The research question

A forecast of tomorrow’s return can be viewed as a probability distribution rather than a single value. My thesis investigated how ideas from data-driven optimal transport could inform that distribution, and how dynamic programming could connect a forecast to a sequence of decisions.

The mathematical approach

The work draws on the distributional barycenter problem and data-driven flows. It considers conditional density estimation, transport maps, and numerical methods for relating observed samples to an unknown distribution.

The thesis builds on work by Esteban G. Tabak, Giulio Trigila, and Wenjun Zhao. It describes a formulation that replaces a more difficult minimax optimization with a single minimization, and investigates how the approach might be applied to financial time series.

Financial setting and scope

The thesis discusses Apple, Tesla, Google, and Amazon price data collected through Tiingo, with numerical exploration in a Jupyter Notebook. Dynamic programming is considered as a way to use a conditional distribution of future returns in a financial decision problem.

This was preliminary academic work. The thesis develops an approach and its possible application; it does not establish a validated trading strategy or a claim of investment outperformance.

Explore the geometry.

This small interactive example shows how probability mass moves between two one-dimensional Gaussian distributions.

The geometry of changeFIG. 01

Moving between distributions.

SourceTransportTarget
One-dimensional Gaussian 2-Wasserstein interpolation0.00.20.4-6-3036Probability density
Move through the transportt = 0.50
Source distributionTarget distribution
Analytical illustration · 2-Wasserstein geometry

Drag the control to move a Gaussian distribution from the source to the target. Its mean and standard deviation interpolate linearly; this moves probability mass rather than mixing the two endpoint densities.

Mean 0.10Standard deviation 1.05Endpoint W₂ distance 4.230
The mathematics

For one-dimensional normal distributions with means μ₀, μ₁ and positive standard deviations σ₀, σ₁, the displacement interpolation is:

μₜ = (1 − t)μ₀ + tμ₁
σₜ = (1 − t)σ₀ + tσ₁
W₂² = (μ₀ − μ₁)² + (σ₀ − σ₁)²

The illustration uses μ₀ = −2, σ₀ = 0.8, μ₁ = 2.2, and σ₁ = 1.3. It is a teaching example created for this website, separate from the financial experiments discussed in the thesis.

Reference: Peyré & Cuturi, Computational Optimal Transport ↗
Read the original thesis (PDF)Explore all research