Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Text2TimeSeries: Enhancing Financial Forecasting through Time Series Prediction Updates with Event-Driven Insights from Large Language Models

Litton Jose Kurisinkel, Pruthwik Mishra, Yue Zhang

TL;DR

This work tackles the challenge of forecasting stock prices by integrating textual event signals from large language models with traditional time-series predictions. It introduces TimeS, a modular framework where a per-stock time-series model forecasts future prices, an LLM predicts discrete price-change indicators for upcoming days from event text, and a forecasting updater merges these insights to yield refined predictions. Through ExtEDT data and T5-based label prediction, the authors demonstrate that LLM-informed updates can improve RMSE/MAE over strong baselines, particularly for short-horizon event windows, while acknowledging limitations tied to model choice and data sources. The approach offers a principled pathway for multimodal financial forecasting that leverages event-driven insights to augment numeric predictions in real time.

Abstract

Time series models, typically trained on numerical data, are designed to forecast future values. These models often rely on weighted averaging techniques over time intervals. However, real-world time series data is seldom isolated and is frequently influenced by non-numeric factors. For instance, stock price fluctuations are impacted by daily random events in the broader world, with each event exerting a unique influence on price signals. Previously, forecasts in financial markets have been approached in two main ways: either as time-series problems over price sequence or sentiment analysis tasks. The sentiment analysis tasks aim to determine whether news events will have a positive or negative impact on stock prices, often categorizing them into discrete labels. Recognizing the need for a more comprehensive approach to accurately model time series prediction, we propose a collaborative modeling framework that incorporates textual information about relevant events for predictions. Specifically, we leverage the intuition of large language models about future changes to update real number time series predictions. We evaluated the effectiveness of our approach on financial market data.

Text2TimeSeries: Enhancing Financial Forecasting through Time Series Prediction Updates with Event-Driven Insights from Large Language Models

TL;DR

This work tackles the challenge of forecasting stock prices by integrating textual event signals from large language models with traditional time-series predictions. It introduces TimeS, a modular framework where a per-stock time-series model forecasts future prices, an LLM predicts discrete price-change indicators for upcoming days from event text, and a forecasting updater merges these insights to yield refined predictions. Through ExtEDT data and T5-based label prediction, the authors demonstrate that LLM-informed updates can improve RMSE/MAE over strong baselines, particularly for short-horizon event windows, while acknowledging limitations tied to model choice and data sources. The approach offers a principled pathway for multimodal financial forecasting that leverages event-driven insights to augment numeric predictions in real time.

Abstract

Time series models, typically trained on numerical data, are designed to forecast future values. These models often rely on weighted averaging techniques over time intervals. However, real-world time series data is seldom isolated and is frequently influenced by non-numeric factors. For instance, stock price fluctuations are impacted by daily random events in the broader world, with each event exerting a unique influence on price signals. Previously, forecasts in financial markets have been approached in two main ways: either as time-series problems over price sequence or sentiment analysis tasks. The sentiment analysis tasks aim to determine whether news events will have a positive or negative impact on stock prices, often categorizing them into discrete labels. Recognizing the need for a more comprehensive approach to accurately model time series prediction, we propose a collaborative modeling framework that incorporates textual information about relevant events for predictions. Specifically, we leverage the intuition of large language models about future changes to update real number time series predictions. We evaluated the effectiveness of our approach on financial market data.
Paper Structure (37 sections, 8 equations, 12 figures, 7 tables)

This paper contains 37 sections, 8 equations, 12 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. TimeS: Overall Method
  4. $T_{s}$:Time Series Model
  5. $F$:Stock state computation using Indicators Predicted by large Language Models
  6. Amplification Prediction using Temporal stock State $S_{t}$
  7. U: Updating time Series Price Predictions
  8. Loss:
  9. Experiments
  10. Datasets
  11. ExtEDT: Extended EDT Dataset with News Events and Time Series Data
  12. Dataset: Training Time Series Models
  13. Fine tuning LLM for Price Change Label Prediction
  14. Settings:
  15. Evaluation and Results
  16. ...and 22 more sections

Figures (12)

  • Figure 1: Stock Price Dynamics: Event Induced Changes in Time Series
  • Figure 2: TimeS: In the lower portion of the diagram, the LLM utilizes stock and event data as inputs to forecast price change indicators for the subsequent $n$ time intervals, which are then employed to determine stock states. In the upper portion of the diagram, time series are updated using price amplification values derived from these stock states.
  • Figure 3: Time series depicted as a Random Walk on a 2D Grid, where at every time point, it may either increase, decrease, or remain neutral, denoted by 1, -1, and 0, respectively.
  • Figure 4: SentiEvent: Base Model Setting for Price Amplification Prediction Using Bert
  • Figure 5: CASE STUDY1:Accurate Prediction During Moderate Upward Price Movement,Stock:FICO, ,Event:"FICO Recognized by Chartis as Category Winner in Innovation, AI Applications, and Financial Crime-Enterprise Fraud; Ranked Sixth Overall in the 2021 Chartis RiskTech 100 Report Position Reflects FICO's Analytic Innovation Strategy and Ability to Help Organizations Manage the Complexity of Their Analytic Assets. SAN JOSE, Calif., Nov. 30, 2020 /PRNewswire/ -- Highlights: FICO ranked sixth in this year's RiskTech 100 a comprehensive study of the world's major solution providers in risk and compliance technology FICO was recognized as category winner in Innovation for the fourth consecutive year FICO also won category awards for AI Applications and Financial Crime - Enterprise Fraud Global analytics software provider FICO, today announced that it has ranked sixth in Chartis Research's annual RiskTech100 report of world's leading risk technology providers. FICO also won category awards for Innovation, AI Applications, and Financial Crime Enterprise Fraud. ""FICO's top-ten ranking reflects its innovation strategy"", said Sid Dash, research director at Chartis Research." Expected Labels:INC_5 INC_16 INC_17, Predicted Labels:INC_6 INC_11 INC_16 "
  • ...and 7 more figures