TR EN

MSc Thesis Defense: Nisa Defne Aksu, MODELING THE EFFECTS OF MARKET NEWS AND AI-EMULATING ADVISORS ON FAIRNESS AND FINANCIAL INCLUSION IN DEFI CREDIT AND TRADING SYSTEMS, Date & Time: June 17, 2026 – 1:00 PM, Place: FENS G025

MODELING THE EFFECTS OF MARKET NEWS AND AI-EMULATING ADVISORS ON FAIRNESS AND FINANCIAL INCLUSION IN DEFI CREDIT AND TRADING SYSTEMS

 

Nisa Defne Aksu
Computer Science & Engineering, MSc Thesis, June 2026

 

Thesis Jury

     Prof. Selim Saffet Balcısoy (Thesis Advisor)

  Prof. Galina Andreeva

  Asst. Prof. Dilara Keküllüoğlu

 

 

Date & Time: June 17, 2026 – 1:00 PM

Place: FENS G025

Zoom Linkhttps://sabanciuniv.zoom.us/

Meeting ID: 590 502 4125

 


Keywords : Decentralized Finance (DeFi), Agent-Based Simulation (ABS), Behavioral Clustering, Algorithmic Fairness, LLM-like Advisory Agents

 

Abstract

 

Millions of users with various behavioral profiles have been onboarded via decentralized finance (DeFi) protocols, but little is known about the fairness implications of AI-mediated financial advising in these environments. This thesis examines whether large language model (LLM)-like advisory agents in two DeFi contexts — credit and token trading — produce systematically uneven financial outcomes through news sentiment propagation. A dual agent-based simulation framework is implemented using real on-chain data from 18,901 MakerDAO vault records and 2,405,340 ERC-20 addresses. Each simulation deploys 500 heterogeneous agents across three advisor types — fear-based, neutral, and optimistic — and nine empirically derived news sentiment clusters over 20 cycles. The primary finding is that advisor effects are systematically heterogeneous across user behavioral profiles — concentrated 77 times more strongly among moderate-activity users than high-volume users, a result not predetermined by simulation design. A secondary finding is that fear-based advisors produce higher credit scores and portfolio values under declining market conditions, reflecting alignment between cautious guidance and DeFi protocol incentive structures rather than universal advisor superiority. Fairness audit reveals that standard demographic parity and disparate impact criteria are met, yet the equal opportunity analysis of variance (ANOVA) fails at p = 0.0018, indicating that access-based metrics can pass while outcome-distribution audits detect significant disparities. These findings support an expanded fairness auditing framework that evaluates not only credit access but how advisory systems systematically shape user financial trajectories over time.

Home

Orta Mahalle, 34956 Tuzla, İstanbul, Türkiye

Telefon: +90 216 483 90 00

Fax: +90 216 483 90 05

© Sabancı Üniversitesi 2023