The research team behind Uvistium comprises thirteen people with backgrounds in quantitative finance, machine learning, and market microstructure. What they actually do.
Uvistium runs three separate reinforcement-learning agents, each optimised for a specific market regime:
A meta-selector classifies the prevailing market regime every 60 seconds and routes decisions to the appropriate agent. Classification uses a combination of Hurst exponent, realised volatility, and cross-asset correlation structure.
Model type, in detail. All three agents are Proximal Policy Optimisation (PPO) with a modified reward function that maximises Sharpe ratio over a 60-trading-day horizon, with explicit penalty for drawdowns above 8%. Neural network architecture: 4-layer MLP with LayerNorm and ELU activation.
Data drawn from more than 40 sources — commercial and open:
Only walk-forward validation, no in-sample backtests. Every model is trained on a rolling 24-month window and tested on the subsequent 6 months. A model enters production only after robust out-of-sample performance across several consecutive periods.
Current production models have been walk-forward validated across 2007–2026 — spanning the 2008 financial crisis, the 2011 Eurozone crisis, the Brexit vote of 2016, the March 2020 COVID crash, the 2022–2023 inflation shocks, and the sterling-crisis mini-budget aftermath of autumn 2022.
The research team is organisationally separated from sales, marketing, and the executive board. Research findings — including negative ones that lead to model decommissioning — are communicated internally without commercial consideration.
When a model demonstrates persistent underperformance in a market phase, it is decommissioned. This happened once this calendar year (deactivation of the first-generation Range-V2 agent in April 2026), without delay and without marketing consultation.
Alongside the quantitative models, the team continuously monitors the macro indicators most relevant to UK and global markets: Bank of England MPC decisions and minutes, ONS inflation and labour-market data, PMI surveys, quarterly GDP prints, and gilt-market signals. These feed into the models as features but are not the primary signal source.
Uvistium