Posts
-
HMSPC: A Hybrid Mechanistic-Stochastic Physical-Continuous Model for Battery Dynamics
By replacing purely autonomous latent dynamics in Neural ODEs and Latent ODEs with input-conditioned, gated continuous-time evolution, HMSPC better captures regime-dependent structure in battery voltage time series data. This inductive bias, combined with variational inference and uncertainty-aware decoding, leads to consistent gains over Latent ODEs (Rubanova et al., 2019) in both reconstructive accuracy and calibration.
-
Investigating the Invisible: Analyzing Market Behavior and Divergence using an IDEC Autoencoder
Can an autoencoder learn the hidden structure of markets? I trained one on cross-sectional equity returns and found three stable regimes, and a reconstruction error that spiked on Liberation Day before I even knew why.
-
Beyond Black Boxes: A Framework for Building Rigorous AI for Volatility Forecasting
I built a hybrid GARCH-LSTM for volatility forecasting, then tried to prove it wasn't just memorizing noise. Monte Carlo permutation testing across 8 stocks revealed something uncomfortable: not all stocks are actually predictable.