Systems shipping in production today.
A representative slice of recent Perceptronix work. Some projects are publicly published with academic partners; others are described at architectural level only, under client NDAs.
MASCET — Macroeconomic Analytics System for Cross-cutting Economic Trends
MASCET is a modular AI platform for macroeconomic forecasting, developed collaboratively with the National Institute of Economic and Social Research (NIESR) and the University of Birmingham, supported by the EPSRC.
At its core sits a multi-recurrent neural (MRN) ensemble trained on a wide panel of macroeconomic indicators. The system produces UK CPI inflation forecasts to within ±0.2% across multi-month horizons, and on multiple occasions has detected US inflation turning points five months earlier than the Survey of Professional Forecasters.
MASCET's outputs have been published by NIESR (Winter 2025, Spring 2024) and in peer-reviewed journals.
MLB Scion — Major League Baseball outcome modelling
MLB Scion is a calibrated outcome-modelling system for Major League Baseball, developed for a private analytics client. The pipeline ingests and aligns more than 23,000 historical MLB games, augmented with player-, team-, and venue-level features.
A stacked ensemble — combining LightGBM with deep neural networks for player-level signals — produces probability-calibrated match outcomes suitable for downstream decisioning, with explicit uncertainty estimates.
The system is wired into a CI/CD pipeline that retrains nightly, publishes performance dashboards, and surfaces drift before it bites.
Detailed architecture and results are confidential under NDA.
FOREX Scion — Multi-recurrent neural ensemble for FX trading
FOREX Scion is a live algorithmic trading system for FX CFDs. Its signature is a dual-MRN ensemble:
- → Macroeconomic MRN — trained on FRED and OECD macro panels to capture regime-level signals.
- → Technical MRN — trained on tick-aligned price, volume, and microstructure features for short-horizon directional signals.
- → Majority-vote arbitration — a deterministic ensemble layer reconciles the two MRNs into a single position recommendation with explicit confidence.
Live data flows in from Finage, OANDA, MetaTrader, and FRED. State persists in MariaDB on AWS EC2/EBS, deployment runs through GitHub Actions, and every model revision is walk-forward validated before being allowed to trade.
Specific signals, returns and risk parameters are confidential under NDA.
Have a forecasting or classification problem that needs to work in production?
Tell us about it. First conversations are confidential, no-obligation, and usually end with a clear view of feasibility, data needs, and time-to-value.