A single model is a single opinion. Our production stack runs 42 of them in parallel, across five architectural families, and asks the same question at every tick: is this the moment to act?
The ensemble, not the oracle
The temptation with modern AI is to search for the one model that is right most often. In market data, that model does not exist for long — regimes change, and yesterday's champion is tomorrow's noise. Ensembles trade a small amount of peak accuracy for a large amount of stability.
- 12 transformer variants trained on tick-level microstructure
- 8 temporal convolutional networks on session-window features
- 10 gradient-boosted trees on cross-asset correlations
- 6 state-space models focused on regime detection
- 6 rules-based sanity checks with a hard veto
The tie-break
When the vote is close, the router defers. Doing nothing is a valid output. A signal is only released when a supermajority agrees within a bounded confidence band — and every release ships with an attribution map that a human analyst can read after the fact.
// Principle
A quiet model is a healthy model. Silence is a feature, not a failure.