Research

Testing what apparent predictive performance actually means.

Research across finance, machine learning and model reliability, designed to expose caveats, stronger baselines and the point where a promising result stops surviving.

Featured Transformer research

A score is the beginning of the investigation.

01

MSc Dissertation · Individual ResearchFinancial ML · Model Reliability · Markets

When Financial Transformers Look Predictive

The model looked predictive when all assets were pooled together. A simple asset-identity baseline did even better, and performance fell to roughly random when the model had to predict changes within the same asset.

Transformer pooled ROC-AUC: 0.790
Training-only static asset prior: 0.824
Pair-weighted within-asset ROC-AUC: 0.492

View case study

A 60-session market window enters the Transformer; pooled, static-prior, within-asset, identity, chronology and simulation diagnostics then test what supports the reported score.

Independent retrieval research

Search quality, complexity and interpretability.

Team Project + Independent ExtensionInformation Retrieval · Search · Model Evaluation

Lexical, Dense & Hybrid Information RetrievalKeyword and meaning-based search found complementary evidence. Combining them improved retrieval on two public benchmarks; a fixed neural reranker then made the strongest system slower and worse.

+0.0431SciFact ΔMAP, hybrid vs active lexical stage

Quantitative research

Strategies rebuilt under harder evidence.

Individual ResearchSystematic Investing · Backtesting · Robustness

Systematic Equity Factor ResearchAn attractive 28-stock backtest weakened when the universe broadened and concentration fell. Fresh data was still positive, but not statistically conclusive.

32.41%Original concentrated CAGR

Individual ResearchPairs Trading · Walk-Forward Testing · Robustness

Statistical Arbitrage & Walk-Forward ValidationThe original result did not survive a broader development test. Fresh confirmation later turned positive, but uncertainty remained too wide for a strong claim.

+20.72%Original total return

Individual ResearchAsset Allocation · Risk · Robustness

Portfolio Construction & Risk AllocationTwo portfolios can hold similar assets while concentrating risk very differently. The study compares both weights and risk contribution, then tests whether sophisticated optimisation earns its complexity.

Academic work

Learning under severe class imbalance.

Team ProjectComputer Vision · Class Imbalance · Deep Learning

Long-Tail Deep LearningA small number of classes contain many examples, while thousands of rarer classes have very little training data. The work tests techniques that stop overall accuracy from hiding failure on rare species.

96.20%CIFAR-10 accuracy