Rahul Rao Juvvadi’s SAP S/4HANA research proposes a practical way to help finance and audit teams identify the journal entries that deserve attention first
For large companies, closing the books is a routine event with unusually high consequences. In the final days of a reporting period, finance teams may post accruals, reclassifications, corrections and other manual journal entries under intense time pressure. Most of those entries are legitimate. Some contain mistakes. A small number may be used to conceal manipulation. The difficulty is not understanding that risk exists. It is deciding where to look when an enterprise accounting system may contain millions of transactions.
That is the problem Rahul Rao Juvvadi set out to examine in his 2022 paper, “Machine Learning for Anomaly Detection in the Financial Close: A Journal Entry Risk-Scoring Framework for SAP S/4HANA,” published in the International Journal of Communication Networks and Information Security. Juvvadi’s proposal is straightforward in purpose: assign a risk score to individual journal entries and use that score to help controllers and auditors decide which transactions should be reviewed first.
The framework does not attempt to replace professional judgment. It does not approve, reject, reverse or clear an accounting entry. It is designed to organize attention.
Traditional methods are useful but limited. Fixed rules can identify obvious exceptions, such as unusually large entries, weekend postings or transactions created by users with conflicting access. Statistical sampling can reduce an enormous population to a manageable set. But neither approach is well suited to identifying risk that emerges from several factors acting together.
A transaction may appear ordinary when its amount, timing or account combination is considered separately. It may look different when all three are viewed together, particularly when the entry is posted near period-end, involves an unusual pairing of general-ledger accounts and is later reversed. That is the kind of pattern Juvvadi’s framework is intended to surface.
The research is built specifically for SAP S/4HANA, the enterprise resource planning system used by many large organizations. It draws from ACDOCA, commonly known as the Universal Journal, where financial and management-accounting line items are stored in a shared structure. From that data, the framework evaluates four broad categories of risk: unusual posting behavior, uncommon account combinations, reversal activity and round-amount transactions.
This system-level focus is one of the study’s strengths. Rather than discussing artificial intelligence in general terms, the paper identifies the transaction fields needed for analysis and explains how a risk score could be returned to the financial-close workflow. That makes the work relevant to the people who would have to use it: finance leaders, internal auditors, external auditors and enterprise-technology teams.
The model combines three forms of machine learning. Two are designed to find transactions that look unusual even when there are few past examples to learn from. The third learns from journal entries that previous audits or control reviews had already identified as suspicious, non-compliant or erroneous. Their outputs are then combined into one risk probability for each entry.
That combination addresses a practical weakness in historical audit data. Auditors do not examine every transaction. An entry that was never selected for review cannot automatically be treated as a confirmed clean entry. By combining anomaly detection with learning from past findings, the framework is intended to identify both familiar risk patterns and unusual activity that does not closely resemble earlier cases.
Juvvadi evaluated the approach using approximately 2.5 million journal-entry line items from one company code within a multinational manufacturer’s SAP S/4HANA environment. The data covered twelve fiscal periods and was divided chronologically so that later periods were used to test a model trained on earlier activity. That choice matters because posting behavior can change over time, and a random split may create a less realistic picture of future performance.
On the held-out test data, according to the paper, the combined model reported precision of 0.94, recall of 0.91, an F1 score of 0.92 and a ROC-AUC of 0.97. The paper also states that the combined model outperformed each of the individual models tested in the study.
The paper is careful not to present those figures as universal. The pilot involved one company code, and the labels came from earlier audit findings and internal-control reviews. A different organization may use another chart of accounts, different document types and different close procedures. The reported results therefore show what the framework achieved in the study population, not what every company should expect automatically.
The research is equally candid about the burden of false positives. Even if only five percent of all entries were flagged for review, that would still amount to roughly 125,000 alerts across 2.5 million entries — far more than any close team could investigate within a normal reporting cycle.
Juvvadi’s answer is to treat the model as a ranking tool rather than a pass-or-fail system. A company could review the highest-scored entries that fit the team’s actual capacity. In the study, the top one percent of entries produced precision of 0.81 against the available audit-derived labels. In practical terms, the value lies in improving the order in which limited review time is used.
“The model never decides whether an entry is right or wrong,” Juvvadi said. “It decides what a person should look at first. The aim isn’t to automate judgment — it’s to make sure the limited hours an audit team has are spent where the risk actually is, and that we can always explain why an entry ended up in front of them.”
The distinction is important. A high score is not a finding of fraud. It is a reason to look more closely. A controller or auditor would still examine the supporting documents, authorization history, accounting purpose and business context before reaching a conclusion.
The framework also addresses a question that often determines whether machine learning can be trusted in a control environment: why did the model flag this transaction? An unexplained probability is of limited use to an auditor. Juvvadi’s design includes an explanation layer that identifies both the part of the combined model that influenced the result and the transaction characteristics that contributed most to the score.
A reviewer could therefore see that an entry was elevated because of an unusual account pairing, its timing near period-end, a reversal pattern or a combination of factors. That does not make the decision automatic, but it gives the reviewer a clearer starting point and creates a more defensible record of why the entry entered the review queue.
The paper goes beyond prediction accuracy to consider governance. Once a model begins influencing financial-reporting review, the model itself becomes part of the control environment. Its data, source code, access rights, model versions and decision thresholds cannot be allowed to change without oversight.
Juvvadi proposes controls around those risks. Changes to the model would require documented approval. Access would be restricted and logged. Failed scoring or retraining jobs would be monitored. Reviewer overrides would be recorded with the identity of the reviewer, the time of the decision and the reason for the override. New model versions would be tested before replacing the version already in use, and prior versions would be retained for rollback.
This governance component is significant because enterprise finance teams are adopting machine learning faster than many control frameworks are evolving. A model may improve one review activity while creating a new weakness if nobody can explain how it changed, who approved it or whether it is still performing as expected. Juvvadi’s framework treats accountability as part of the technical design rather than as an afterthought.
The research also recognizes that past audit findings contain their own bias. Auditors find what they choose to examine, and those choices are influenced by existing risk assumptions. A model trained only on past findings may therefore repeat earlier audit habits instead of revealing the full population of risky activity.
To reduce that risk, Juvvadi recommends periodically selecting a blind sample of low-scored entries for substantive review and adding the results back into the training data. This would test whether the model is overlooking categories of transactions simply because similar entries were not examined in the past.
Taken together, the work makes a broader contribution to the discussion of how machine learning should be used in enterprise finance. It connects SAP transaction data, audit judgment, anomaly detection, explainability, review capacity and model governance in one operating framework. The proposed result is not an automated accounting decision. It is a more focused and more transparent review process.
The immediate application is SAP S/4HANA, but the underlying problem is not limited to one platform. Large financial systems routinely produce transaction volumes that exceed the capacity of manual review. Adapting the approach to another environment would require different field mappings, local validation and controls suited to that organization, but the central idea remains the same: use technology to help professionals decide where their attention is most needed.
Further testing will be required before the framework can be considered broadly validated. Independent studies across multiple companies, industries and system configurations would help establish how well the results generalize. Evidence from sustained production use would also show whether the model remains effective as business processes, personnel and posting behavior change.
Even with those limitations, Juvvadi’s research addresses a consequential problem in modern financial reporting. Audit and finance teams cannot examine every entry, but they remain responsible for identifying the transactions most likely to contain error or manipulation. His proposal does not remove that responsibility. It gives those teams a more disciplined way to decide where the review should begin.
