{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,6]],"date-time":"2026-10-06T04:53:08Z","timestamp":1791262388234,"version":"4.1.0"},"posted":{"date-parts":[[2026]]},"group-title":"SSRN","reference-count":0,"publisher":"Elsevier BV","license":[{"start":{"date-parts":[[2026,1,1]],"date-time":"2026-01-01T00:00:00Z","timestamp":1767225600000},"content-version":"unspecified","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>Calibrating an agent-based limit-order-book (LOB) simulator is a prerequisite for credible simulated backtests of execution algorithms and market making, yet every evaluation is expensive, stochastic, and derivative-free. We study this calibration under a fixed budget of 50 simulator evaluations in an 11-dimensional parameter space. Six numerical baselines (Sobol sampling, a Treestructured Parzen Estimator (TPE), Bayesian Adaptive Direct Search (BADS), Gaussian-process Bayesian optimization (GP-BO), multiobjective TPE, and component-aware GP-BO) are compared with Optimization by PROmpting (OPRO) driven by the large language model (LLM) Claude Sonnet 5, with and without frozen domain heuristics. We ask (1) whether an LLM proposal policy can allocate a fixed simulator budget competitively and (2) whether domain heuristics improve OPRO. A deterministic evaluator scores ticker-specific calibrations for five B3 equities with a standardized, grouped simulated-method-of-moments (SMM) loss over 29 moments of returns, spreads, imbalance, trade size, trade timing, and depth. All optimizers share one agent configuration, search domain, initial design, and budget, and selected parameters are evaluated on held-out dates. Averaged over five tickers and five repetitions, heuristic OPRO attains the lowest mean calibration endpoint (3.375) and held-out Test Book loss (4.575), against 5.002 and 5.973 for plain OPRO and 4.004 and 5.938 for component-aware GP-BO, the strongest numerical method. In this low-budget regime, OPRO is competitive with common black-box optimizers, and domain heuristics lower both its mean loss and its run-to-run variability.<\/jats:p>","DOI":"10.2139\/ssrn.7557238","type":"posted-content","created":{"date-parts":[[2026,10,6]],"date-time":"2026-10-06T04:24:47Z","timestamp":1791260687000},"source":"Crossref","is-referenced-by-count":0,"title":["Do Domain Heuristics Help LLM Optimizers Calibrate an Agent-Based Limit Order Book?"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0005-2975-6594","authenticated-orcid":true,"given":"Bernardo","family":"Teixeira","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0009-0245-2805","authenticated-orcid":true,"given":"Joao Pedro","family":"Alexandrino","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-9284-8310","authenticated-orcid":true,"given":"Renan","family":"Avila","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,10,6]],"date-time":"2026-10-06T04:24:47Z","timestamp":1791260687000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7557238"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7557238","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}