Dr. Matteo Ottaviani

This work proposes using the feature of large language models (LLMs) to learn about the "data bias" injected by diverse SDG classifications into bibliometric data by exploring five SDGs. We build a LLM that is fine-tuned in parallel by the diverse SDG classifications inscribed into the databases' SDG classifications. Our results show high sensitivity in model architecture, classified publications, fine-tuning process and natural language generation. The wide arbitrariness at different levels raises concerns about using LLM in research practice.

For the Bayesian learner, updating priors based on observation is the path that improves the beliefs and this, in turn, affects the efficiency of prices. Or so was believed until agents with no learning advantage, that is zero-intelligence, produced near-optimal levels of efficiency. We design a market environment characterized by model misspecification in which agents update their initial probabilistic guesses with increasing sophistication in incorporating observations. In this designed environment, the market penalizes divergence from the truth, measured in relative entropy. We show that a non-linear U-shaped relation brings together individual rationality in learning and quality of pricing (i.e. informative efficiency).

Agent-based models successfully allow to computationally set up socio-economic environments populated by heterogeneous agents interacting with one another and their environment based on internal behaviours, social norms, and data learned from past observations. Nowadays, the large availability of data on many dimensions considerably enriches the dynamic description. However, parameter space exploration, calibration, and sensitivity analysis (SA) have become computationally prohibitive. I propose a novel, general, multi-purpose, and trustworthy procedure (i.e., Monte Carlo-Once-At-a-Time); either it is an advanced local SA method itself or it serves as a reliability check before building the surrogate of an ABM for performing global SA.

The paper explores market selection in an Arrow-Debreu economy with complete markets where agents learn over misspecified models. Standard Bayesian learning loses formal justification, and biased learning processes may provide a selection advantage. The ecology of traders in the market affects selection dynamics and long-run asset valuation. Model misspecification makes it difficult to rank learning behaviors based on survival prospects. Prediction averaging has an advantage when the true data generating process belongs to the same family of models, but this advantage disappears when the true model belongs to a more general class. Rules guaranteeing survival exploit imitative mechanisms.