From Price Noise to Market Semantics: A Self-Adapting Large Language Model Framework for Financial Intelligence

Overview

Financial markets are fast changing with time-series data, but large language models (LLMs) are trained only with a point-in-time data. The purpose of this project is to create a self-learning LLM that’s fully aware of financial markets in real-time so investors can make decisions quickly and accurately. In Step-1, we transform raw market data into symbolic representations that enables continuous self-improvement through controlled fine-tuning and self-edits. Raw 1-minute OHLCV data is converted into Renko blocks and stored in the database. In Step 2, Renko data is transformed into a three-layer representation comprising: symbolic tokenization, market structure state, and a time-indexed event log. This symbolic representation allows the LLM to reason over market dynamics, as the self-learning framework operates on symbolic, not numeric tokens. In Step 3, the LLM receives this data, it produces an output containing Market Meaning, Strategy Class, Confidence, Explanation, and Self-Critique. In Step 4, we enter a self-evalution loop, where it proposes self-edits, synthetic improvements to its reasoning and decision framework. These edits are evaluated through inner and outer optimization loops with reinforcement learning. Weight updates are permanently applied if performance improves; otherwise, the model reverts to the prior stable state. In conclusion, after the first round of self-evalution loop our LLM's interpretation quality value is 88%. The goal of the loop is not infinite fine-tuning, but to guide learning that increases the model’s accuracy.

From Price Noise to Market Semantics: A Self-Adapting Large Language Model Framework for Financial Intelligence

Overview

Competition history

Related projects

AI-Powered Financial Literacy Assistant

Deriving Unsupervised Fluid Intelligence from Schema-Less Polymorphic Unstructured Data

Improving the Rationale for Stock Market Investments to Help Middle Income Households

Keep Your Data Close, but Your Failures Closer: Failure-Driven Adversarial Self-Evolution of Language Models

When Words Move Markets: How the Tone of Fed Communications Shape Stock Market Returns and Volatility

Using Multivariate Time Series Forecasting Neural Networks With Sentimental Analysis To Determine Future Stock Prices

Emergent Behavior in Constrained and Unconstrained Financial Environments: Portfolio Volatility and Overtrading

Statistical Inferences on Non-stationary Increments Self-similar Stochastic Processes via Lamperti Transformations

Browse more like this