← Back to Blog Insights

Inside Engram: Navrisk's Approach to SME Risk Scoring

1 January 2025

Engram is a predictive risk scoring framework developed for assessing the credit risk of UK small and medium-sized enterprises (SMEs). Rather than relying solely on traditional financial ratios, the model combines financial performance with management behaviour to provide a broader view of business risk.

The framework consists of two complementary models:

Together, these models aim to provide lenders with a more accurate assessment of default risk over a 12-month horizon.

Purpose

SME lending starts from a common challenge: smaller businesses often have limited financial history and less publicly available information than larger corporates, making traditional credit assessment difficult. The objective of Engram is to improve risk assessment by combining multiple sources of information into a single, explainable risk score rather than relying on financial statements alone.

SME Focus

The model is designed specifically for UK SMEs. It defines its target population using company size, turnover and legal structure, while excluding businesses whose financial information is unsuitable for consistent modelling. This ensures that the resulting models are trained on businesses with similar operating characteristics rather than attempting to apply one methodology across all company sizes.

Data Collection and Preparation

Predictive models are only as good as the data used to build them. The underlying data included:

Significant effort was devoted to cleaning and validating the dataset. Companies with incomplete financial records or inconsistent accounting relationships were removed to reduce noise and improve model accuracy. Director information was extracted from unstructured text using natural language processing techniques, allowing variables such as appointments, resignations and management tenure to become usable model inputs.

Defining Default

A key challenge was determining what constitutes business failure. Rather than relying solely on whether a company became inactive, the model combines trading status with a measure of financial distress. This produces a more meaningful definition of default by distinguishing between businesses that close naturally and those that fail because of financial difficulties. True default events are relatively rare, creating an imbalanced dataset that must be addressed during model development.

Feature Engineering

Considerable emphasis was placed on selecting variables that are both statistically useful and economically meaningful.

For the financial model, the selected variables include:

These broadly represent liquidity, profitability, operating efficiency and solvency.

The management model focuses on variables such as:

Combining these factors provides a richer picture of business quality than financial metrics alone.

Model Development

The two models use different machine learning techniques.

Engram Financial

The financial model uses Logistic Regression, chosen because it offers a strong balance between predictive performance and interpretability. A transparent statistical approach was preferred over marginal performance gains from more complex algorithms.

Engram Management

The management model uses a Random Forest algorithm. Unlike logistic regression, Random Forest can capture complex, non-linear relationships within management data that would otherwise be missed. Although less inherently interpretable, it offers stronger predictive performance for this dataset. Explainability is later addressed using Shapley values.

Handling Imbalanced Data

Because only a small proportion of companies default, SMOTE (Synthetic Minority Oversampling Technique) is used to rebalance the training data. This creates synthetic examples of defaulted companies so that the models can learn patterns associated with failure without becoming biased towards the much larger population of healthy businesses. Resampling is only applied during training, ensuring testing remains representative of real-world performance.

Model Performance

Performance is measured primarily using the Receiver Operating Characteristic (ROC) score.

The Financial model achieves an in-sample ROC of 0.952 and an out-of-sample ROC of 0.953. The Management model achieves an in-sample ROC of 0.936 and an out-of-sample ROC of 0.942.

Historical validation across previous reporting periods also demonstrates relatively consistent performance, suggesting the models generalise well rather than simply fitting one dataset. Both models successfully distinguish between higher-risk and lower-risk companies.

Explainability

Recognising that lenders require transparency, Shapley Values are used to explain model outputs. Rather than presenting a “black box” prediction, the framework identifies which variables contributed most to a company’s risk score and whether each factor increased or reduced predicted risk. This makes the results easier for credit analysts to understand, justify and communicate during lending decisions.

Risk Bands

Rather than presenting raw probabilities, Engram converts predictions into practical risk categories:

The cut-off points are determined mathematically using the distribution of active and defaulted companies, creating thresholds that separate lower-risk from higher-risk borrowers as effectively as possible. This makes the output suitable for operational lending workflows and portfolio monitoring.

Overall Assessment

Engram is a data-driven credit risk framework designed specifically for UK SME lending. Its main contribution is combining traditional financial analysis with management-related indicators to improve predictions of business default. The methodology follows recognised statistical and machine learning practices, including careful data preparation, feature engineering, validation and explainability.

While the Financial model prioritises transparency through logistic regression, the Management model uses Random Forest to capture more complex behavioural signals. Together, they provide lenders with a structured, explainable risk score that can support underwriting, portfolio monitoring and early identification of deteriorating businesses — a thoughtful balance between predictive accuracy, practical usability and regulatory explainability.