Journal Entry Tests (JETs) Overview

• Journal Entry Tests (JETs) are formalized, rule-based procedures applying finite audit rules to every ledger entry to maximize anomaly detection recall.
• JETs employ threshold, ratio, round-number, and temporal-pattern tests to flag suspicious entries while maintaining strict audit protocols.
• Recent enhancements integrate federated, privacy-preserving frameworks and AI-augmented methods, significantly improving detection precision and lowering false positives.

Journal Entry Tests (JETs) are formalized, full-population, rule-based procedures for detecting suspicious activity in double-entry bookkeeping records, especially within tax-related general ledgers. Each test applies a finite set of audit rules to every journal entry, producing flags for entries that deviate from expected patterns. As foundational elements of financial auditing, JETs are designed to maximize anomaly detection recall but are also associated with high false-positive rates and limited capacity to identify subtle or context-sensitive irregularities. Recent research explores more advanced, privacy-preserving, and AI-augmented frameworks to enhance both detection accuracy and operational efficiency while maintaining strict data confidentiality (Mashiko et al., 22 Jan 2025, Kadir et al., 2 Dec 2025).

1. Formal Structure and Mechanisms of Journal Entry Tests

A Journal Entry Test operates over a ledger $\mathcal{L} = \{x_1, x_2, \dots\}$, where each journal entry $x$ encompasses numeric (e.g., amounts), categorical (e.g., account codes), temporal (e.g., dates), and textual (e.g., descriptions) features (Kadir et al., 2 Dec 2025). The core structure consists of a finite set of indicator rules $\{R_j\}_{j=1}^m$, each implemented as a function that evaluates whether $x$ meets or violates a specified criterion.

Rules in standard JET architectures include:

• Threshold checks: Detect large amounts (e.g., $\mathrm{amt}(x) > \theta_{\mathrm{amt}}$) or tax-rate anomalies.
• Ratio checks: Compare entry amounts to distributional statistics (e.g., $$\frac{\mathrm{amt}(x)}{\mu_{\mathrm{amt}}} > \theta_{\mathrm{ratio}}$$).
• Round-number tests: Flag entries with suspiciously “round” amounts.
• Temporal-pattern tests: Identify back-dating or off-period postings.

Mathematically, each rule yields a binary anomaly indicator $s_j(x) \in \{0, 1\}$. Two principal combination schemes are used:

• Logical OR: $JET(x) = 1$ if $\exists j: s_j(x) = 1$; otherwise $0$. This prioritizes recall.
• Weighted voting: $S(x) = \sum_j w_j s_j(x)$, with a flag threshold $k$.

JETs are applied to every ledger entry before auditor review.

2. Performance Metrics and Baseline Effectiveness

JETs are typically assessed using precision ($P$), recall ($R$), F1 score, and confusion matrix entries (TP, FP, FN, TN) (Kadir et al., 2 Dec 2025). For example, on synthetic datasets with known anomalies:

Method	Precision ($P$)	Recall ($R$)	$F_1$	False Positives (FP)
JET (rule-based)	0.53	0.90	0.50	942

JETs deliver high recall but low precision; a substantial fraction of flagged entries are false positives (e.g., $\approx 20\%$ flagged with only $1\%$ being true anomalies). This imposes significant manual triage overhead on auditors (Kadir et al., 2 Dec 2025).

3. Limitations of Rule-Based JETs

The canonical weaknesses of JETs include:

• High false-positive rates: Rigid thresholds capture many innocuous but statistically rare entries.
• Limited contextual awareness: JETs are unable to detect distributed, structured, or context-adjustable anomalies (e.g., “structuring” via multiple small credits).
• Inflexibility for novel or evolving fraud patterns: Static rule-sets cannot learn from historical context or adapt to emergent threats.
• Auditor workload: The bulk of flagged entries require manual inspection, reducing efficiency.

These limitations motivate the fusion of JETs with advanced statistical and machine learning–based approaches (Mashiko et al., 22 Jan 2025, Kadir et al., 2 Dec 2025).

4. Enhancement via Federated, Non–Model-Sharing Anomaly Detection

Recent advances incorporate federated, non–model-sharing frameworks—specifically Data Collaboration (DC) analysis—to address both performance and confidentiality issues (Mashiko et al., 22 Jan 2025). The DC protocol proceeds as follows:

1. Dimensionality Reduction: Each organization $i$ reduces its private data $X_i$ with a mapping $\phi_i$ (e.g., PCA), yielding $H_i = \phi_i(X_i)$.
2. Anchor Embedding: Each $\phi_i$ is also applied to a common “anchor” dataset $X_{\rm anc}$, yielding $H_i^{\rm anc}$.
3. Collaboration Space Alignment: Transformations $G_i$ align each $H_i^{\rm anc}$ to a shared $Z$ via minimization:

$$\min_{Z,\{G_i\}} \sum_{i=1}^c \|H_i^{\rm anc} G_i - Z\|_F^2$$

with a closed-form solution via SVD.

1. Unified Representation for Model Training: Final representations $Z_i = H_i G_i$ are aggregated for a combined autoencoder detector.

This framework outperforms both single-organization baselines and iterative model-sharing FL (FedAvg), especially in non-i.i.d. data settings—mirroring real-world inter-firm heterogeneity. Communication is minimized to a single round, raw data are never exposed, and local data transforms are privately computed and only used post-collaboration (Mashiko et al., 22 Jan 2025).

5. AI-Augmented Approaches: LLMs and Hybrid Workflows

AI-augmented systems such as AuditCopilot introduce LLMs as flexible, context-aware anomaly detectors (Kadir et al., 2 Dec 2025). This workflow is characterized by:

• Prompt-based audit reasoning: No model fine-tuning; raw entry data and context statistics are passed as JSON-like prompts, with LLMs guided by system-defined audit criteria.
• Context-rich inputs: Combining structured features, global summary statistics (mean, percentiles), and classical anomaly scores (e.g., Isolation Forest hints).
• Scored outputs: LLMs deliver a binary anomaly flag, a confidence score, and a natural-language explanation per entry.

In comparative benchmarks, LLMs surpass traditional JETs and classical ML in both precision and $F_1$:

Method	Precision ($P$)	Recall ($R$)	$F_1$	False Positives (FP)
Gemma-7B (AuditCopilot)	0.71	0.99	0.79	68
Mistral-8B (AuditCopilot)	0.90	0.98	0.94	12
Isolation Forest	0.61	0.98	0.68	169
JET (Traditional)	0.53	0.90	0.50	942

Moreover, interpretability is enhanced through concise rationales ("Amount lies above the 99th percentile and transaction posted outside normal hours"), supporting auditor trust and review (Kadir et al., 2 Dec 2025).

6. Practical Deployment and Comparative Considerations

For real-world integration, these enhanced JET frameworks require:

• Anchor data and parameter selection: Sustainable agreement on anchors $X_{\rm anc}$, mapping $\phi_i$, and collaboration dimensionality (Mashiko et al., 22 Jan 2025).
• Secure communications: Only intermediate embeddings and anchor projections are transmitted; raw data remain local.
• Calibration of thresholds: The anomaly cutoff $\tau$ is derived empirically, balancing recall and precision with auditor tolerance for false positives.
• Regulatory alignment: All data-handling and communications must meet sectoral data retention and encryption requirements.

A typical pipeline combines classic JETs for broad, easily interpretable rule-coverage, with AI/ML models handling cases requiring context, subtlety, or adaptive logic. Human auditors remain the ultimate decision-makers, using both quantitative flags and LLM explanations as decision-support (Mashiko et al., 22 Jan 2025, Kadir et al., 2 Dec 2025).

7. Research Frontier and Ongoing Challenges

Open issues in the evolution of JETs include:

• Information loss due to dimensionality reduction: Utilizing PCA on sparse, one-hot encoded data may filter relevant fine-grained features; the exploration of autoencoder-based dimension reduction and more representative anchor data is ongoing (Mashiko et al., 22 Jan 2025).
• Robust thresholding and explainability: Determining practical $\tau$ values requires careful empirical calibration. The interpretation of LLM outputs remains subject to prompt design and audit context (Kadir et al., 2 Dec 2025).
• Hybrid workflow optimization: The combination of rule-based, federated, and LLM-based detection presents trade-offs among detection accuracy, interpretability, and operational burden.
• Privacy versus utility: Balancing the utility of collaborative detection with stringent confidentiality requirements continues to motivate advances in federated and privacy-preserving computation (Mashiko et al., 22 Jan 2025).

These emerging paradigms position JETs not only as rule-based compliance screens but as components within collaborative, AI-augmented audit systems, enhancing cross-firm anomaly detection while rigorously protecting sensitive financial information.