Latent

Latent is a convolutional autoencoder designed around spatio-temporal methods for multivariate financial time series. The architecture captures both the temporal dynamics within each asset and the cross-asset spatial relationships across the input at each timestep.

The encoder takes a rolling 720-hour window of the 28 input instruments and compresses it into a fixed 128-dimensional vector. The decoder reconstructs the input from this embedding, training the latent space to retain only the structure most critical for explaining cross-asset co-movement and regime transitions.

The resulting embedding is a compact, continuous representation of the current market state, analogous to a coordinate in a learned regime space.