金融研究arXiv cs.LG14 Oct, 12:00 pm SGTSIGNAL DD58F1

Code-Interactive Candlestick Reasoning: Fin-Thinker Makes Evidence Verifiable

ORIGINAL / Code-Interactive Visual Reasoning for Financial Candlestick Understanding

Existing financial candlestick QA often measures answer accuracy alone, without checking whether predictions are supported by chart evidence. This work introduces code-interactive financial candlestick reasoning: the model generates and executes code to obtain observations and perform computations, then uses returned results to guide subsequent reasoning. It also builds 32K replay-verified trajectories and a 1,200-question benchmark that separately assesses answer correctness

01 ABSTRACT

The authors propose a code-interactive financial candlestick reasoning task and present Fin-Thinker, based on Qwen3.5-9B, which generates and executes code to obtain observations and perform calculations before continuing reasoning. To train this capability, they construct FinKLine-32K with 32,184 replay-verified code-observation trajectories and use three-stage training: foundational visual-code fine-tuning, financial process supervision, and execution-feedback reinforcement learning. On FinKLineBench, they report 13.1 percentage-point improvement in strict accuracy over the best direct-answer baseline and 33.2 points over its base model with sandbox access; however, all numbers come from the authors' own benchmark and self-reported experiments and need independent replication.

02 KEY FINDINGS

The task requires locating visual elements in candlestick charts, aligning them across trading dates and panels, and combining numerical and financial reasoning to answer.
Fin-Thinker uses code to obtain observations and computations; the FinKLine-32K training data includes successful solutions and error-recovery trajectories, emphasizing verifiable processes.
FinKLineBench contains 1,200 questions across 138 financial task types and separately evaluates answer correctness, code execution, and evidence grounding; ablations show successive gains from process supervision and execution-feedback reinforcement learning.

03 RELEVANCE

If financial chart QA only scores final answers, it can reward guessing or memorization rather than genuine use of chart evidence. This work uses code execution as a channel for evidence acquisition and computation and evaluates evidence grounding separately, helping move candlestick understanding from black-box answers toward inspectable reasoning processes, which is more relevant for high-stakes financial analysis.

04 EVIDENCE BOUNDARY

The reported improvements and conclusions are based on the authors' self-built FinKLineBench and self-reported experiments, with no independent third-party replication or out-of-sample validation across markets and time periods shown. The benchmark covers 138 financial task types, but data sources, market scope, time ranges, and annotation quality control details are not fully described in the abstract, so the evidence boundary is limited to this benchmark. Strict accuracy and grounded accuracy depend on exact answers and GPT-5.6 judging, respectively, and the latter may introduce evaluator bias. The model is based on Qwen3.5-9B and requires sandboxed code execution; deployment cost, execution safety, and suitability for low-latency scenarios are not discussed. These results do not constitute investment advice or proof of trading effectiveness.

Return to the primary source

AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG

Read original ↗