model card
A convolutional autoencoder that compresses 28 financial instruments into a 128-dimensional latent vector, capturing regime-level cross-asset behaviour.
at a glance
architecture
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.
training data
Hourly price data across all 28 instruments, spanning 2008 to 2017. The training window covers the 2008 GFC and recovery, the European sovereign debt crisis and multiple FX and commodity regime shifts, giving the model a broad sample of cross-asset stress and correlation behaviour.
Equities & Volatility
- FTSE100
- SP500
- NIKKEI
- VIX
FX
- DXY
- EUR-USD
- GBP-USD
- USD-JPY
- USD-CHF
- USD-CAD
- AUD-USD
- NZD-USD
- EUR-JPY
- EUR-GBP
- EUR-AUD
- EUR-CHF
- GBP-JPY
- GBP-AUD
- AUD-JPY
- NZD-JPY
Rates
- US2Y
- US10Y
- UK2Y
- UK10Y
- Germany2Y
- Germany10Y
Precious Metals
- XAU-USD
- XAG-USD
intended use
- 01Regime detection and classification across market cycles
- 02Cross-asset correlation analysis and contagion mapping
- 03Risk signal generation ahead of structural market shifts
- 04Feature input for downstream financial ML models
- 05Research into latent structure of multi-asset markets
limitations
- 01Trained on historical data — may not generalise to unprecedented market structures
- 02Embedding dimensions have no guaranteed semantic interpretation
- 03Does not directly predict prices, returns or volatility
- 04Coverage limited to the 28 included instruments; idiosyncratic signals from other assets are not captured
- 05Regime labels are inferred post-hoc; the model is unsupervised and does not output named regimes