Applied AI Research Lab
The hardest problems in complex domains are not prediction problems. They are representation and decision problems under uncertainty.
Problem Space
These systems are chaotic, partially observed, and non-stationary. Before you can optimise decisions, you need to learn good representations of the underlying state.
| Domain | Surface Problem | True Problem |
|---|---|---|
| Financial markets | What is happening? Is there structure? Is this a crisis? | Learning cross-asset representations and planning over representations. |
| FX timing | When to convert | Optimal stopping under uncertainty |
| EEG signals | Generalisation across patients | Learning invariant latent representations |
The HL Framework
01
Learn a latent space
02
Optimise decisions
03
Optimise representations
Each system follows the same pattern. The latent space is not an intermediate step. It is the foundation everything else is built on.
Systems
Unified encoder across 28 financial assets (FX, bonds, equities, commodities) compressed into a 128-dimensional embedding. The state representation layer for all downstream systems.
hestonlabs.com/latent →SMBs lose 3–5% annually from poor currency exchange timing. Forava reframes this as a fixed-horizon optimal stopping problem, solved over Latent embeddings of the financial universe.
forava.xyz →No standardised benchmark exists for evaluating models trained on cross-asset data. CARL establishes evaluation protocols across market-state retrieval, correlation structure, and portfolio construction.
Paper forthcomingCollaborate
If you're working on a problem where the raw data is not the right formulation (chaotic, partially observed, non-stationary), we'd like to hear about it.
Start a conversation →Accelerator
Contact
contact@hestonlabs.com