkpointer, which tracks interaction steps via checkpoints_aio and checkpoint_writes_aio.

Agent Tools Overview

The Market Assistant Agent relies on a modular toolset, divided into portfolio-specific analysis and general market intelligence:

Portfolio-Specific Tools

• market_analysis_reports_vector_search_tool: Retrieves semantically relevant market reports using Voyage-finance-2 embeddings, balancing vector similarity and recency to return both portfolio-wide and asset-specific insights.

• market_news_reports_vector_search_tool: Finds and summarizes recent news relevant to portfolio assets, including sentiment classification (positive/neutral/negative) and overall news impact.

• get_portfolio_allocation_tool: Displays the current investment distribution across portfolio assets, with tickers, descriptions, and allocation percentages.

• get_portfolio_ytd_return_tool: Calculates and formats year-to-date performance based on historical returns, stored in portfolio_performance.

• get_vix_closing_value_tool: Provides the most recent VIX closing price as a quick indicator of market volatility and risk.

General Financial Intelligence

• tavily_search_tool: Complements internal tools by retrieving up-to-date external financial data and news via web search, particularly useful for assets not currently in the portfolio.

Embeddings and Vector Search

The system uses the Voyage-finance-2 embedding model—optimized for financial text—to generate semantic representations for reports. These embeddings are stored in:

• reports_market_analysis (market reports)

• reports_market_news (news articles)

Both collections support semantic querying via MongoDB Atlas Vector Search, enabling high-relevance retrieval across market and news contexts.

Visit our GitHub repo: Leafy Bank UI and create this demo for yourself by following the instructions. (Note: Each part of the solution has its own repo, you’ll find them in the README as the “Capital Markets” services).

Why MongoDB

A Modern Data Foundation for AI-Driven Financial Applications

MongoDB is a unified, high-performance data platform built for the complexity of financial services. It replaces fragmented point solutions with a single system capable of handling structured, time series, unstructured, and vectorized data—making it ideal for powering intelligent, adaptive applications in capital markets, wealth management, and beyond.

Schema Flexibility and Document Model

MongoDB’s JSON-like BSON model aligns naturally with agent state and object-oriented design, simplifying data modeling for accounts, portfolios, transactions, and AI metadata. Developers can:

• Evolve data models without disruptive migrations thanks to dynamic schemas.

• Store related data together (e.g., user profiles and trade history) for faster, more efficient queries.

• Handle diverse data types—from primitives to arrays and embedded documents—in a unified structure.

Time Series Collections for Market Data

Purpose-built Time Series collections ingest and analyze large volumes of market data efficiently. Features include:

• High-throughput ingestion for equities, bonds, and commodities.

• Automatic compression, indexing, and data expiration for performance and cost control.

• Native support for windowed analytics like moving averages and volatility tracking.

Atlas Vector Search for Semantic Understanding

MongoDB Atlas Vector Search enables semantic understanding across financial content using vector embeddings. This solution leverages Voyage-Finance-2, a finance-specific model it powers:

• Context-aware retrieval of financial news using interactive prompts and similarity-based matching.

• Semantic similarity search enhanced with metadata filtering for more precise results.

MongoDB Checkpointer for Agent Memory

The Checkpointer pattern allows AI agents to persist their internal state across interactions, supporting:

• Context retention and long-term reasoning.

• Full visibility into past AI decisions, supporting easy audits and regulatory compliance.

Aggregation Pipelines and Atlas Charts for Insight Generation

MongoDB’s aggregation framework delivers real-time insights, from trend analysis to macroeconomic correlations. Combined with Atlas Charts, teams can build dynamic dashboards that surface:

• AI-driven portfolio signals.

• Market anomalies.

• Real-time performance indicators.

Together, these capabilities make MongoDB a robust, future-ready data layer for AI-driven financial solutions—delivering the flexibility, scalability, and intelligence modern finance demands.

Key Learnings

• AI powered solutions in investment portfolio management unlock a new level of insights that was not possible before. Portfolio managers are now equipped with the power of AI Agents that can analyze market data, market news, market events and propose insightful and actionable recommendations about portfolio assets and risks in real time.

• MongoDB is a key enabler in the Agentic AI journey for investment portfolio management. MongoDB’s flexible document model natively stores unstructured and dynamic market data, including news, events and social media data. MongoDB Vector Search retrieves contextual market similarities and correlations, while seamlessly integrating with embeddings and generative AI models.

• Agentic AI is a fast evolving paradigm in AI-driven decision-making. AI Agents continuously adapt, adjusting their actions and strategies based on real-time market conditions and market fluctuations. AI Agents learn from past experiences, refining their performance over time.

Technologies and Products Used

Authors

• Peyman Parsi

• Ainhoa Múgica

• Julian Boronat

• Andrea Alaman Calderon