Dr. Ignacio Deza

Predictive Reliability & Model Stress-Testing

Standard machine learning often assumes abundant data and Gaussian noise. When industrial systems violate these assumptions, models become fragile. My work focuses on stress-testing predictive architectures to identify hidden failure modes—such as temporal myopia—and implementing rigorous alternatives grounded in statistical physics.

Physics-Grounded Data Science

Complex systems theory and non-equilibrium statistical mechanics provide the tools to extract signal from "pathological" residuals. These methods are essential for subscription markets, aerospace maintenance, and any high-stakes environment where the cost of a false assumption is prohibitive.

Rapid Domain Absorption

By recognising universal patterns across different scales, I have moved from climate dynamics to energy forecasting and aerospace. A strong theoretical foundation allows for the quick mastery of domain-specific challenges through the lens of structural data patterns.

Rare Event & Extreme Risk Forecasting

When outcomes are infrequent but critical—such as equipment failure or rapid subscriber churn—standard classification models systematically under-evaluate risk. I develop architectures that respect true class imbalance.

Applications: Aerospace maintenance, subscription markets, climate extremes

Low-Sample & Sparse Data Modelling

Building reliable predictive systems for feature spaces with thousands of dimensions but limited samples. This requires dimensionality reduction that maintains physical interpretability while avoiding overfitting.

Applications: Industrial IoT, medical diagnostics, sensor networks

Non-Gaussian & Heavy-Tailed Risk

Standard models fail when data exhibits power laws or Lévy flights. I implement methods that account for the heavy-tailed behaviour where actual commercial risk typically resides.

Applications: Financial time series, turbulence, network traffic