[Paper Review] Financial Transaction Retrieval and Contextual Evidence for Knowledge-Grounded Reasoning

Artem D. Sakhno, Daniil Tomilov+6 more|arXiv (Cornell University)|Mar 16, 2026

Financial Distress and Bankruptcy Prediction0 citations

TL;DR

FinTRACE introduces a retrieval-first framework that converts raw financial transactions into a structured knowledge base of feature essences, behavioral patterns, and white-box rules to ground LLM reasoning, achieving strong zero-shot and few-shot results and enabling knowledge-grounded instruction tuning.

ABSTRACT

Nowadays, success of financial organizations heavily depends on their ability to process digital traces generated by their clients, e.g., transaction histories, gathered from various sources to improve user modeling pipelines. As general-purpose LLMs struggle with time-distributed tabular data, production stacks still depend on specialized tabular and sequence models with limited transferability and need for labeled data. To address this, we introduce FinTRACE, a retrieval-first architecture that converts raw transactions into reusable feature representations, applies rule-based detectors, and stores the resulting signals in a behavioral knowledge base with graded associations to the objectives of downstream tasks. Across public and industrial benchmarks, FinTRACE substantially improves low-supervision transaction analytics, doubling zero-shot MCC on churn prediction performance from 0.19 to 0.38 and improving 16-shot MCC from 0.25 to 0.40. We further use FinTRACE to ground LLMs via instruction tuning on retrieved behavioral patterns, achieving state-of-the-art LLM results on transaction analytics problems.

Motivation & Objective

• Address the challenge of applying LLMs to time-distributed, heterogeneous financial transactions by creating a reusable, grounded knowledge base of behavioral evidence.
• Develop a retrieval-first pipeline that transforms raw transaction logs into feature essences, behavioral patterns, and task targets, linked by transparent rules.
• Enable LLM grounding and instruction tuning using the KB to improve low-supervision transaction analytics across public and industrial benchmarks.

Proposed method

• Transform raw transaction histories into a structured knowledge base with three semantic layers: feature essences, behavioral patterns, and downstream targets.
• Use explicit white-box rules (AutoWoE-based) to connect essences to patterns and patterns to task outcomes, forming traceable evidence chains.
• Construct KB-grounded prompts for LLMs to produce predictions conditioned on grounded evidence, replacing row-wise serialization.
• Perform KB-grounded instruction tuning by converting KB instances into grounded reasoning triplets for supervision.
• Support zero-shot and few-shot adaptation by retrieving relevant behavioral patterns without target data during KB construction and by inference-time adaptation (self-reflection) with limited labeled examples.
• Evaluate with zero-shot and few-shot LLMs (GPT-based and Llama-based) on public and proprietary transaction datasets.

Figure 1. Generalization gap between LLM and specialized pipelines. In few-shot settings (0/4/16 labels), LLM methods (GPT-OSS, Llama-3-Instruct, Reasoning Bank) plateau below 0.30 MCC, while specialized models (TabPFN, CatBoost, LLM4ES, CoLES) exceed 0.48 MCC only with full supervision. The shaded

Experimental results

Research questions

• RQ1 Can a structured, retrievable knowledge base of financial behavioral evidence improve LLM-grounded reasoning on transactional analytics under low supervision?
• RQ2 Does KB-grounded instruction tuning yield state-of-the-art results on open financial benchmarks while preserving general language capabilities?
• RQ3 How does zero-shot retrieval-based grounding compare to traditional prompt-only or fully supervised tabular models on churn and demographic tasks?
• RQ4 Can inference-time, self-reflection mechanisms support few-shot adaptation to unseen financial tasks without parameter updates?

Key findings

• FinTRACE doubles zero-shot MCC on Rosbank from 0.19 to 0.38, demonstrating strong grounding over prompt-only methods.
• In 16-shot settings, FinTRACE reaches MCC of 0.40 on Rosbank and 0.77 F1 on DataFusion, outperforming strong baselines.
• Knowledge-grounded instruction tuning with FinTRACE achieves MCC of 0.48 on Rosbank and 0.53 on Gender while preserving text capabilities, matching or surpassing task-specific models on some metrics.
• Zero-shot FinTRACE on a proprietary dataset yields the highest MCC (0.10), outperforming base GPT-OSS and TabLLM baselines under extreme label scarcity.
• Ablation results show LLM-guided behavioral pattern selection and white-box knowledge are key drivers of performance gains.
• Context constructed via white-box KB significantly outperforms simple distributional summaries or feature importance augmentations.

Figure 2. FinTRACE overview. Raw transaction logs (a) are transformed into a structured knowledge base (b) with three semantic layers: feature essences, behavioral patterns, and downstream targets. Explicit white-box rules connect these layers (c), allowing the LLM to produce grounded predictions su