How we built LendScore: Turning cash flow data into credit risk insights

Brett Manning
Product Manager

Understanding the challenge: Cash flow underwriting is complex, but doesn’t have to be hard 

Credit underwriting has long relied on credit bureau data, which provides many years of structured, standardized information on loans, repayments, and delinquencies. But these files have blind spots, saying little about individuals without significant credit exposure and missing early warning signals that might be seen in other financial behaviors. 

Bank transaction data, or cash flow data, tells a richer story. It reflects how consumers earn, spend, save, and manage money day to day. It’s dynamic, real-time, and much broader financial behaviors. When used alongside bureau data, cash flow data can enable lenders to spot hidden risk, identify missed opportunities, and more accurately price loans.

However, cash flow data introduces new complexities:

1. Unstructured data: Transactions come as text strings and amounts that lack consistent formatting across institutions or even across time.

2. High variability: Incomes fluctuate, discretionary spending shifts seasonally, and behaviors evolve.

3. Regulatory transparency: Cash flow data exposes lenders to Fair Lending and explainability concerns, which require careful treatment to avoid regulatory pitfalls. 

Building a credit risk model on such data means solving problems across data processing, feature generation, and decisioning—all while meeting compliance and interpretability expectations. 

And that’s where LendScore comes in. We set out to create a solution that distills complex data to a single market-ready score to improve credit risk assessment when used alongside traditional data.

Let’s dive into how we built our score and what makes it unique from other cash flow-based scores on the market.

Step 1: Processing raw data: Structuring the cash flow signal

At the foundation of LendScore is Plaid’s consumer-permission transaction data—billions of records representing consumers’ inflows, outflows, and balances over time. The first challenge was to transform this unstructured data into a stable analytical base. Plaid’s AI-powered transaction categorization models interpret raw text like:

“ACH CREDIT JPM PAYROLL” → Income
“STARBUCKS 02412 NY” → Discretionary Spending

Each transaction is labeled into high-level categories such as “Credit Card Repayments” or “Home Improvement Spending.” Then we aggregate these into higher-level categories like essential/discretionary spending or credit repayment, and also cluster transaction streams to separate recurring and one-off expenses.

Once categorized, we aggregate across time windows, e.g., 30, 60, and 90 days, to create features such as:

• Average monthly inflows

• Standard deviation of depository balances

• Share of spending on discretionary vs. essential categories

• Ratio of income to total outflows

This aggregation turns unorganized transaction streams into structured, behavioral signals suitable for modeling.

From these structured signals, we engineered hundreds of attributes and chose the top 145 most predictive to include in our score to balance the trade off between model lift and complexity, reducing any concerns about explainability. About 81% of the predictive power comes from cash flow features, while the remaining 19% originate from Plaid’s Network Insights.

Cash flow attributes

Cash flow features describe how money moves:

• Credit balance volatility: How variable are credit card balances month-to-month?

• Depository stability: Are balances consistently maintained above a threshold?

• Income regularity: Are inflows predictable and recurring?

• Loan payment behavior: How consistent are repayments and payment amounts?

Plaid Network Insights

Plaid’s network data offers a unique layer of predictive power. Because 1 in 2 US consumers connect their financial accounts through Plaid (nearly 1 million connections daily), we can observe network-level patterns that correlate with repayment outcomes—while keeping track of consents to ensure that consumers understand how their data is being used.

Network Insights includes attributes such as:

• The number of connections to certain types of financial apps (e.g., lenders, savings tools, earned wage access products), e.g., more lending accounts connections increase risk.

• Patterns of connection that distinguish between recent and historic interactions and differentiate stable relationships vs. risk seeking behaviors, e.g., recent lending connections are riskier than long standing ones. 

These signals are Plaid-only features that come from our vast network of 12,000 financial institutions and 7,000 apps, and are not available through traditional credit bureaus or other cash flow underwriting data providers. They help differentiate risk among consumers who look identical based on traditional credit or cash flow data.

Step 3: Modeling and calibration

LendScore is built using XGBoost, a gradient-boosted decision tree algorithm that balances accuracy and interpretability. Each feature is constrained to maintain a monotonic relationship with credit risk, ensuring, for example, that higher balances or more regular inflows always correspond to lower predicted risk.

The model was trained on 1.44 million tradelines, combining Plaid transaction data with credit bureau performance labels (e.g., whether a consumer went 90+ days past due in 12 months). It outputs a score from 1–99, with higher scores indicating lower risk.

Model validation (Kolmogorov-Smirnov statistic)

LendScore demonstrated strong orthogonality to bureau models (i.e., it captures distinct signals) and higher lift in risk segmentation, particularly for near-prime and credit-invisible populations.

Step 4: Decisioning: Compliance and fairness by design

Using cash flow data in credit risk models introduces unique fair lending risks that differ from traditional bureau-based underwriting. Because transaction data reflects day-to-day financial activity, it can inadvertently correlate with socioeconomic or demographic factors. For instance, spending patterns linked to geography, job sector, or household size. The biggest concerns can include:

• Proxy variables: Certain behavioral indicators (such as merchant types or payment type frequency) can unintentionally proxy for protected classes such as race or gender if not carefully managed.

• Outcome alignment: Ensuring that the predicted outcome (e.g., 90+ DPD) and its use in decisioning are equitable across groups.

To mitigate these risks, Plaid designed LendScore with rigorous safeguards:

• Feature-level review to exclude variables correlated with protected class proxies.

• Monotonic constraints that enforce consistent, intuitive relationships between features and risk.

• Independent third-party fairness audits that evaluate disparate impact and representativeness of the training data.

The independent review by FairPlay confirmed that LendScore’s features are relatively uncorrelated with protected class attributes, and that applying LendScore thresholds above 0.5 does not result in disparate approval outcomes across protected groups.

Explainability and adverse action reasons

One of the core regulatory requirements under FCRA and ECOA is that lenders must provide consumers with an adverse action notice explaining the key reasons behind a credit decision. To support this, LendScore generates five reason codes alongside every score.

These reason codes are derived directly from the model’s underlying feature importance and contribution scores at the individual level. Specifically:

• Each feature’s directional impact on the applicant’s score is quantified using SHAP (Shapley additive explanations) values within the gradient-boosted framework.

• The top five features that most negatively influenced the score are selected and mapped to standardized, human-readable reason code descriptions (e.g., "High volatility in account balances" or "Low inflow frequency relative to expenses").

• These reason codes are provided in the LendScore API response, allowing lenders to automatically include them in adverse action letters.

This approach ensures explainability-by-design, allowing credit risk professionals and compliance teams to audit the decision process while maintaining alignment with regulatory standards.

Step 5: Putting it together: A new layer of credit risk insight

The final LendScore model provides a rank-ordered risk score (1–99) that measures a consumer’s probability of default (90+ days past due within 12 months). It captures dynamic financial behavior unavailable in traditional credit models, helping lenders expand access while maintaining strong risk performance.

What’s next

LendScore rounds out Plaid’s suite of credit solutions that support lenders across the entire loan lifecycle. We’re committed to empowering our customers with innovative, compliant, and insight-rich tools that lead to more informed credit decisions. To achieve this, we’re adding new cash flow and network insights, modeling upgrades, and new segment-specific scores. 

Stay tuned for more exciting developments, check out our deep dive video, and sign up for our waitlist to be one of the first lenders to get access to LendScore and other new products.