using machine learning to segment UK mortgages – Bank Underground

Joe Grimshaw

Who are the UK’s mortgage borrowers, and how do their characteristics differ? Despite extensive literature on mortgage profiles, loan-level segmentation remains limited, existing work relies on aggregates or predefined categories. I address this gap by applying unsupervised machine learning to 20 years of data, allowing the model determine segments without prior assumptions. Three clusters emerge: one with low leverage, and two with high leverage but notably different income profiles. Lending composition has shifted gradually. High leverage, high-income borrowers now account for a larger market share, and first-time buyers increasingly fall into more leveraged segments. Machine learning is crucial for financial stability, revealing concentrations of characteristics, and trends, that aggregates and simple splits cannot, offering richer and earlier indications of potential vulnerabilities.

How can we sort UK mortgage borrowers into meaningful groups?

To understand how borrower characteristics are distributed across UK owner-occupier mortgages, I apply k-means clustering to the FCA’s Product Sales Database, a loan-level data set covering all mortgages issued between 2005 and 2025. The algorithm processes 10 variables in total: loan to value (LTV), loan to income (LTI), gross income, loan value, property value, interest rate, mortgage term, borrower age, debt-servicing ratio (DSR), and net income. The groupings are then characterised using three core features: LTV, LTI, and term length. In my diagnostics, these show the strongest influence on how observations separate into distinct groups, and are often recognised as key determinants of household vulnerability.

K-means is an unsupervised machine learning algorithm that groups loans so those within a cluster are more similar to each other than to those in other clusters. Think of it as an algorithm that looks for natural groupings in the data, rather than being told in advance what those groupings should look like. It iterates, reassigning loans to their nearest cluster centre and recalculates centres until groupings stabilise. All inputs are standardised before fitting to ensure fair comparisons across different scales, and outliers are excluded by removing observations above the 99.9th percentile for each variable.

I test alternatives. DBSCAN identifies clusters based on the density of nearby observations, making it well suited to irregular shapes but sensitive to parameter choices. Hierarchical clustering builds a tree of nested groupings, useful for visualising structure but computationally demanding at this data scale. K-means proves the most robust and interpretable for this task, with stable allocations across reruns and clear separation that can be communicated to both technical and non-technical audiences. The number of clusters is guided by the need for tractable interpretation, and the usage of two diagnostics: the elbow method, which identifies where adding more clusters yields diminishing improvements, and silhouette analysis, which checks how cleanly each loan fits its assigned group relative to the others as well as. All point to three clusters as the natural solution.

Three distinct borrower segments emerge from the data

Chart 1 shows the cluster centre for each group across the three core features (LTI, LTV, mortgage term) and gross income.

I assign cluster labels after estimation, based on the three core metrics. The cluster with the lowest average is labelled Group A; the highest, labelled Group C; and the other, labelled Group B. While these labels are determined mechanically, the resulting groups exhibit clear and stable profiles. Group A consistently corresponds to low‑leverage lending, Group B combines relatively high-leverage with higher-incomes, and Group C captures high‑leverage borrowing concentrated among lower‑income borrowers.

Group A (Low Leverage): This segment is defined by more conservative borrowing and accounts for about a third of the flow of mortgage lending. The median LTV is approximately 39%, the median LTI is 1.9, and the median mortgage term is 15 years. The share of lending at higher thresholds, such as loans exceeding 90% LTV or 4.5 times income, is negligible. Borrowers in this group are typically older, with higher incomes and substantial deposits, and are less likely to be first-time buyers.

Group B (High Leverage, High Income): Between 10% and 15% of new lending each quarter falls into this segment. Mortgages in this group have a median LTV of around 67%, a median LTI of 3.2, and a median term of 23 years. Around 7.5% of loans exceed 90% LTV, and nearly 11% are above 4.5 times income. This cluster mainly consists of more affluent borrowers accessing higher-value properties, leveraging their income and extending terms to do so.

Group C (High Leverage, Low Income): Over half of new lending falls into this segment each quarter, displaying a median LTV of 80%, a median LTI of 3.4, and a median term of 28 years. Around 13% of loans are above 90% LTV, and about 10% exceed 4.5 times income. While both groups have a similar proportion of high LTI loans and comparable average LTIs, this group has higher LTV and term lengths. Borrowers are younger, with lower incomes and smaller deposits, and the group includes a significant share of first-time buyers.

For three of the four metrics in Chart 1, cluster centres follow a sequential order from Group A to Group C. Income is the exception. Group B has the highest median income, followed by Group A, with Group C lowest. Group B borrowers have the financial capacity to service larger loans but take on more leverage to access higher-value properties. Consider a dual-income couple in London buying a £700,000 flat with a £550,000 mortgage. Group C borrowers take similar leverage but with lower incomes, like a single first-time buyer purchasing a £200,000 home on a modest salary. Both groups are highly leveraged, but their financial profiles differ markedly.

Chart 1: Average characteristics by cluster (2022–25)

Notes: Chart shows median values for each cluster. LTV and LTI are expressed as ratios. Gross income is in £s. Mortgage term is in years.

How has the composition of lending changed over time?

Group C has consistently represented the largest share of new lending, while Group B has been the smallest. However, Group B has grown in prominence, increasing its share from around 7% in the mid-2000s to 11% in recent years (Chart 2), resulting in a slow reduction in the market share of Group A over time. This gradual shift reflects changes in both borrower behaviour and market conditions.

Chart 2: Share of completions over time

Both Groups B and C, now include a greater share of first-time buyers than before the financial crisis (Chart 3). Group C in particular has seen its first-time buyer share grow over the past 10 to 15 years. This pattern is closely linked to the increase in house price to income ratios. As affordability pressures have mounted, first-time buyers have increasingly needed to take on greater leverage and mortgage terms to access the market.

Chart 3: First-time buyer share over time by cluster

Mortgage terms have lengthened across all segments (Chart 4). Since 2015, the median term has increased by two years in each group. Compared to the mid-2000s, the increase is even more pronounced, up to five years longer in Groups B and C. Longer terms allow borrowers to spread repayments, reducing monthly outgoings and easing affordability pressures. However, they also mean higher total interest costs and longer exposure to market fluctuations.

Chart 4: Average mortgage term length over time by cluster

Regional differences reveal distinct borrower compositions

London and the South East have the lowest share of Group C borrowers (Chart 5), but also have the lowest share of Group A borrowers. Why? Because, a material share of lending in these regions falls into Group B, the high leverage, high income segment, which is much smaller elsewhere. Higher property prices mean borrowers often need both high incomes and large loans relative to their earnings, with longer mortgage terms common to manage repayments. Group B’s prominence in London and the South East is not a recent development but a longstanding feature of the UK mortgage landscape.

Elsewhere, the picture is more mixed. The North of England, Northern Ireland, and Wales have a greater share of Group A lending, reflecting lower house prices, LTVs, and LTIs. The Midlands, Scotland, and the South West show more Group C lending. A region may appear unremarkable in average ratios, yet mask a very different compositional story beneath. Rather than summarising borrowers by a single average, loan-level segmentation reveals the distinct groups driving it.

Chart 5: 2025 distribution of clusters within each region

What does this mean for financial stability?

Understanding who is borrowing, where, and how much, is central to assessing mortgage market patterns. This approach complements existing frameworks by letting borrower segments emerge from the data rather than imposing predefined categories. Simple splits by income or region can indicate that leverage is rising; they cannot tell you that higher leverage and income are increasingly clustering together, or that first-time buyers are concentrating in more stretched segments, or that regions with similar average ratios hold very different borrower mixes beneath. These combinations matter. Surfacing them early, before they appear in aggregate statistics, strengthens the toolkit and data available for safeguarding financial stability in one of the UK’s most systemically important markets.

Joe Grimshaw works in the Bank’s Macro-Financial Risks Division.

If you want to get in touch, please email us at bankunderground@bankofengland.co.uk or leave a comment below.

Comments will only appear once approved by a moderator, and are only published where a full name is supplied. Bank Underground is a blog for Bank of England staff to share views that challenge – or support – prevailing policy orthodoxies. The views expressed here are those of the authors, and are not necessarily those of the Bank of England, or its policy committees.