SBP’s Digital Bank Rollout: Assessing Credit Risk Models for Pakistan’s Unbanked Population

The Information Asymmetry of Pakistan's Informal Economy

Pakistan does not suffer from a lack of loanable capital; it suffers from an information asymmetry that renders 57% of its adult population financially invisible (World Bank, 2021). The launch of the State Bank of Pakistan (SBP) digital bank licensing framework represents a structural attempt to bridge this divide. Yet, as five newly licensed digital banks prepare to deploy their balance sheets, they confront a stark reality: the traditional tools of credit risk assessment are useless in a country where the informal economy accounts for approximately 35% to 40% of gross domestic product (PIDE, 2023). Without formal credit histories, tax filings, or verifiable payrolls, the unbanked population cannot be assessed by conventional credit bureaus.

This analytical paper evaluates the viability of alternative credit risk models within the context of SBP’s Digital Bank Rollout: Assessing Credit Risk Models for Pakistan’s Unbanked Population. It examines how machine learning algorithms, non-traditional data pipelines, and domestic technology infrastructure can be synthesized to price risk accurately. The inquiry is not merely academic. The survival of these digital banks depends on their ability to underwrite low-income borrowers without triggering a wave of non-performing loans (NPLs) that could destabilize the broader financial system. As Pakistan's software exports grow—reaching $3.22 billion in fiscal year 2023–24 according to the Pakistan Software Export Board (PSEB, 2024)—the domestic technological capacity to build these models has matured, but the regulatory and structural integration of alternative data remains fragmented.

The Regulatory Genesis of Digital Banking in Pakistan

The State Bank of Pakistan’s regulatory journey toward digital banking began in earnest with the issuance of the Licensing and Regulatory Framework for Digital Banks in January 2022. This framework established two categories of licenses: Digital Retail Banks (DRBs) and Digital Full Banks (DFBs). The primary objective was to promote financial inclusion, encourage cheap credit, and foster a digital payments ecosystem. By October 2023, the SBP had granted in-principle approvals to five consortia, including Easy Paisa DB, HugoBank, KT Bank, Mashreq Bank, and Raqami. These entities are mandated to maintain a minimum capital requirement of PKR 1.5 billion, which will gradually scale to PKR 4 billion over a five-year transition period.

This regulatory push occurs against a backdrop of low private sector credit-to-GDP, which stood at a mere 15% in 2023 (State Bank of Pakistan, 2023). Traditional commercial banks in Pakistan have historically preferred to park their liquidity in risk-free government securities (PIBs and T-Bills) rather than extend credit to small and medium enterprises (SMEs) or agriculture. This crowding-out effect has starved the real economy of capital. For the unbanked, the situation is even more acute. The lack of physical collateral—such as land titles or urban property—prevents them from accessing formal credit. Consequently, they rely on informal, high-cost credit networks, such as local money lenders or committee systems, which extract high rents and offer no path to formal wealth creation.

The Mechanics of Alternative Credit Risk Models

To underwrite the unbanked, digital banks must replace the traditional credit score with Alternative Credit Scoring (ACS) models. These models rely on machine learning algorithms—such as gradient-boosted decision trees (XGBoost) and neural networks—to process non-traditional data streams. The core hypothesis of ACS is that digital behavior is a proxy for financial character and repayment capacity. In Pakistan, where mobile teledensity stands at 79% (PTA, 2024), telecom data is the most abundant source of alternative information.

The causal chain of alternative underwriting operates through specific transmission channels. For example, telecom data produces credit insights via airtime top-up frequency and call detail records (CDRs). A borrower who consistently tops up their mobile balance in small, regular increments demonstrates cash-flow stability, whereas erratic, large top-ups may indicate volatile income. Furthermore, mobile wallet velocity—measured by the frequency and volume of transactions on platforms like Easypaisa or JazzCash—provides direct evidence of disposable income. According to the State Bank of Pakistan, mobile wallet transactions grew by 30% year-on-year in 2023–24, generating billions of data points that can be ingested by credit risk models (SBP, 2024).

However, the transition to algorithmic underwriting introduces second-order risks. The first-order effect of using alternative data is increased credit access for marginalized groups. The more consequential second-order effect is the risk of algorithmic bias and systemic exclusion. If a machine learning model is trained on historical data from urban centers, it may penalize rural borrowers whose transaction patterns differ. For instance, agricultural cash flows are highly seasonal, characterized by large inflows during harvest periods and long periods of inactivity. A rigid algorithm designed for urban salaried workers would misclassify a viable rural farmer as high-risk, thereby reinforcing the very financial exclusion it was designed to solve.

"In a market where 40% of the economy is informal, a credit risk model that relies solely on formal banking history is not conservative; it is obsolete."

Data Fragmentation and the Open Banking Deficit

The primary structural constraint facing alternative credit risk models in Pakistan is data fragmentation. Unlike India, which developed the Unified Payments Interface (UPI) and the Account Aggregator (AA) framework, Pakistan lacks a centralized, consent-based data-sharing architecture. Digital banks are forced to negotiate bilateral data-sharing agreements with telecom operators, utility companies, and provincial revenue authorities. This fragmentation prevents the creation of a holistic view of a borrower's financial obligations.

Consider the case of utility data. In Pakistan, electricity distribution companies (DISCOs) maintain separate, non-digitized databases of consumer payment histories. A borrower's utility bill is a powerful indicator of their repayment capacity and willingness to pay. Yet, because this data is not integrated into a centralized credit registry, digital banks cannot access it in real-time. The State Bank of Pakistan has attempted to address this through the implementation of Raast, Pakistan’s instant payment system. While Raast has successfully lowered transaction costs, it does not yet function as a data-sharing registry for credit underwriting. Consequently, digital banks operate in information silos, increasing the risk of over-leveraging borrowers who take out multiple nano-loans from different platforms simultaneously.

Macroeconomic Volatility and Algorithmic Calibration

The performance of alternative credit risk models is highly sensitive to macroeconomic conditions. Pakistan’s economy has been characterized by severe stagflation, with consumer price index (CPI) inflation averaging 23.4% in 2023–24 and interest rates peaking at 22% (State Bank of Pakistan, 2024). In such an environment, historical data becomes a poor predictor of future repayment behavior. An algorithm trained during a period of relative economic stability will fail to predict defaults when inflation erodes the real purchasing power of low-income borrowers.

This vulnerability is illustrated by the microfinance sector’s experience during the 2023 economic shock. Non-performing loans (NPLs) in the microfinance sector, which serves a demographic identical to the unbanked target market of digital banks, escalated to 8.4% (Pakistan Microfinance Network, 2024). The transmission mechanism was direct: rising fuel and food prices forced micro-entrepreneurs to divert cash flows from debt servicing to basic survival. For digital banks, this highlights the necessity of building dynamic, real-time feedback loops into their credit models. Static algorithms must be replaced by adaptive models that adjust credit limits and interest rates dynamically based on high-frequency macroeconomic indicators, such as fuel price adjustments and localized consumer spending patterns.

The Counter-Case: The Limits of Algorithmic Underwriting

The argument that traditional banking methods are superior fails to account for the changing technological landscape of Pakistan. The country's IT sector, supported by the Pakistan Software Export Board (PSEB), has demonstrated the capacity to build sophisticated data pipelines. In FY24, Pakistan's IT exports reached $3.22 billion, driven by software development and fintech solutions (PSEB, 2024). This domestic engineering talent is highly capable of developing localized machine learning models that account for the unique cash-flow patterns of the Pakistani informal sector. The barrier is not technological capability, but regulatory access to public data registries.

To unlock the potential of these models, the State Bank of Pakistan must collaborate with the National Database and Registration Authority (NADRA) and utility companies to establish a secure, consent-based open data exchange. For example, integrating NADRA’s biometric verification system with real-time utility payment registries would allow digital banks to verify a borrower's identity and cash-flow consistency instantly. This would significantly reduce the cost of customer acquisition and underwriting, allowing digital banks to offer lower interest rates to low-income borrowers.

Conclusion & Way Forward

The State Bank of Pakistan’s digital bank rollout is a bold regulatory experiment, but its success depends on the mathematical integrity of its credit risk models. To prevent a systemic rise in non-performing loans, digital banks must move beyond basic telecom data and build multi-dimensional alternative credit scoring models. This requires a coordinated policy effort. The SBP must mandate a comprehensive Open Banking Framework, forcing traditional banks, utility companies, and telecom operators to share data through secure, standardized APIs. Furthermore, the domestic IT sector, supported by the Pakistan Software Export Board, must continue to develop localized machine learning models that can adapt to Pakistan's volatile macroeconomic environment.

For a deeper dive into Pakistan's digital economy and financial sector reforms, see our Technology Section and our analysis of the Pakistan Macroeconomic Outlook. The transition from collateral-based lending to algorithmic underwriting is not merely a technical upgrade; it is a structural necessity for sustainable economic growth in Pakistan.