Temporal Relational Reasoning of Large Language Models for Detecting Stock Portfolio Crashes

TL;DR

This work tackles the problem of detecting stock portfolio crashes during rare, unprecedented events by enabling zero-shot reasoning over a dynamically built temporal-relational graph constructed from daily news. The proposed Temporal Relational Reasoning (TRR) framework comprises four cognitive-inspired phases—brainstorming, memory, attention, and reasoning—that generate and prune a temporal graph linking news articles to intermediate entities and portfolio stocks, before a final LLM-based crash prediction. TRR demonstrably outperforms state-of-the-art baselines across crisis periods (2007, 2010, 2020) and maintains robustness during stable periods, with an additional demonstration of extending the approach to macroeconomic crisis detection. The framework highlights the potential of structured, memory-informed, graph-based reasoning over evolving information streams to predict complex, interdependent market phenomena and to serve as a pragmatic tool for practitioners and policymakers alike.

Abstract

Stock portfolios are often exposed to rare consequential events (e.g., 2007 global financial crisis, 2020 COVID-19 stock market crash), as they do not have enough historical information to learn from. Large Language Models (LLMs) now present a possible tool to tackle this problem, as they can generalize across their large corpus of training data and perform zero-shot reasoning on new events, allowing them to detect possible portfolio crash events without requiring specific training data. However, detecting portfolio crashes is a complex problem that requires more than reasoning abilities. Investors need to dynamically process the impact of each new piece of information found in news articles, analyze the relational network of impacts across different events and portfolio stocks, as well as understand the temporal context between impacts across time-steps, in order to obtain the aggregated impact on the target portfolio. In this work, we propose an algorithmic framework named Temporal Relational Reasoning (TRR). It seeks to emulate the spectrum of human cognitive capabilities used for complex problem-solving, which include brainstorming, memory, attention and reasoning. Through extensive experiments, we show that TRR is able to outperform state-of-the-art techniques on detecting stock portfolio crashes, and demonstrate how each of the proposed components help to contribute to its performance through an ablation study. Additionally, we further explore the possible applications of TRR by extending it to other related complex problems, such as the detection of possible global crisis events in Macroeconomics.

Figures (9)

• Figure 1: Illustration of thought-based frameworks (e.g., ToT sun2023think, GoT besta2024graph), search-based frameworks (e.g., ToG sun2023think), and our proposed Temporal Relational Reasoning (TRR) framework, with Relational Only and Relational+Temporal variants.
• Figure 2: The components of our Temporal Relational Reasoning (TRR) framework. TRR emulates the human cognitive capabilities used for solving complex problems, that include (1) brainstorming, (2) memory, (3) attention and (4) reasoning skills.
• Figure 3: Parameter selection of $q$.
• Figure 4: Examples of generated graphs during the crash periods in the 2007, 2010 and 2020 dataset.
• Figure 5: Examples of the economic crisis indicator generated from our TRR framework. The line coloured red represents the ground truth crisis labels, i.e., a value of 1 represent a crisis; 0 otherwise. The other coloured lines represent the crisis indicators.