IRIS - The Engine

Built for markets as they are.

Finance-first AI, by design Axyon IRIS is our proprietary AI prediction framework - purpose-built to navigate the complexity, noise, and non-stationarity of real financial markets.

The Axyon IRIS Framework

IRIS processes data through four sequential, rigorously controlled stages — each designed to eliminate bias and maximise signal quality.

01.Data Sources

"Garbage in, garbage out"

High-quality data is the foundation of any successful AI application, as algorithms cannot turn noise into signal. Our approach is inherently data-centric, with an automated AI factory designed to continuously enhance data breadth, depth and quality.

10+ Data Sources

Morningstar · FactSet · LSEG · FTSE Russell · RavenPack · CUSIP Global Services

Market Data

End-of-day and intraday prices, indices, commodities, VIX, and currency exchange rates.

Fundamentals & Macro

Corporate fundamentals, macroeconomic indicators, analyst forecasts, and options data.

Sentiment & Alternative

Sentiment indicators extracted from news and social media. Proprietary client data is supported in select partnerships.

02.Data Processing

Up to 1bn data points per model

The processing pipeline starts by ingesting, cleaning, and organising diverse data streams so our models learn from reliable, high-quality information. It transforms raw inputs into ML-ready features while eliminating lookahead bias.

Point-in-time validation

+2 Millions Rows of Data ingested per day (Annual average data in 2025)

Lookahead bias correction

Handling survivorship bias and corporate actions

Feature normalisation and z-scoring

Outlier detection and treatment

03.AI Modeling

Engineered for rigour, automation and scale

The engine identifies persistent patterns across markets and generates predictive signals on the relative future performance of securities. It leverages Learning-To-Rank (LTR) models, ensemble methods and scalable AutoML architectures.

10,000-100,000 AI models trained for each new investable universe

+1000 Heterogeneous AI models in production

~2 ^ 10000 AI ensembling search space size

Scalable hybrid infrastructure leveraging High Performance Computing (HPC) and cloud clusters

04.Strategy Building

Translating predictive relative performance signals into systematic, investable strategies.

AI-based signals are analysed and used to build systematic investment strategies (e.g. long-only, long/short), supporting bottom-up aggregation and end-to-end explainability.

Smart rebalancing — threshold-based (signal-driven) or calendar (weekly, monthly, quarterly)

Simulated fees & transaction costs modeled directly in the backtest

Investable universe — sector, geography, long-only or long-short

Position limits, sector caps, ESG overlays and other mandate constraints

How Predictions Are Made

Our approach is grounded in a fundamental truth: markets cannot be "solved." They are adaptive, complex, and constantly evolving. Our technology reflects this reality.

The engine that learns to rank

Axyon IRIS uses Learning-to-Rank (LTR) — a family of ML techniques originally developed for search engine ranking, here applied to rank assets by expected relative performance.

Point Wise

Pairwise

Listwise

A Models Factory Approach

Rather than relying on a single all-encompassing algorithm, we develop and manage a diverse factory of AI models - each designed to capture different signals across different market conditions.

Robustness Across Regimes

Capture of Diverse Signals

Enhanced Adaptability

No black box

Every prediction is decomposed via SHAP into 12 feature categories. Investment managers can see exactly why an asset ranked high or low. Supports regulatory audit trails under MiFID II. Examples:

Momentum+0,18

Earnings Quality+0,12

Sentiment+0,07

Valuation-0,5

Macro-0,3

The foundation behind the signal.

Axyon AI's compute infrastructure was architected in partnership with CINECA, one of EU's largest high-performance computing centre - providing the raw GPU power needed for AutoML at institutional scale.