Financial Bubble Detection with HLPPL

Co-first-authored research on identifying and quantifying financial bubbles in U.S. equity markets using a novel extension of the Log-Periodic Power Law (LPPL) framework. The Hyped LPPL model integrates sentiment-driven behavioral signals with the classical LPPL oscillatory crash-hazard structure, producing a dual-stream architecture for bubble detection and crash-time forecasting.

Research Question

Classical LPPL models capture the super-exponential price growth and log-periodic oscillations that characterize speculative bubbles, but they rely solely on price dynamics and are sensitive to fitting windows. This work asks: can incorporating NLP-derived sentiment signals improve both the detection accuracy and the timing precision of bubble identification? The HLPPL framework proposes a confidence-weighted sentiment stream fused with the traditional LPPL technical stream.

Methodology

The model has three components. First, a 7-parameter LPPL model is fitted via multi-start constrained optimization to identify bubble regimes and estimate critical times. Second, a confidence-weighted sentiment analysis pipeline using FinBERT and BERTopic extracts behavioral signals from financial text. Third, a Dual-Stream Transformer architecture fuses the LPPL-derived features with sentiment features through a regime-dependent BubbleScore that quantifies bubble intensity. The full pipeline is implemented in Python with PyTorch, featuring YAML-driven configuration and checkpoint-based resumability.

Results and Contribution

The Dual-Stream Transformer achieves an MSE of 0.087 and a correlation of 0.625 on BubbleScore prediction. Trading strategies derived from the model produce 34.1% annualized returns with a Sharpe ratio of 1.13 in backtesting. The work demonstrates that sentiment-augmented bubble detection meaningfully improves upon pure price-based LPPL models in both identification precision and economic value.

Publication and Presentation

Published as a preprint on arXiv (2510.10878). Presented at the 21st Quantitative Finance Conference 2025 in Rome and QuantMinds International 2025 in London. Co-authored with Zheng Cao, Xingran Shao, and Helyette Geman. The full implementation is available as an open-source Python package with comprehensive documentation.