Calibrating Trust in the Age of LLMs: Mitigating Anchoring and Automation Bias in Financial Decision Making
Domains
Tech Stack
Project Summary
Abstract
Retail investors increasingly use LLMs to interpret markets, but fluent AI explanations can also distort judgment through anchoring, automation bias, and misplaced trust.
I built MarketSim, a full-stack trading simulator with a local RAG pipeline and a deliberately frictional interface: split Bull/Bear commentary, a gated final-synthesis reveal, and independent technical risk nudges designed to slow blind acceptance.
In a controlled study of 18 participants, 216 trading decisions, and 12 stocks, the system showed that AI meaningfully changes behavior: it reduced risky buys for speculative users by about 23% while increasing risky buys for aggressive investors from 20% to 44%.
What I Built
- AI support changed behavior by investor persona rather than acting as a neutral assistant across 18 participants and 216 trades.
- Users strongly preferred compressed AI summaries, with 81% of AI-enabled trades using the final-synthesis reveal flow.
- The experiment found weak anchoring rather than blind automation, with decision-range variance narrowing by 11.75%.
Impact
- Reduced risky buys for speculative users while revealing where the same AI layer can amplify risk for aggressive investors.
- Demonstrated why responsible FinTech products need adaptive cognitive friction instead of one-size-fits-all AI guidance.
Page Info
The Challenge
Retail investors increasingly turn to LLMs for financial interpretation, creating risks around information overload, automation bias, and anchoring on fluent AI explanations.
The Cognitive Firewall
Built MarketSim as a cognitive sandbox with dialectical Bull/Bear commentary, hidden final synthesis, and algorithmic risk nudges that operate independently of the LLM.
Experiment and Findings
Ran a controlled study with 18 participants, 216 trading decisions, and 12 stocks across AI-enhanced and control environments to measure confidence, risk perception, and trust calibration.