Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning

Yijia Xiao; Edward Sun; Tong Chen; Fang Wu; Di Luo; Wei Wang

Home/Archive/arXiv

arXiv · 2509.11420

Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning

Yijia Xiao·Edward Sun·Tong Chen·Fang Wu·Di Luo·Wei Wang

View original source ↗Read source document ↗Export metadata ↓

Abstract

Developing professional, structured reasoning on par with human financial analysts and traders remains a central challenge in AI for finance, where markets demand interpretability and trust. Traditional time-series models lack explainability, while LLMs face challenges in turning natural-language analysis into disciplined, executable trades. Although reasoning LLMs have advanced in step-by-step planning and verification, their application to risk-sensitive financial decisions is underexplored. We present Trading-R1, a financially-aware model that incorporates strategic thinking and planning for comprehensive thesis composition, facts-grounded analysis, and volatility-adjusted decision making. Trading-R1 aligns reasoning with trading principles through supervised fine-tuning and reinforcement learning with a three-stage easy-to-hard curriculum. Training uses Tauric-TR1-DB, a 100k-sample corpus spanning 18 months, 14 equities, and five heterogeneous financial data sources. Evaluated on six major equities and ETFs, Trading-R1 demonstrates improved risk-adjusted returns and lower drawdowns compared to both open-source and proprietary instruction-following models as well as reasoning models. The system generates structured, evidence-based investment theses that support disciplined and interpretable trading decisions. Trading-R1 Terminal will be released at https://github.com/TauricResearch/Trading-R1.

Explore related subjects

q-fin.TR cs.AI cs.CE cs.CL cs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yijia Xiao, Edward Sun, Tong Chen, Fang Wu, Di Luo, Wei Wang. 2025-09-14. Trading-R1: Financial Trading with LLM Reasoning via Reinforcement Learning. https://arxiv.org/abs/2509.11420

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

The Marginal Effects of Ethereum Network MEV Transaction Re-Ordering

Two MEV builders now produce nearly 80\% of Ethereum blocks. Block builders have the ability to reorder transactions on the blockchain in a way that can be harmful to participants. We estimate participants would pay in the aggregate nearly \$7.2 million per month to guarantee that they remained in the first quartile of the block. Sandwich attacks, in which a transaction is front run, are frequent, averaging more than one every two blocks. Gas fees on these transactions pay for nearly 9.6\% of the MEV payments to the validator. Reforms such as gas fee priority or private transaction pools might be helpful.

q-fin.TR↗

arXiv2026

Computable Countermarkets and the Limits of Universal Trading

We explain why no trading algorithm can guarantee profit in every market. For each deterministic program that always returns a finite-precision position, we construct a fixed, algorithmically generated price path on which every active position loses and inactivity earns nothing. This holds with positive, continually changing prices, costless trading, and unlimited computation time. Separate arguments limit learning market rules, certifying future events, and establishing randomness from finite data. Useful strategies may exploit market structure, information, or compensation for risk, while benchmark performance need not imply profit. Reversing and rearranging price histories within the assumed market class provide practical stress tests, distinguishing conditional success from universal guarantees.

q-fin.TR↗

arXiv2026

Adapting the Actor Model of Concurrency for High-Frequency Trading: Synchronous Message Delivery (fast_send) and a Tick-to-Book Latency Study

The actor model - state isolation, data-race freedom, deadlock resistance, and sequential single-message reasoning - has long been dismissed as unsuitable for high-frequency trading (HFT): actors seem to imply many threads, a mailbox per actor, and a heap-allocated message plus a context switch per interaction, overhead incompatible with a microsecond budget. This paper argues the dismissal is wrong for co-located actors, and supports it both analytically and with a deployed, measured implementation: kaspar-hft, an open-source C++20 framework. Four extensions adapt the model for HFT: fast_send, a synchronous delivery mechanism in which the sending thread runs the receiver's handler inline and returns the reply as a value; actor groups, which co-schedule actors on one thread behind a shared mailbox; per-actor selectable mailbox queues; and a memory pool. fast_send has receiver transparency: the handler cannot tell whether delivery was synchronous or asynchronous, or which thread runs it. A grouped synchronous chain runs on one thread, cutting scheduler context switches from O(N) to O(1), and a thread-local call-chain test catches cyclic invocation before any lock is taken. Microbenchmarks put the synchronous round trip at tens of nanoseconds. On a live CME market-data feed (ES, NQ, ZN futures), socket-to-book latency decomposes into a ~7 microsecond decode-and-book floor plus a per-message slope; the framework's own contribution is under 1% of the floor. The tail is set not by the actor machinery but by the market's non-Poisson, clustered arrival process, characterized in a companion paper. The shared-queue group also yields a production/simulation duality: the same actor code runs unchanged in live trading and deterministic backtest.

q-fin.TR↗