Financial Market Prediction with LLMs

machine-learningfinancial modelling

Overview

Financial market prediction is one of the most difficult and computationally expensive machine learning (ML) objectives. Current research shows that ML models for market prediction can greatly benefit from exogenous signals such as processed sentiment. In this dissertation research project, I explore the three popular forecasting models, develop sentiment extraction methods, provide a robust feature engineering pipeline and optimise hyperparameters. Results show a 16% MAE reduction.

Technical Approach

The multimodal pipeline consists of several components, designed in isolation to promote best testing and engineering practices. Below is an outline of those main components:

  • Data Collection Unit - this component gathers large quantities of hourly market data for several ETF (Exchange Traded Fund) tickers using Alpaca API web socket. Sentiment data, which consists of US presidents' tweets, is collected via scraping and open-source datasets.
  • Pre-Processing & Feature Engineering Unit - a component responsible for deriving technical indicators (e.g., moving averages, RSI, etc.), integrating sentiment scores, and ingesting temporal features (calendar-based features like day of the week)
  • Sentiment Extraction Unit – a component responsible for deriving sentiment scores from unstructured tweets using FinBERT and FinGPT models. FinGPT model was custom-prompted to produce a 6-way classification schema.
  • Modelling Unit – a component consisting of LSTM/GRU/TFT models responsible for producing multi-horizon hourly predictions.
  • Hyperparameter Tuning Unit – a component responsible for tuning the hyperparameters of the ML models on training and validation datasets.
  • Evaluation Unit – a component responsible for evaluating the pipeline on the held-out test dataset.
  • CLI Interface – an easy-to-use CLI for managing experiments.

Results

  • Developed a novel custom classification schema with gating mechanics for FinGPT LLM, allowing for a more robust capture of sentiment signals.
  • Reduced the MAE by 16% compared to the baseline setup without sentiment inclusion.

Links

Future Plans

Since this project was part of my Master's degree and was time-constrained, there are several extensions I would like to implement. I have many ideas, but the core ones are:

  1. Extend the scope of the data by substituting ETF data for crypto data and presidential sentiment for financial news and reddit/twitter sentiment. Current version uses a relatively low amount of data, and, by extending the scope, I am hoping to push the pipeline generalisation abilities even further
  2. Extend the project to deal with live streaming of data. Live inference has always been a major goal of this project, and I'm hoping it will result in an interesting and useful trading tool.
  3. Developing backtesting strategies to test out the pipeline in real-world scenarios. I'm hoping to see how it compares on live market data and how much profit (most likely loss) can be ripped.
  4. A live dashboard. Dashboards are cool!