Havrylchyk, O., Mariotto, C., Rahim, T. and Verdier, M. (2021). The Expansion of Peer to Peer Lending, Review of Network Economics, 19 (3), 145-187
(Winner: Best Paper Award, Toronto FinTech Conference)
Evaluation and Fine-Tuning of Generative Artificial Intelligence Models using Discrete Choice Experiments, 2025
Recent advances in Generative AI models have resulted in many practical applications, but the evaluation of such models remains an unsolved problem. While consensus is converging on aligning outputs with human preferences, existing methodologies lack solid foundations in statistical decision theory. I propose a method to evaluate the subjective output of Generative AI models by adapting Discrete Choice Experiments conducted in the field. The estimated parameters of the choice function can then be used to solve for optimal hyper-parameters of the model by addressing the firm’s custom-defined optimization problem. This method adapts the Discrete Choice Experimental Survey framework for real-time application on software data products like Generative AI models, whose performance depends on user preferences revealed through user choices. The approach rests on solid foundations in statistics and choice theory.
Demand Estimation for Differentiated Goods in LLM-enriched Characteristic Space, 2025
I develop novel ways to incorporate LLM-generated embeddings into estimation methods of demand for differentiated products. The embeddings are generated from textual descriptions of products using a pre-trained Large Language Model (LLM) and successfully capture rich semantic meanings from language by mapping the text into the space of real numbers. Such information is usually not captured by human encoding or by traditional deterministic methods. I contrast different approaches of incorporating these high-dimensional embeddings into demand models and show their relative strengths in predictive performance and in removing bias from the main coefficients of interest, such as price and other interpretable product characteristics. I demonstrate the method with an application to demand for grocery products
Can Online Platforms Improve Resource Allocation by Controlling Prices? , 2022
In many Peer-to-Peer (P2P) online platforms, the prices are set by platforms themselves to clear the market which restricts the market's ability to aggregate information and reflect it in prices. In this paper, I use micro data from an online P2P credit platform to show evidence of a better allocation of credit when prices are set by the platform instead of by the market. I develop a structural econometric model of loan demand and repayment in the presence of asymmetric information. The model exploits exogenous variation in the platform's pricing schedule to identify the effects of prices on borrower choices. I use the estimated model to conduct a counterfactual experiment in which borrowers were offered prices determined by the market. The counterfactual price distribution was generated by using machine learning and exploiting a past change in the pricing mechanism. I find that when the market sets prices instead of the platform, the prices are higher by 3.74% on average and this difference is increasing in borrower riskiness. Consequently, borrowers demand smaller loans, switch to shorter maturity contracts, and are more likely to default. In the aggregate, demand and repayment of loans decreased by 10% and 2%, respectively.
The Impact of Reducing Asymmetric Information on Platform Structure, 2022
I estimate the impact of reducing asymmetric information by implementing finer risk-based pricing in the context of a Peer-to-Peer (P2P) credit market. I exploit an exogenous change in the platform’s credit scoring policy which increased the number of risk categories from 8 (coarse scoring) to 49 (fine scoring). The centralized price-setting rules of the platform ensured that the one-to-one relationship between credit scores and prices remained intact unlike in a traditional credit market where it is broken. I find the effects of the policy to be three-fold: (i) it had positive effects on loan size, the fraction of loan amount repaid, and rates of return for investors, (ii) it skewed the distribution of platform revenues from lenders to borrowers, (iii) among borrowers, it shifted platform revenue distribution from high-risk to low-risk borrowers.