TEXT ANALYSIS: Agentic Trading Paper

TEXT START

"A growing body of work explores how Large Language Models can be embedded in trading systems as agents that perceive market information, retrieve context, reason about decisions, emit tradable actions, and adapt under market feedback."

1. THE DISSECTION

This is a methodological audit dressed as a survey. It is not an advocacy piece selling LLM trading agents; it is a forensic snapshot of a field performing the pre-collapse theater of rigor — discovering, in aggregate, that the entire subfield is a reproducibility ghost town.

The paper does something unintentionally revealing: it maps the corpse before naming the cause of death. 77 studies, and only 2/19 meet even the most basic extraction threshold. The authors don't say it, but they are documenting a field that has produced architectural theater, not evaluable systems.

The core framing — "Architecture-Capability-Adaptation as a working analytical lens" — is a sophisticated way of saying: we can't measure what matters, so we'll describe what we can see.

2. THE CORE FALLACY

The paper treats protocol incomparability as a methodological problem when it is, under DT logic, an ontological signal. The field cannot reproduce itself because:

• There is no stable benchmark substrate. Markets are non-stationary, adversarial, and reflexive. Any "evaluation" is a snapshot of one regime, one market micro-structure, one slippage model. The absence of reproducibility isn't a fixable reporting gap — it's the market fighting back against the intrusion.

• The evaluative standards are borrowed from a dead paradigm. "Closed-loop evaluation" and "split protocols" were designed for supervised learning on stable targets. LLM trading agents operate in an environment where the act of deployment changes the evaluation signal (the agent's trades move prices, which changes the regime, which invalidates the protocol). This isn't a reporting checklist problem. It is structurally unsolvable.

• The 11/19 documenting "execution timing or semantics" is treated as a modest positive. It is, in fact, damning. Eleven of nineteen studies — in a domain where milliseconds and slippage can erase alpha — still don't document when and how trades actually execute. This is not rigor in formation. This is cargo-cult science with a financial veneer.

3. HIDDEN ASSUMPTIONS

• Assumption 1: LLM agents in trading will converge toward something evaluable at scale. No evidence is presented for this. The assumption is that methodological maturation will produce comparable protocols. Given that market regimes are adversarial and non-stationary, the opposite is structurally more likely: as agents proliferate, the evaluation substrate becomes more unstable, not less.

• Assumption 2: Architecture-Capability-Adaptation taxonomy is a useful lens. It treats "architecture" as a primary differentiator. Under DT logic, architectural choices in LLM-based systems are temporary moats at best — competitive advantages measured in months, not years. What matters is not the architecture but who owns the compute, the data moat, and the execution infrastructure. The paper treats the problem as an engineering problem; it is a political economy problem.

• Assumption 3: The bottleneck is "reporting." This is the most dangerous assumption. The authors frame 58 of 77 studies as "background and design context" — effectively excusing 75% of the literature as context-setting rather than evidence. This is the epistemic hedge of a field that hasn't accepted it is building on sand.

4. SOCIAL FUNCTION

This paper performs legitimation work for a field that cannot yet be legitimate. It is a transition management document: it acknowledges the chaos, provides a framework for appearing rigorous, and clears the path for continued funding and publication. The checklist contribution ("reporting checklist") is the most revealing element — it signals that the field understands it has a rigor crisis and responds by offering a reporting standard, not a theoretical or empirical fix.

The "Evidence Ledger, Reproducibility Audit, and Reporting Checklist as main contributions" is, structurally, the field's self-exoneration document. It says: we audited the mess, we found it's a mess, and we've given you a checklist to make future messes look more comparable. The market reality — that LLM agents in trading are racing toward a competitive commons that will destroy alpha as it is discovered — goes unexamined.

This is also prestige signaling to the academic CS/AI audience: a survey of 77 papers with protocol coding and reproducibility audits is credentialing material, not a threat to the research program.

5. THE VERDICT

The paper is a corpus callosum cut on a field already in systemic shock. It documents that the LLM-agent-in-trading research program cannot yet produce a single reproducible result across its entire primary evidence base. This is not a transitional phase. Under P1 (Cognitive Automation Dominance), financial markets are precisely the domain where AI first achieves durable superiority — high-frequency, pattern-heavy, data-saturated, low-physical-constraint. The paper's finding that 15/19 studies are R0 (not even computationally reproducible) means the entire field is operating on narrative alpha, not demonstrated alpha.

The real discontinuity this paper accidentally exposes: the evaluation infrastructure for AI trading agents cannot keep pace with the systems being evaluated. This is not a reporting gap. This is the market beginning to devour the researchers trying to study it. When even the meta-studies can't establish a baseline, the system has entered terminal velocity.

Survival read for humans: If you are in this research space, the paper's own findings are a structural warning. The field will consolidate around whoever controls execution infrastructure and proprietary data, not whoever publishes the most reproducible survey. The checklist is not a life raft. It is a certificate of death with a case number.