Master’s thesis · TU Dortmund with Barkow Consulting
Contextual Anomaly Detection in Cash Access
A machine-learning framework that scores each German region against what its own context predicts, then attributes each deviation to the regional factors behind it. It joins geospatial, socioeconomic and financial data and finds gaps a national threshold would miss.
Unusual compared to what?
A rural region with few ATMs may be adequately served, while a city with many may not. The count alone does not settle it.
The thesis reframed this into two questions: which regions are unusual compared with socioeconomically similar regions, and which factors most influence how cash-access resources are distributed. There was no ground truth for the first question — no reference dataset marks a region as over-served or under-served.
Building the dataset
A complete, centralised view of cash-access locations across Germany did not exist, so the dataset came first. I scraped 75,000+ locations spread across 10+ bank locator websites and cleaned them into one verified source-of-truth database. It is the first consolidated database of Germany’s cash-access network.
Context was defined through eight socioeconomic groups, among them household and sectoral economics, unemployment, taxation, workforce and demographics. Those groups form the comparison: a region is judged against places with a similar profile rather than against the national average.
Choosing a method
I researched and implemented several contextual anomaly detection frameworks, comparing them on German data and 16 benchmark datasets.
Interpretability was the deciding criterion: the results had to make sense to business stakeholders, so a framework that flagged a region without saying why was not usable.
What the model flags, and why
Raw counts favour big cities. Adjusted for context, the picture moves in both directions: some rural regions are well served, and parts of Thüringen, Brandenburg and Rheinland-Pfalz show possible under-provision.
For each flagged region, the output names the regional characteristics that push it away from its own expected level of access. SHAP values and feature importance rank the contextual factors behind the deviation.
That gave the business an explainable way to prioritise which regions to investigate.
Limits and next steps
The analysis is cross-sectional: it describes the structure of access, not how it changes over time. It also depends on the granularity and freshness of the published regional data, which limits how local the conclusions can be.
The database is built from bank locator websites, so its completeness is bounded by what those lists publish.
A natural continuation would add a temporal dimension to track whether gaps widen or close, and test the framework against other definitions of access, such as travel time.