Till innehåll på sidan

Mårten Frösslund: Discrete Time Simple Directed Stochastic Graph Model with Degree Dependent Connection Probabilities

Master thesis

Tid: Må 2026-11-23 kl 11.00

Plats: KTH, Room 3418

Respondent: Mårten Frösslund

Exportera till kalender

Deriving analytical graph model probability formulas and comparing model implementation on Tokyo Stock Exchange Index data to structural model

Abstract: Interconnectedness between companies affect credit risk and should be considered in systemic risk analyses. Previous works have explored constant connection probabilities, transition probabilities between different states, and existing network types.

This paper defines and analyses a novel simple directed stochastic graph model with degree dependent connection probabilities.

Nodes are classified as either activated, defaulted, or unactivated. Activated nodes are non-isolated nodes, defaulted nodes are isolated nodes that were activated, and unactivated nodes are isolated nodes yet to be activated. Defaulted nodes cannot be activated. Unactivated nodes are activated by gaining an edge to any other non-defaulted node. The next graph is found by using defined matrix operators which explicitly depend on the chosen edge probability functions and the current system state.

Directed connection probabilities can be described by any probability density function dependent on the degrees of the incident nodes. Using independent standard uniform random variables the edge probability functions are sampled and the graph is stochastically evolved. Analytical formulas for the system are derived on any set of nodes for the next time step. Centrally for the graph application on companies, the probability of nodes defaulting is examined for any number of time steps. The graph model and derived probability formulas were implemented in Python and was tested by calculating explicit values from Tokyo Stock Exchange Index data from August 2021 through August 2026. The resulting time series of probability of system default was compared to an established structural model as a proof of concept. The novel graph model presents a generalised framework on how to model discrete time stochastic processes found in the real world.