Tokenless

The AI that solves the market: A new era in forecasting with natural language explainability

LG AI Research introduces its advanced financial forecasting framework, which powers a US equities market ETF (LQAI) and a new "Master Score with Commentary" product with LSEG. The system uniquely combines structured financial data with unstructured text from news and reports, using the proprietary Exaone LLM and a multi-agent architecture to deliver explainable, accurate, and actionable market predictions across the entire US stock market.

3 min read
The AI that solves the market: A new era in forecasting with natural language explainability
source: https://www.youtube.com/watch?v=zbmXulPIJpo

LG AI Research has leveraged its foundational research capabilities and proprietary 32-billion parameter LLM, Exaone, to build a sophisticated AI-powered forecasting framework. Initially proven through internal LG projects—such as supply and demand forecasting for LG Electronics and raw material price forecasting for LG Energy Solutions—this framework has demonstrated significant real-world value, saving millions of dollars and optimizing complex industrial processes.

These internal applications enabled the development of core assets and knowledge in AI-based forecasting, specifically:

  • Hybrid Data Integration: Combining structured numerical data with unstructured text from news reports, corporate filings, and social media.
  • Deep Document Understanding (DDU): A model for extracting insights from the complex multimodality of financial documents, including tables, charts, and graphs.
  • Advanced Time-Series Analysis: Utilizing Transformer-based models and methodologies like mixed-frequency, multivariate ensemble learning.
  • Explainable AI (XAI): Using the same LLM that makes a prediction to generate natural language explanations, addressing the “black box” problem prevalent in finance.

A Multi-Agent Framework for Financial Forecasting

The core of the system is the Exaone Business Intelligence AI, a multi-agent framework designed for comprehensive future insights. It systematically processes information through several layers:

  • Data Layer: Ingests a wide array of data, including market price data, macroeconomic indicators, company fundamentals, real-time news (over 5,000 articles daily), analyst reports, and earnings call transcripts.
  • Model Layer: Features a suite of models including deep learning time-series forecasters, the Exaone LLM, a reasoning model (Exaone Deep), the DDU model for unstructured data, and various embedding and search models.
  • Module Layer: Handles the critical and difficult task of parsing and pre-processing financial data, followed by embedding, RAG (Retrieval-Augmented Generation), forecasting, ranking, and portfolio construction.
  • Agent Layer: Composed of role-based agents that mimic a human financial team:
    • AI Journalist: Identifies and contextualizes news.
    • AI Economist: Performs forecasting and sentiment analysis.
    • AI Analyst: Generates explanations and rationale for decisions.
    • AI Decision Maker: Synthesizes all information to make a final decision (e.g., a trade, a score, or a portfolio adjustment).

Real-World Application and Performance

This framework is not just theoretical; it underpins two major financial products:

  1. LQAI ETF: An AI-powered ETF listed on the NYSE that manages a portfolio of 100 US large-cap stocks, rebalanced monthly entirely by the AI. The fund has successfully outperformed its S&P 500 benchmark by 300 basis points in 2024. The AI generates a monthly report explaining its investment decisions, much like a human portfolio manager’s letter, providing crucial transparency.

  2. Master Score with Commentary: A new product developed in partnership with global financial data provider LSEG (formerly Refinitiv).

    • It provides a forecasted return score (from 1 to 100) for a given stock over the next four weeks.
    • Crucially, it includes an AI-generated commentary that explains the rationale behind the score, making the abstract number actionable for clients.
    • The service provides unprecedented coverage of the entire US equities market (~5,500 stocks), including nano-cap stocks that human analysts cannot cover.

Back-testing from 2021 to 2024 shows strong performance. Stocks in the top score quintile (80-100) saw an average 20-day return of +3.1%, while those in the bottom quintile (1-20) returned -4.0%. The model’s directional accuracy, or “hit rate,” is between 51% and 59%, a highly favorable rate in quantitative finance.

The ultimate goal is to mimic human cognitive processes—combining quantitative data with qualitative insights from text—but to do so at a scale and speed that is impossible for humans. By integrating all available information, the system can anticipate market shifts, enable confident decisions, and optimize resource allocation.

Share post

Stay In The Loop! Subscribe to Our Newsletter.

Get updates straight to your inbox. No spam, just useful content.

Recommended for You

Unlocking Unstructured Data with LLMs

Unlocking Unstructured Data with LLMs

Shreya Shankar of UC Berkeley discusses DocETL, a MapReduce-style framework that leverages LLMs to extract, analyze, and structure insights from unstructured enterprise data. The conversation covers practical architecture patterns, the role of non-determinism, strategies for model selection (including fine-tuning and multi-LLM pipelines), and the importance of user experience in this emerging field.

Building Production-Grade RAG at Scale

Building Production-Grade RAG at Scale

Douwe Kiela, CEO of Contextual AI, explains the evolution from basic RAG to "RAG 2.0", an end-to-end, trainable system. He argues that this system-level approach, which integrates optimized document parsing, retrieval, reranking, and grounded models, is superior to relying on massive context windows alone and is a fundamental tool for next-generation AI agents.