More than ex-post fitting: log-periodic power law and its AI-based classification

Ganghyeok Lee, Minhyuk Jeong, Taeyoung Park, Kwangwon AhnHumanities 2025 · DOI

Abstract

The log-periodic power law (LPPL) model can be used to model the formation and evolution of endogenous bubbles in asset prices to predict the critical time of bubble bursts; however, it has a limitation in that the parameter interpretation and reliability are contingent on the researchers’ subjective judgments. We introduce an artificial intelligence-based framework to address the reliability issue of the LPPL model. Specifically, we use an AI classification model to produce a reliability score, which is then incorporated into a novel risk metric designed to provide a more reliable early warning indicator of potential financial crashes. To train classification models, we created over 13 million labeled LPPL parameter sets estimated for daily closing prices of large-cap stocks in the United States (covering 24 years). We consider a simple formula to reformulate the crash information provided by the LPPL outcomes to create our risk metric, distance-to-crash with AI (DTCAI). When interpreting the LPPL results through our proposed risk metric, greater confidence can be placed on the predicted crash time as the metric inherently incorporates reliability information. Finally, the economic significance of the DTCAI metric is subsequently demonstrated through its application in dynamic portfolio-rebalancing strategies. The results show that strategies guided by our DTCAI metric significantly outperform the benchmark buy-and-hold strategy, regardless of the investor’s risk preference. This study suggests that integrating artificial intelligence into financial models like LPPL can enhance the accuracy of bubble detection and risk quantification, offering a more robust approach to anticipating financial crises and optimizing investment strategies.

Introduction

Predicting financial bubbles and their subsequent crashes remains a central challenge in quantitative finance. The Log-Periodic Power Law (LPPL) model is a prominent mathematical tool designed for this purpose, predicated on the theory that irrational herding behavior among market participants leads to unsustainable, faster-than-exponential price growth accompanied by accelerating oscillations. This process eventually culminates in a "critical time" or market crash.

However, the LPPL model faces a significant practical limitation: it is prone to "ex-post fitting." Because the model is highly non-linear and requires the estimation of seven distinct parameters, it often produces crash predictions for any given time series, regardless of whether a genuine bubble exists. Historically, researchers have relied on subjective manual filtering to distinguish true warnings from mathematical artifacts, leading to high false alarm rates. This study introduces an AI-based framework to automate and objectify this reliability assessment.

Methodology

The proposed framework replaces subjective human judgment with an objective AI classification pipeline. The process began with the generation of over 13.8 million LPPL parameter sets using a Genetic Algorithm (GA) applied to historical S&P 500 data (1995–2018). These sets were labeled based on their proximity to actual historical crashes: a set was labeled "reliable" if the predicted crash date fell within 10 days of an actual event.

To handle the massive class imbalance—where reliable crash signals are rare—the researchers employed the Synthetic Minority Oversampling Technique (SMOTE). They then trained various classifiers, with the Artificial Neural Network (ANN) emerging as the top performer due to its ability to capture complex non-linear relationships between parameters.

Architecture of the Artificial Neural Network (ANN) used for classifying LPPL parameter reliability.

The core innovation is the introduction of a new risk metric: Distance-to-Crash with AI (DTCAI). While the standard Distance-to-Crash (DTC) metric measures temporal proximity to a predicted critical time, DTCAI multiplies this value by the reliability probability score generated by the AI. If the AI identifies a spurious fit, the probability approaches zero, effectively neutralizing the false alarm.

Results and Economic Significance

The integration of AI led to a dramatic reduction in false signals. In an out-of-sample test covering the volatile 2019–2024 period (including the COVID-19 crash and the 2022 decline), the raw LPPL model generated 360 crash alerts. The AI-enhanced DTCAI metric reduced this to 199 alerts, significantly improving the signal-to-noise ratio and reducing the false alarm rate from 85.8% to 34.7%.

Comparison of crash alerts between the raw DTC metric and the AI-enhanced DTCAI metric, demonstrating a significant reduction in false alarms.

Beyond statistical accuracy, the study demonstrated the economic utility of the DTCAI metric through dynamic portfolio backtesting. Strategies that rebalanced from stocks to bonds based on DTCAI signals consistently outperformed a standard buy-and-hold benchmark across various risk profiles. For instance, an aggressive strategy guided by DTCAI achieved a Sharpe ratio of 0.0592, compared to 0.0440 for the benchmark.

A notable finding was the model's performance during the early 2020 COVID-19 crash. Although the pandemic was an exogenous shock, the DTCAI metric flashed alerts beforehand, suggesting that the market was already in a state of endogenous fragility—the pandemic merely served as the trigger for an existing bubble.

Conclusion

This research demonstrates that the inherent "quality" of LPPL fits can be learned by AI models. By automating the classification of parameter reliability, the DTCAI framework transforms the LPPL from a sensitive theoretical tool into a robust, actionable early warning system. While the current model was trained on large-cap US stocks, future work will focus on generalizing these models to other asset classes and high-frequency data to further refine crash detection capabilities.

Summary updated

Abstract

The log-periodic power law (LPPL) model can be used to model the formation and evolution of endogenous bubbles in asset prices to predict the critical time of bubble bursts; however, it has a limitation in that the parameter interpretation and reliability are contingent on the researchers’ subjective judgments. We introduce an artificial intelligence-based framework to address the reliability issue of the LPPL model. Specifically, we use an AI classification model to produce a reliability score, which is then incorporated into a novel risk metric designed to provide a more reliable early warning indicator of potential financial crashes. To train classification models, we created over 13 million labeled LPPL parameter sets estimated for daily closing prices of large-cap stocks in the United States (covering 24 years). We consider a simple formula to reformulate the crash information provided by the LPPL outcomes to create our risk metric, distance-to-crash with AI (DTCAI). When interpreting the LPPL results through our proposed risk metric, greater confidence can be placed on the predicted crash time as the metric inherently incorporates reliability information. Finally, the economic significance of the DTCAI metric is subsequently demonstrated through its application in dynamic portfolio-rebalancing strategies. The results show that strategies guided by our DTCAI metric significantly outperform the benchmark buy-and-hold strategy, regardless of the investor’s risk preference. This study suggests that integrating artificial intelligence into financial models like LPPL can enhance the accuracy of bubble detection and risk quantification, offering a more robust approach to anticipating financial crises and optimizing investment strategies.
BibTeX
@inproceedings{lee2025more,
  author = {Ganghyeok Lee and Minhyuk Jeong and Taeyoung Park and Kwangwon Ahn},
  title = {{More than ex-post fitting: log-periodic power law and its AI-based classification}},
  booktitle = {Humanities},
  year = {2025},
  doi = {10.1057/s41599-025-05920-7},
}
Figure