From Pixels to SignalsHierarchical Vision Transformers for Return Prediction
I study whether modern vision models can extract return-predictive signals from price charts, and examine which parts of the chart matter for their predictions.
PhD student · Department of Economics and Business Economics · Aarhus University
Aarhus University · Started February 2025
Aarhus University · Completed in three semesters
Aarhus University
Main supervisor: Jesper N. Wulff · Co-supervisor: Bezirgen Veliyev
I study whether modern vision models can extract return-predictive signals from price charts, and examine which parts of the chart matter for their predictions.
A price chart and a table can represent the same market history. This paper investigates what accounts for differences between models trained on these representations.
I examine how numerical, graphical, and textual presentations of market histories shape the financial forecasts supplied by a common language model.
Preliminary draft — work in progress. Please do not cite. Comments welcome.
Instructor; developed all course materials
Teaching assistant, Economics programme, Aarhus University
Supervision of six master’s theses and two internship reports.
Research visit with Claes Bäckman
Presented From Pixels to Signals
Presenter; chair, Deep Learning for Financial Time Series
Presented From Pixels to Signals
Computer Vision Application in Financial and Energy Markets
Philip Bond · Aarhus University
Jesper N. Wulff · Aarhus University
Dong Li · Aarhus University
Bryan Kelly, Dacheng Xiu & Semyon Malamud · Yale School of Management
Aarhus University
Jes Frellsen, Pierre-Alexandre Mattei & Jakub Tomczak · Sophia Antipolis