Abstract
The rapid spread of information over social media influences quantitative trading and investments. The growing popularity of speculative trading of highly volatile assets such as cryptocurrencies and meme stocks presents a fresh challenge in the financial realm. Investigating such "bubbles" - periods of sudden anomalous behavior of markets are critical in better understanding investor behavior and market dynamics. However, high volatility coupled with massive volumes of chaotic social media texts, especially for underexplored assets like cryptocoins pose a challenge to existing methods. Taking the first step towards NLP for cryptocoins, we present and publicly release CryptoBubbles, a novel multi-span identification task for bubble detection, and a dataset of more than 400 cryptocoins from 9 exchanges over five years spanning over two million tweets. Further, we develop a set of sequence-to-sequence hyperbolic models suited to this multi-span identification task based on the power-law dynamics of cryptocurrencies and user behavior on social media. We further test the effectiveness of our models under zero-shot settings on a test set of Reddit posts pertaining to 29 "meme stocks", which see an increase in trade volume due to social media hype. Through quantitative, qualitative, and zero-shot analyses on Reddit and Twitter spanning cryptocoins and meme-stocks, we show the practical applicability of CryptoBubbles and hyperbolic models.
Introduction
Cryptocurrency markets represent a unique frontier in finance, characterized by volatility levels often ten times higher than traditional equities. This volatility is frequently driven by "bubbles"—periods of rapid price escalation fueled by social media contagion and investor exuberance. Traditional financial forecasting models often struggle in this environment because they fail to account for the non-linear, hierarchical nature of social media influence.
This research addresses two primary gaps: the lack of high-quality, large-scale datasets for cryptocurrency-related NLP and the inadequacy of Euclidean models in capturing the power-law dynamics of market hype. We introduce CryptoBubbles, a comprehensive dataset and a novel "multi-span identification" task designed to detect the specific start and end dates of multiple, potentially overlapping bubbles within a given lookahead window.
The CryptoBubbles Dataset and Task
To facilitate research in this area, we curated and released the CryptoBubbles benchmark, which includes:
- Scale: Data for 456 cryptocoins across 9 major exchanges over a five-year period (2016–2021).
- Social Media Integration: Over 2.4 million tweets corresponding to these assets.
- Ground Truth Labeling: Bubbles were identified using the Phillips-Shi-Yu (PSY) model, a standard statistical tool for detecting exuberance, and further validated by five experienced financial analysts (achieving a high Cohen’s $\kappa$ of 0.93).
- Cross-Platform Generalization: A separate test set of 25,000 Reddit posts covering 29 "meme stocks" (e.g., GameStop, AMC) was curated to evaluate zero-shot performance.
Unlike simple regression or ranking tasks, the multi-span identification task requires models to output the exact temporal boundaries of market anomalies, providing actionable risk management data for investors.
Methodology: Multi-Bubble Hyperbolic Network (MBHN)
The core innovation of this work is the Multi-Bubble Hyperbolic Network (MBHN). The model moves away from standard Euclidean neural networks, which assume "flat" relationships, toward Hyperbolic Learning.
Why Hyperbolic Space?
Social media influence and market shocks follow power-law distributions. These structures are inherently hierarchical and "scale-free," where a few influential users have a disproportionately massive impact. Hyperbolic space (specifically the Poincaré Ball) expands exponentially, making it mathematically superior for representing these tree-like hierarchies compared to flat Euclidean space.
Model Architecture
- Hyperbolic GRU (HGRU): The model processes temporal data (prices and tweets) directly within the hyperbolic manifold using a Gated Recurrent Unit implemented with Möbius operations.
- Temporal Attention: An adaptive mechanism learns to distinguish between market "noise" and high-impact signals by weighting hidden states based on their relevance to price movement.
- Multi-Span Extraction: The model predicts the probability of each day being a bubble "start" or "end." It then employs Non-Maximum Suppression (NMS)—a technique adapted from computer vision—to select the most likely non-overlapping bubble periods.
Experimental Results
Our evaluation demonstrates that the MBHN significantly outperforms traditional and state-of-the-art baselines:
- Predictive Accuracy: MBHN achieved an F1-score of 0.53 on bubble span prediction, a massive improvement over traditional ARIMA models (0.02) and advanced financial NLP models like SN-HFA.
- Zero-Shot Generalization: Despite being trained on Twitter and cryptocurrency data, the model generalized remarkably well to Reddit and meme stocks. For instance, it identified bubbles in the CCIV stock with an F1-score of 0.73, despite never seeing that asset during training.
- Data Efficiency: The model is highly efficient, requiring only a 5-day lookback window ($\tau = 5$) to make accurate predictions, whereas conventional methods often require 100+ days of historical data.
Discussion and Conclusion
The success of MBHN suggests that the "geometry of hype" is fundamentally hyperbolic. By capturing the hierarchical nature of social media contagion, the model can identify the "surprise" or sentiment shifts that precede bubble bursts—signals that price-only models frequently miss.
While the model provides high predictive power, future work will focus on improving the interpretability of hyperbolic transformations and incorporating non-public data such as institutional trading volumes. Beyond finance, this hyperbolic multi-span approach holds potential for other domains characterized by power-law dynamics, including epidemiology and real-time healthcare monitoring.
Summary updated