How alternative data lending is reaching Southeast Asia's unscored MSMEs

The challenge in micro-lending is that small business data does not fit into standard banking templates.

For a major corporation, getting a loan involves a predictable sequence of paperwork. The company provides audited financials, tax returns, and collateral, allowing the bank’s underwriting team to evaluate the risk.

However, in Southeast Asia where over 70% of the adult population remains unbanked or underbanked, this system fails market vendors across the region, as well as small logistics operators in rural Mongolia and other remote locations. Traditional banking models view an undocumented history as an unacceptable risk, which leads to immediate rejection.

But, these business owners often record sales by hand, manage irregular cash flows, and lack both formal credit scores and physical collateral.

This dynamic creates a massive economic roadblock.

In Southeast Asia, micro, small, and medium enterprises (MSMEs) account for over 97% of all businesses and employ nearly 70% of the workforce. Despite their economic weight, more than 60% of them cannot secure capital when they need to expand. Furthermore, over 70% of the adult population in the region remains unbanked or underbanked.

Bridging this gap requires lenders to change how they process unformatted, real-world data rather than lowering their underwriting standards.

Processing data friction in emerging markets
The primary challenge in micro-lending stems from the fact that small business data does not fit into standard banking templates. Machine learning algorithms require structured inputs, yet informal businesses generate smudged paper receipts, mixed-language invoices, and erratic bank statements.

Overcoming this requires a highly specific data engineering pipeline designed to convert messy physical logs into clean credit signals:

Computer vision and regional calibration
Before an algorithm can score an applicant, intelligent document processing (IDP) must digitise the raw paperwork. This presents a complex computer vision challenge. IDP platforms require training on regional data subsets to accurately parse localised handwriting variations, non-standard layouts, and degraded paper documents. This processing step can help streamline inputs, making it significantly faster than manual data entry.

Agentic verification systems
Once the system extracts the data, an artificial intelligent (AI) validation agent cross-references the documents to detect anomalies. The system automatically cross-checks rent expenses visible in a digitised bank statement against the dates and terms specified in a physical lease agreement. By analysing documents as interconnected points in a network, the system verifies business legitimacy without relying on central credit bureau registries.

Algorithmic income reconstruction 
The stated income of an MSME rarely aligns perfectly with their bank statements because informal cash flows follow complex paths. Instead of searching for a single net income figure, the machine learning model reconstructs financial health by analysing transaction frequencies, average daily balances, and seasonal cash cycles. The algorithm weights these specific behaviours based on statistical significance.

For instance, the frequency of weekly cash deposits often carries more predictive power regarding repayment than an official business registration certificate.

Resolving the cold-start problem
Lenders face a "cold-start problem" when an applicant has zero digital footprint, leaving the algorithm with no historical data to evaluate. The solution requires transitioning from static, one-time assessments to dynamic data relationships.

Under a graduated lending framework, the platform approves an initial, low-risk micro-loan. This transaction serves as a practical data-gathering exercise. As the borrower repays, the machine learning model tracks their real-world consistency and monitors how their cash flow adapts to minor economic dips.

The model updates its risk assessment with every completed cycle. The borrower’s real-time performance data serves as a practical replacement for physical collateral. As the algorithm refines its confidence based on this live track record, the system increments the available loan limits.

In Mongolia, the Asian Development Bank (ADB) reports that whilst nearly 90% of registered businesses are SMEs, only about 10% of approximately 37,000 SMEs regularly access financing through banks. MSMEs constitute over 97% of all businesses in Southeast Asia and employ roughly 69% of the workforce.

Yet according to a study by the Tech for Good Institute, over 60% of surveyed MSMEs were unable to get a loan when they needed financing.

This shows that the opportunity is immense for the region, as this framework can function at scale. Technology that leverages lending informed by behavioural data, automated scoring, document intelligence, digital end-to-end origination is directly applicable to these markets.

My career began in mathematics and artificial intelligence research, but I find the most compelling application of these tools in the tangible social impact they generate for underserved populations.

When a business owner in a remote province receives her first loan approval through a smartphone, she gains access to capital that no traditional bank would have historically extended. Witnessing complex mathematical models translate into direct economic empowerment for these individuals remains the most significant outcome of this work.