Abstract Research overview
This study examines the potential and limitations of automated risk analysis using open municipal payment data. A seven-module Python system was applied to 5.4 million transactions from the City of Gothenburg, then selected cases were verified against public records. Statistical flags predominantly reflected legitimate payment patterns. A complementary semantic screen led to a case with 13 documented control deficiencies. The findings highlight the importance of contextual review and the limits imposed by missing invoice-approval data.
Keywords Open data · Anomaly detection · Municipal finance · Internal control
Can open data support independent scrutiny?
Gothenburg publishes its supplier payments as open data. For my bachelor's thesis in Industrial Engineering and Management, I built a Python system to examine all 5,433,668 published transactions from 2021–2024: 18,774 suppliers and SEK 107.6 billion.
The study follows a Design Science approach: build an analytical system, then test its practical value by checking flagged cases against contracts, original invoices and approval records obtained through requests for public documents.
From 48 CSV files to seven analysis modules
Prepare the data. Harmonize mixed encodings, column names, dates and text-formatted amounts. Keep all transactions, including credit invoices, and validate supplier identifiers.
Look for patterns. Profile suppliers using variation in amounts, concentration near procurement thresholds and possible split purchases. Complement the scoring model with identity checks, Benford analysis and tools for manual review.
Evaluate and iterate. Nine versions reduced high-risk flags from 1,516 to 31. That is a reduction in review volume, not a demonstrated improvement in fraud detection. The ten highest-ranked suppliers examined in the detailed results all had legitimate explanations, including fixed care rates, framework agreements and rent.
Examine the context. A complementary semantic screen compared supplier activity with the service indicated by the account description. It identified 16 associations billing administrative services. Interpreting those combinations required human review.
Follow the signal all the way to the documents
Fourteen requests for public documents supported the investigation. Of six semantically flagged cases taken to external verification, four had legitimate explanations, one showed a price anomaly without demonstrated irregularity, and one contained 13 documented control deficiencies.
4 explained cases
Contracts and invoices explained the flagged patterns. Broad account descriptions had obscured legitimate services.
1 price anomaly
A markedly higher hourly rate warranted attention, but the documentation did not establish an irregularity.
1 case with deficiencies
Document review established gaps in invoice requirements, approval procedures and record keeping.
A small transaction opened a larger investigation
A payment of SEK 80,000 led to the review of an association that had received SEK 2.16 million in municipal funding. The study documented 13 control deficiencies through invoices, contracts, written responses and bankruptcy records. The statistical model had assigned the association a low score.
An independent whistleblower investigation at the same administration described similar patterns in other cases. It supported the comparison of control weaknesses; it did not investigate or individually confirm this study's 13 findings.
The findings concern documented control deficiencies. The study does not establish whether a crime occurred.
The missing fields matter as much as the model
Regular payments often reflect legitimate business arrangements. A meaningful review also needs context: what was purchased, what the contract required and who handled and approved the invoice.
The thesis proposes pseudonymised invoice-approval data to make patterns such as the same person controlling multiple steps visible to external reviewers. This is a proposal for further development; the study did not implement or validate an approval-analysis system.
The results come from one municipality and a small, deliberately selected set of cases. The semantic screen's value is supported by one case with documented deficiencies, which calls for further testing across municipalities.