Research on the impact of LLM borrower action sequencing on executed credit in simulated lending markets
Get up to speed
Victoria Ruojie Li has released a revised working paper finding through simulated collateralized lending market experiments that shared action ordering among LLM borrowers significantly affects the scale of executed credit. The experiment set up 12 ordinary order markets and 12 balanced order markets, with an exact randomization test yielding a two-sided p-value of 2/8100; the credit difference between leveraged-first and deleveraged-first ordinary order markets was 1.12 units of initial unused reference borrowing capacity. In the experiment, three small open-source language models faced borrowing limits based on collateral valued at pool prices, and only two of the model families participated in trading. No liquidations or actual bad debts occurred in all 24 markets; under the preset 30% fixed-price shock benchmark, individual high loan-to-value loans still generated scenario write-offs even though the overall loan-to-value ratio remained below the write-off threshold, with most write-offs coming from collateral consumed by liquidation bonuses.
AI-generated from reports · 1 days agoUpdated
Coverage timeline
Follow the coverage to understand different sides of the event.
- arXiv q-finStudy on the Impact of Shared Action Ordering on Executed Credit for LLM Borrowers in a Simulated Lending Market
An experiment on a simulated collateralized lending market shows that shared action ordering significantly affects the scale of executed credit for LLM borrowers. The experiment set up 12 ordinary order markets and 12 balanced order markets, using an exact randomization test to obtain a two-sided p-value of 2/8100, with a credit difference of 1.12 units of initial unused reference borrowing capacity between leverage-prioritized and deleveraging-prioritized ordinary order markets. No liquidations or actual bad debts occurred in all 24 markets; under the fixed-price 30% shock benchmark, most write-offs came from collateral consumed by liquidation bonuses.
Trend of this event
Not enough consecutive observations yet; trend is not drawn.