{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,9,29]],"date-time":"2026-09-29T20:13:27Z","timestamp":1790712807187,"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-nc-nd\/4.0\/"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"abstract":"<jats:p>Abstention by a language-model trading agent is not neutral: the agents do not disagree with the rule that has their inputs, they abstain, and for an agent that holds a position an abstention is a sale avoided. We run a preregistered paper-trading system in which three agents decide daily trades in five ETFs beside the moving-average rule on the same price inputs. In 262 live and 2,173 anonymised historical decisions the agents opposed the rule in 0% and 1.1% of cases, with no detectable gain. A larger model, on 300 points that over-represent days on which the close is within two per cent of a long-term (150-day) moving average, opposed it 1 time and abstained on 80% of those days. A calendar of scheduled macroeconomic events changed 64 decisions, 59 of them into abstentions; this is a change of behaviour, and we find no improvement in quality. Live, the agent ledgers and the rule ledgers hold different positions on about one asset-day in five, and the consensus gate between agents blocked 22 decisions, all of them sales. The abstention depends on one sentence of the prompt: replacing it with an instruction to commit lowers the abstention rate from 55% to 19%, and the released decisions follow the rule. The same input sent twice changes 9% of decisions. We propose a minimal evaluation protocol for agent papers. The results cover two models from one provider and a short live record.<\/jats:p>","DOI":"10.2139\/ssrn.7541520","type":"posted-content","created":{"date-parts":[[2026,9,29]],"date-time":"2026-09-29T19:48:56Z","timestamp":1790711336000},"source":"Crossref","is-referenced-by-count":0,"title":["Abstention Is Not Neutral: Ablating the LLM in an LLM Trading Agent"],"prefix":"10.2139","author":[{"ORCID":"https:\/\/orcid.org\/0009-0007-7513-2920","authenticated-orcid":true,"given":"MinGyu","family":"Lee","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"78","container-title":[],"original-title":[],"deposited":{"date-parts":[[2026,9,29]],"date-time":"2026-09-29T19:48:56Z","timestamp":1790711336000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.ssrn.com\/abstract=7541520"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026]]},"references-count":0,"URL":"https:\/\/doi.org\/10.2139\/ssrn.7541520","relation":{},"subject":[],"published":{"date-parts":[[2026]]},"subtype":"preprint"}}