💻 computer science

A Multi-Agent Generative AI Framework for Explainable Stock Market Analysis in Indian Equity Market

This paper presents a five-agent generative AI framework that integrates technical, fundamental, and sentiment analysis with anomaly detection and LLM-based reasoning to provide explainable, confidence-aware stock market predictions for Indian equities, achieving 72% directional accuracy and 78% cross-agent consensus.

Original authors: R Ashok Kumar, Prashanth Gowda G D

Published 2026-07-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: R Ashok Kumar, Prashanth Gowda G D

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to decide whether to buy a specific stock in the Indian market. Usually, you'd have to act like a detective, a mathematician, a psychologist, and a historian all at once. You'd look at price charts (technical), read balance sheets (fundamental), scan the news for gossip (sentiment), and compare the company to its neighbors (sector benchmarking). Doing all this alone is exhausting, and often, the clues contradict each other.

This paper introduces a solution: a digital team of five expert robots (agents) that work together to give you a clear, honest answer about a stock, complete with a "why" that you can actually understand.

Here is how this "team" works, broken down into simple parts:

1. The Five Experts (The Agents)

Instead of one giant, confusing computer brain, the system uses five specialized agents, each with a specific job. They all work at the same time (like a relay race where everyone starts running together) to save time.

The Chart Reader (Technical Agent): This agent looks at the stock's price history. It checks if the stock is moving up or down, if it's "overheated" (too expensive too fast), or if it's a good time to buy based on momentum. It's like a surfer watching the waves to know when to jump in.
The Accountant (Fundamental Agent): This agent digs into the company's financial health. It checks if the company is making money, if it has too much debt, and if its price is fair compared to its earnings. It's like a home inspector checking the foundation and roof before you buy a house.
The Mood Reader (Sentiment Agent): This agent reads thousands of news articles and social media posts about the company. It uses a special AI (FinBERT) trained on financial language to figure out if the news is good, bad, or neutral. It knows that news from yesterday matters less than news from today, so it weighs recent stories more heavily.
The Neighbor Watcher (Sector Agent): This agent compares the target company to its competitors. Is this company doing better or worse than the rest of its industry? It's like checking if a student is the top of their class or just average.
The Judge (Synthesis Agent): This is the team leader. It takes the reports from the other four, weighs them based on how confident each expert is, and combines them into one final score.

2. The "Self-Healing" Mechanism

What happens if the Chart Reader says "Buy" but the Accountant says "Sell"? In many AI systems, this causes a crash or a confusing answer.

In this system, there is a Referee (Anomaly Detector). If the experts disagree too much, the Referee stops the process and says, "Wait a minute, something is wrong. Let's double-check the data." The system automatically re-runs the check with fresh data. It tries this twice. If the experts still disagree after two tries, the system flags the result for a human to look at, rather than guessing. This makes the system "self-healing" because it tries to fix its own mistakes before giving you an answer.

3. The "Translator" (Explainable AI)

Most AI systems give you a number (e.g., "Buy") but won't tell you why. They are "black boxes."

This system uses a Google Gemini AI not to make the decision, but to translate the decision into plain English. Think of the five agents as the "evidence" and the Gemini AI as a lawyer presenting the case in court. It writes a short paragraph saying, "We recommend buying this stock because the price is rising (Technical), the company has low debt (Fundamental), and the news is positive (Sentiment)." It cites the specific numbers it used, so you can trust the reasoning.

4. How Fast and Accurate is It?

The researchers tested this system on 10 different Indian stocks across 5 different industries (like IT, Banking, and Energy) over a 15-month period.

Speed: It's very fast. It takes about 26 to 32 seconds to analyze a stock and give you a report. That's fast enough to use while you are waiting for your coffee.
Accuracy: When the system predicted the stock would go up or down, it was right 72% of the time over the next five days. This is better than just guessing or using only one type of analysis.
Agreement: The five experts agreed with each other 78% of the time. When they did agree, the system was even more confident in its answer.

The Bottom Line

This paper presents a framework that turns the chaotic, confusing world of stock analysis into a structured, team-based process. It doesn't just give you a "Buy" or "Sell" signal; it gives you a reasoned argument backed by data from multiple angles. It's designed specifically for the Indian market, handling local news and market quirks, and it includes safety checks to prevent the AI from making wild guesses when the data is messy.

In short, it's like having a personal investment committee of five experts who check their work, argue their points, and then write you a clear, easy-to-read report in under 30 seconds.

Technical Summary: A Multi-Agent Generative AI Framework for Explainable Stock Market Analysis in Indian Equity Markets

Problem Statement
Retail equity investors and financial researchers in the Indian market face a critical challenge: analytical fragmentation. Technical, fundamental, and sentiment signals are traditionally processed in isolation, preventing a unified, confidence-aware analysis. Furthermore, existing AI-driven stock prediction tools often prioritize accuracy over interpretability, relying on "black-box" deep learning models (e.g., LSTM, GRU) that provide point predictions without transparent reasoning chains. This lack of explainability hinders regulatory compliance and investor trust, as decision-makers cannot verify the logic behind algorithmic recommendations.

Methodology
The authors propose a heterogeneous, five-agent framework designed specifically for Indian equities (NSE and BSE). The system operates on a modular, state-managed pipeline using Python asyncio for parallel execution and LangGraph for orchestration.

Data Ingestion: The system retrieves OHLCV data and fundamental metrics via yfinance and scrapes financial news from Moneycontrol and Economic Times using BeautifulSoup4. Text data is embedded using sentence-transformers and indexed in dual vector stores (FAISS and ChromaDB) for retrieval-augmented generation.
Specialist Agents (Parallel Execution):
- Technical Analysis Agent: Calculates RSI, MACD, multi-period EMAs (9, 21, 50, 200), and support/resistance levels. It generates a weighted score and confidence based on indicator alignment.
- Fundamental Analysis Agent: Evaluates intrinsic value using valuation (P/E, P/B, PEG), profitability (ROE, ROA), financial health (Debt/Equity, Current Ratio), and growth metrics. It employs Z-score normalization against sector medians.
- Sentiment Analysis Agent: Utilizes ProsusAI FinBERT for three-class sentiment classification, enhanced with an exponential time-decay weighting function ( $\lambda=0.9$ ) to prioritize recent news. A rule-based fallback (Loughran-McDonald lexicon) activates if the LLM fails.
- Sector Benchmarking Agent: Compares the target stock against 2–5 peer securities within 10 predefined Indian sectors, calculating percentile ranks for P/E, returns, and growth.
Orchestration and Anomaly Detection: An orchestrator manages the shared state (AnalysisState). An Anomaly Detector monitors for conflicts (e.g., opposing high-confidence signals, price-news divergence, or low-confidence outputs). Upon detection, it triggers a bounded re-verification loop (maximum two cycles) with extended data windows or fallback sources. If anomalies persist, the system flags the result for human inspection and applies a confidence penalty.
Synthesis and Reasoning:
- Synthesis Agent: Aggregates agent outputs using a confidence-weighted formula. It applies a consensus boosting mechanism: if 3 out of 4 agents agree on a direction with >60% confidence, the signal is upgraded (e.g., BUY $\to$ STRONG BUY).
- Generative Reasoning: The system invokes Google Gemini 2.0 Flash not as an analysis engine, but as a structured renderer. It receives the aggregated scores, evidence lists, and anomaly flags to generate natural-language investment reasoning with explicit evidence citations.
Interfaces: The system exposes a FastAPI REST backend for programmatic access and a Streamlit dashboard with Plotly visualizations for interactive screening.

Key Contributions

Heterogeneous Multi-Agent Architecture: Integration of technical, fundamental, sentiment, and sector analysis into a unified, state-managed pipeline, moving beyond homogeneous ensemble methods.
Confidence-Weighted Synthesis: A mechanism that quantifies multi-dimensional agreement and boosts signal strength when agent concordance exceeds a 75% threshold.
Self-Healing Anomaly Detection: A module that classifies specific conflict types (e.g., Technical-Fundamental Divergence) and executes bounded re-verification loops to resolve data quality issues autonomously.
Domain-Specific Sentiment Tuning: Implementation of FinBERT with exponential time-decay weighting, tailored to the rapid news cycle of Indian markets.
Architectural Explainability: A design where the Large Language Model (LLM) acts as a renderer for pre-verified signals rather than a predictor, ensuring that all generated reasoning is grounded in specific, traceable evidence.

Experimental Results
The framework was evaluated on 10 NSE-listed stocks across five sectors over a 15-month period (Jan 2024 – Mar 2025).

Consensus: The system achieved cross-agent consensus in 78% of test cases.
Directional Accuracy: The model achieved 72% directional accuracy against 5-day market moves, outperforming a buy-and-hold strategy (58%) and pure technical analysis (64%).
Latency: End-to-end inference averaged 32.4 seconds via CLI and 26.1 seconds via REST API, meeting sub-30-second targets for real-time interaction.
Anomaly Handling: Anomalies were detected in 30% of analyses; the re-verification protocol resolved 67% of these cases.
Confidence Correlation: A Pearson correlation of 0.61 was observed between confidence levels and directional accuracy. High-confidence recommendations (>70%) achieved 82% accuracy, compared to 58% for low-confidence ones.

Significance and Claims
The paper positions this work as a reproducible, interpretable architecture that addresses the "black-box" limitation of current financial AI. By integrating heterogeneous analytical perspectives with a self-healing mechanism and LLM-driven evidence citation, the framework offers a transparent decision-support tool for emerging markets like India. The authors claim that this approach democratizes advanced analytics, allowing retail investors to access confidence-aware, explainable insights without relying on opaque deep learning models. The system is presented as a proof-of-concept that can be built using open-source tools and public data, with specific pathways identified for future improvement, such as dynamic weight optimization via reinforcement learning and the development of an "Indic-FinBERT" model for better handling of Indian financial jargon.