Detecting market bubbles: A generalized LPPLS neural network model

Abstract

To enhance bubble detection capabilities, we introduce two significant improvements to the Log-Periodic Power Law Singularity (LPPLS) model: (1) a novel fitting approach, which yields more accurate predictions of critical price distributions within a single sample window; (2) a restructured neural network approach further enhances the estimations of the probability distributions of the critical points across both time and price dimensions, and it can be fine-tuned with real-world data. The simulation and practical applications to typical asset price bubbles in cryptocurrencies, commodities, and equity indices demonstrate that our refined model, the Generalized-LPPLS Neural Network (G-LPPLS-NN), outperforms all other models we examined in terms of predictive accuracy for critical point distributions.