This repository contains the official poster presented at the SIAM Conference on Applications of Dynamical Systems 2025, showcasing our research on LPPLS-Net, a neural network model for forecasting critical points (crashes) in noisy financial markets.
The poster provides a concise summary of our proposed method, synthetic data generation, topological validation, and evaluation on real-world market scenarios (Dot-com and Lehman collapse).
Forecasting financial crashes is vital but challenging.
The Log-Periodic Power Law Singularity (LPPLS) model captures bubble dynamics but suffers from unstable parameter fitting.
We propose LPPLS-Net, a deep neural architecture trained on synthetically generated noisy time series (using white, AR(1), and GARCH(1,1) noise).
Our method achieves fast and accurate detection of critical points, with performance validated using Topological Data Analysis (TDA) and evaluated on real-world scenarios from NASDAQ-100 and S&P 500.
- Neural forecasting model trained on synthetic LPPLS + noise trajectories.
- Use of TDA-based filtering to ensure topological consistency of training data.
- Curriculum learning and soft-penalty losses for stable and accurate parameter recovery.
- Validated on Dot-com and Lehman crash periods using real market data.
To quickly try LPPLS-Net, see demo.ipynb.
This notebook provides:
- Setup and installation instructions
- Example code to generate or load data
- Step-by-step guide to train and using the LPPLS-Net model
- Visualization and experimentation tips
How to use:
- Run
demo.ipynblocally or in Google Colab for an interactive experience.
If you have any questions or would like to collaborate, feel free to reach out via LinkedIn.
© 2025 Varun Biyyala & Marian Gidea
This work was presented at the SIAM Conference on Applications of Dynamical Systems, Miami, 2025.