As part of our commitment to transparency and impact, Tala commissioned a new global study conducted by impact measurement firm 60 Decibels in the spring of 2025. The report, which surveyed nearly 850 repeat Tala customers across Kenya, Mexico, and the Philippines, provides a powerful, data-driven look at how access to digital credit shapes financial lives over time.

The results clearly show the need for accessible financial tools is universal, and when people have the power of choice, they build more resilient and empowered lives.

1. The Universal Challenge of Financial Access

The struggle to access credit is not confined to emerging markets. In the U.S., 49 million adults lack conventional credit scores and about 15% of households have no mainstream credit, leaving them shut out of the financial system. This reality is mirrored in the experiences of our customers. About half of Tala’s borrowers reported having no prior access to a loan similar to ours. 

2. From Financial Stress to Healthcare Access

Financial emergencies and healthcare costs are deeply intertwined, creating a widespread vulnerability that many Americans share with our customers. While 59% of Americans lack savings for a $1,000 emergency, the U.S. healthcare system itself creates instability, leaving 11% of adults unable to afford or access quality care.

This is where financial access becomes a tool for overall well-being. With Tala, our customers are building remarkable resilience; a powerful 80% report an improved ability to cover an emergency expense. Specifically on healthcare, we see a life-changing impact: 55% of our borrowers in Kenya and 41% in Mexico now report an increased ability to seek medical care when needed. Being ready for an emergency gives people peace of mind, and having a fund or access to affordable credit can provide the agency people need to care for their health and build a more secure future. 

3. The Tangible Impact of Financial Control

Beyond managing emergencies, true financial control manifests in day-to-day well-being and an improved outlook on life. According to the study, a staggering 82% of borrowers reported an improvement in their overall quality of life because of Tala. This feeling of progress is validated by a decrease in financial stress, which 81% of our customers reported.

The impact on women is particularly profound and shows that those in the global majority share a similar desire for financial confidence and influence as their American counterparts. In the U.S., while 94% of women believe they will be personally responsible for their finances at some point, only 48% feel confident in their financial abilities. Our data shows that when women in the global majority are given the right tools, they act on this same desire for empowerment with incredible results: 80% of Tala’s women borrowers report increased self-confidence because of our loans, and 59% report gaining more influence in household decision-making. This newfound agency creates a powerful ripple effect, empowering women to improve not only their own lives but the well-being of their families and communities.

4. The Universal Entrepreneurial Spirit

The drive to build a business is a powerful, shared ambition globally. In the U.S., small business owners often rely on personal assets and consumer credit to finance their businesses and are more likely to be denied credit than non-owners. This challenge of securing capital is also a hurdle for entrepreneurs in the global majority. For these business owners, Tala provides a vital path to growth. Globally, 34% of Tala borrowers used their loan for business purposes, with Kenya leading at 40%. Tala is providing a direct investment in their vision, and the results speak for themselves:

• Increased Earnings: A remarkable 87% reported an increase in business earnings.
• Improved Outlook: 82% said their business outlook improved as a result of their loan.
• Investing in Growth: The most common use of business loans was buying inventory (82%), showing a direct path from credit to commerce.

The financial lives of our customers and many Americans are reflections of a shared journey. The desire for a safety net and the ambition to create a better life are universal challenges and aspirations. By building a financial platform that addresses these fundamental human needs, we are creating a more inclusive and connected global digital economy for everyone.

Methodology

Commissioned by Tala, 60 Decibels conducted 847 phone interviews with Tala’s digital loan borrowers in Kenya, Mexico, and the Philippines between April and May 2025. The borrowers were randomly selected from a sample of Tala’s borrower database. 60 Decibels is a global, tech-enabled impact measurement company that brings speed and repeatability to social impact measurement and customer insights.

The post A Shared Journey: How the Financial Struggle of the Global Majority Is Also an American Story appeared first on Tala.

For too long, financial access has been determined by rigid, outdated rules and static ways of understanding people. Since its founding, Tala has been redefining how to deliver financial services to a global population. Tala’s AI-driven technology strategy—which layers real-time decision architecture on top of proprietary data from our customers’ behavior and preferences—has allowed Tala to personalize credit terms and improve customer experience to a degree that we don’t believe is being done in financial services. 

Today, we’re introducing our newest AI model that doesn’t just analyze customers’ past behavior, but actually predicts what loan terms will help customers succeed, on a personalized basis. Think of this new capability as a recommendation system; much like how Netflix customizes content recommendations for each profile in an account based on prior viewing patterns. 

Tala InSight is a groundbreaking, patent-pending decision model that uses causal inference to optimize individual credit limits. The model marks a turning point for the financial services industry by challenging the long-standing assumption that credit score alone should be the determinant of credit limit. Instead, Tala’s new approach identifies the causal impact of changing a customer’s credit limit—awarding each borrower the amount of credit most likely to help them succeed. 

Through AI and now causal inference, Tala InSight will allow us to identify the credit terms that actually help each of our customers grow and then deliver a credit solution for them in real time. 

What makes Tala InSight different: from correlation to causation 

Traditional lending models rely heavily on correlation: if a borrower has a higher credit score, they’re more likely to get approved for larger loans. But correlation doesn’t explain why some people succeed while others don’t. That’s where causal inference comes in.

Causal inference—a set of methods that rose to mainstream attention after economists David Card, Joshua Angrist, and Guido Imbens won the Nobel Prize in 2021—focuses on identifying the true effect of an action. In Tala’s case, the question isn’t just: “What kind of customers succeed?” but “How can personalization help individual customers grow over time with Tala?” 

Tala InSight simulates these counterfactual scenarios across its global customer base using billions of proprietary data points gathered over 10+ years of mobile financial behavior. By doing so, it determines which specific credit limit causes the most beneficial outcomes for each borrower—whether that means increasing, maintaining, or even lowering a loan offer.

This represents a major step-change in credit modeling: a personalized, assumption-free approach that adapts to each customer’s real-world behavior and goals.

Early results: lower default, higher customer value 

The impact has been immediate and measurable. Most impressively, Tala customers across all credit profiles are taking out larger loans and defaulting far less. As a result, Tala InSight is delivering a 10% increase in three-month Customer Lifetime Value compared to Tala’s prior score-based policy—an extraordinary lift for an already scaled product.

This result is a testament to the power of personalization: when customers receive the credit that truly works for them, they are more likely to succeed. This means improved business performance for Tala and the right product at the right time for our customers.

A breakthrough in financial access

Tala, to our knowledge, is the first company in the world to deploy causal inference at scale in the service of financial inclusion. Tala InSight is the result of a cross-disciplinary team with deep expertise in machine learning, behavioral economics, and applied statistics. By combining high-quality behavioral data with advanced statistical infrastructure, Tala is building a modern credit system designed to help people succeed for the long term – not just provide a solution for one point in time. 

The credit decisioning use case is just the beginning. Our goal is to apply assumption-free, personalized causal thinking across all aspects of the Tala experience to unlock even greater customer and business value. 

At Tala, we’re not just building financial products; we’re building a more inclusive and equitable financial future. By harnessing the power of advanced data science and machine learning, we’re proving that personalized, data-driven lending isn’t just a theory—it’s a reality that’s changing lives and setting new standards for the entire industry.

The post Tala Unveils Breakthrough AI Model Using Causal Inference to Expand Financial Access appeared first on Tala.

At Tala, we’ve always believed that financial agency starts with radical trust in people’s potential, especially when they’ve been consistently overlooked by the legacy financial system we’ve inherited. 

For more than a decade, we’ve worked to turn that belief into action: building a global financial infrastructure company that unlocks financial access for millions of people around the world so they can thrive. Today, we’re proud to share new independent research that affirms just how powerful access can be. 

Real-World Impact

The peer-reviewed study conducted by researchers at Harvard, UC Berkeley, Northwestern University, and the University of British Columbia, is the first to show that mobile phone data can be used to objectively measure financial health. 

Published in the Accounting Review – a leading academic journal in the fields of economics, finance, and accounting – the research also shows that our digital-lending has made a significant positive impact on the financial well-being of our customers in Kenya across various measures. 

Using fully de-identified and anonymized datasets, researchers conducted a causal inference analysis of mobile-phone-based indicators of financial wellbeing, such as monetary transactions and balances, mobility, and self-reported income and employment. 

Among the core findings: 

• The power of data: Mobile phone data can objectively assess financial well-being, especially for those without formal financial records.
• Digital lending drives measurable financial improvements: Access to digital credit led to improvements in borrowers’ financial well-being across all measures they explored.
• Income and employment: Self-reported monthly income increased by 20.8% and the likelihood of being employed or self-employed rose by 23.5%.
• Financial transactions and balances: Borrowers’ average transaction amounts increased by 14.9% and they showed improvements in their mobile banking balances.
• Mobility: Access to digital credit resulted in greater mobility, with borrowers traveling to 9.4% more cities.

The findings offer robust, data-driven validation of what our customers have told us anecdotally for years: when you meet people where they are, you unlock real lasting impact. To date, more than 10 million customers have been able to handle financial shocks, manage household expenses, and leverage our credit to start and scale businesses worldwide. 

Our founder & CEO Shivani Siroya has pointed out that everyone likes to say they’re driving financial inclusion, but it’s rarely measured at scale. And this research doesn’t just highlight Tala’s potential – it quantifies it and provides independent, empirical evidence that proves our approach is delivering meaningful change and fulfilling the promise of financial technology.

What’s Next for Tala

Tala’s mission is to unleash economic power of the global majority, and that includes improving people’s financial wellbeing. While Shivani never set out to build a credit lending company, it quickly became clear that credit was the most effective way to test her thesis on the power of data as the foundation of the infrastructure needed to improve the financial lives of the global majority. 

Today, we’re expanding our platform with AI and blockchain to create new value for our customers beyond credit, and we’re excited to deepen our presence in new and existing markets across Latin America, East Africa, and Southeast Asia.

With this research, we’re more confident than ever in Tala’s positioning to be the leading financial platform enabling financial agency across the globe. 

Learn More

Follow the link here to explore the full research report, and be sure to check out our most recent impact report to learn how we measure the financial-wellbeing of our customers year-round. 

Congratulations to coauthors AJ Chen, Omri Even-Tov, Jung Koo Kang, and Regina Wittenberg-Moerman, and huge thanks to our incredible Data Science Team whose work helped make this possible.

The post The Power of Data: New Research Links Tala’s Digital Lending to Improved Financial Well-Being appeared first on Tala.

At Tala, we want to continuously improve our product and user experience. So, to learn whether the changes we make are beneficial, we run experiments. By establishing short-term proxy metrics of long-term value outcomes, we are now able to gain insights more quickly, expedite the experiment process, and make informed decisions more quickly to unlock value to our customers.

We originally deemed naive predictive machine learning models to be poorly suited to the job, as treatment effect estimates gleaned from linear coefficients or feature sensitivity analysis are likely to be biased. We were excited to discover that a statistical modeling technique called the surrogate index addresses this problem head-on. The surrogate index combines early read metrics to estimate long-term treatment effects in a causally sound way more rapidly and, to our surprise, at times more precisely than would be available from a full-duration long-term experiment.  

We reviewed the paper’s replication code and identified a couple of challenges: (1) it was written in Stata (our team writes Python and R code) and (2) it demonstrated the methodology using ordinary least squares (OLS) only, whereas the paper acknowledged that more flexible models might perform better. We overcame both of these challenges by implementing the methodology in Python and extending it with more flexible models.

Understanding and implementing the surrogate index

The surrogate index methodology involves using two datasets, one historical observational and another from an experiment, and has three modeling steps. The two datasets are similar, but the differences between them are crucial. 

Historical observational dataset

The historical observational dataset includes the long-term outcome of interest and early measures of the long-term outcome (aka short-term proxies). It may also include additional predictor variables. Notably, it doesn’t include a treatment group indicator because the data are not from an experiment. For example, if the long-term outcome is the total amount of money repaid minus the total amount of money loaned (aka credit margin) one year after a loan was disbursed to a borrower, then early measures of this outcome might be monthly values of this quantity at one month after loan disbursement, two months after loan disbursement, and so on up to the value at eleven months after loan disbursement. So, for this example, the historical observational dataset would contain the twelve monthly values of credit margin for a set of loans disbursed well before the period of experimentation.

Experimental dataset

The dataset from an experiment includes a treatment group indicator and early measures of the long-term outcome. And like the historical observational dataset, it may also include additional predictor variables. It doesn’t include the long-term outcome because this is the value we don’t want to wait to observe and measure. So, for this example, the experimental dataset would contain a treatment group indicator and the early measures of the long-term outcome for a set of loans disbursed during the period of experimentation.

Three modeling steps

The surrogate index modeling technique involves three steps:

Step 1: Using the historical observational dataset, regress the long-term outcome on one or more short-term predictors and any additional pre-treatment predictors.

Step 2: Using the experimental dataset and the model from Step 1, pass the predictor variables in the experimental dataset (the same predictors used in Step 1) through the model from Step 1 and predict long-term outcome values for the records in the experimental dataset.

Step 3: Using the experimental dataset, regress the predicted long-term outcome values on the treatment group indicator to estimate the average treatment effect (ATE) of the experimental treatment on the long-term outcome.

Identifying the smallest sufficient set of surrogates

One goal of the surrogate index modeling technique is to identify the fewest short-term predictors needed to reliably estimate the average treatment effect because doing so can deliver a more precise estimate.  Therefore, the technique is iterative, at least at the beginning, when determining the number of short-term predictors to include in the models.

For example, given the historical observational dataset described above, with its twelve monthly values of credit margin and no additional predictors, a first model for Step 1 could involve regressing the month-12 credit margin on the month-1 credit margin only. Next, for Step 2, one would pass the month-1 credit margin values in the experiment dataset through the model from Step 1 and predict long-term outcome values for the records in the experiment dataset. Finally, for Step 3, one would regress the predicted long-term outcome values on the treatment group indicator to estimate the treatment’s long-term ATE.

To determine the number of short-term predictors to include in the models, a second model for Step 1 could involve regressing the month-12 credit margin on both the month-1 and month-2 credit margins. In Step 2, one would pass the month-1 and month-2 credit margin values through the model and predict long-term outcome values for the records in the experiment dataset. Finally, for Step 3, one would regress the predicted long-term outcome values on the treatment group indicator to estimate the treatment’s long-term ATE.

This iterative process of constructing models with more and more short-term predictors would continue until the model in Step 1 contained all of the short-term predictors preceding the long-term outcome. In our example, this expanded model in Step 1 would regress the month-12 credit margin on eleven predictors: from month-1 to the month-11 credit margin.

By evaluating this technique using a historical experiment where the long-term ATE has been measured, one can compare each of the eleven estimated ATEs to the actual, observed long-term ATE. One can then see which model estimates are similar to the observed ATE. According to the authors of the paper, the model with the fewest predictors that still recovers the observed ATE is preferred because “by identifying suitable surrogates…one can strip out the residual variation arising from downstream factors that create noise” thereby resulting in a more precise estimate. Figure 3 from the paper illustrates this iterative process; varying the number of short-term predictors included in the model identifies the model with the fewest predictors that still recovers the observed ATE.

Internal evaluation and validation

We were excited to discover a published technique that purported to solve our problem. However, before using it in live experiments that would influence our lending and repayment strategies and product features, we decided to validate it with our historical experiments. 

These validation analyses allowed us both to evaluate the performance of models that are more flexible than OLS and to build the modeling technique in Metaflow, a Python library that makes it straightforward to develop, deploy, and operate various kinds of data-intensive applications.

To validate the technique, we evaluated it across a broad range of experiments and model types:

• We built many types of models, including Elasticnet (Enet), Generalized Additive Models (GAM), Random Forest (RF), Gradient-Boosted Trees (GBM), and Bayesian Additive Regression Trees (BART), and evaluated their performance with and without additional predictors.
• We evaluated its performance with historical experiments spanning across lending and repayment strategies, product features, and countries.

ATE estimates are unbiased

All of our validation analyses recovered the long-term ATEs we had observed in our historical experiments. That is, our estimated long-term ATEs were aligned with the actual ATEs, and our estimates’ bootstrapped 95% confidence intervals contained the actual ATEs. These results gave us and our business partners confidence in the technique, which paved the way for conversations about using it in future experiments.

ATE estimates are at times higher precision

Moreover, we observed the increase in the precision of the average treatment effects estimates that the authors discussed in their paper. The authors noted that the surrogate index “yields a substantial increase in precision…a 35% reduction in the standard error of the estimate” (in their application).

The authors explain this result can occur because “the surrogacy assumption brings additional information to bear on the problem – namely that any variation in the long-term outcome conditional on surrogates is orthogonal to treatment and hence is simply noise that increases the residual variance of the outcome and reduces precision.” So, “by identifying suitable surrogates…one can strip out the residual variation arising from downstream factors that create noise.” Therefore, “surrogates purge the most noise when the residual variances in treatment and the long-term outcome given surrogates are high, thereby yielding larger efficiency gains.”

As shown in the above plot, the 95% confidence interval for the observed average treatment effect on our long-term outcome of interest (i.e., the green dotted lines) ranged from approximately 0 to 12. Whereas, the 95% confidence interval for the 1-month surrogate index estimate of the ATE (i.e., the leftmost blue vertical lines) ranged from approximately 5 to 10. That is, in our application, like in the authors’, the surrogate index yielded a substantial increase in precision — in our case, a 58% reduction in the width of the confidence interval. We observed similar increases in precision across our remaining validation analyses. As the authors pointed out, “the results imply that it is optimal to use the smallest set of surrogates that satisfy the surrogacy assumption to maximize efficiency.”

Succeeding with shorter, reliable, reproducible experiments

We have learned a lot by extending and validating the technique with our historical experiments, particularly the benefits of collaborating with business partners from the beginning, using flexible models, and building the technique in an infrastructure that makes it easy to develop, deploy, and operate.

First, by collaborating with our business partners to select the historical experiments, define the key variables and time frames, and review the results, we ensured our business partners were invested in the validation effort, understood the analyses, and were equally excited about the favorable results. Second, by extending the technique with flexible models and additional predictors, and comparing these models to OLS models with fewer variables, we were able to identify models we are confident will provide reliable estimates of long-term ATEs in future experiments. Third, by building the technique in a robust infrastructure, we have reliably reproduced results, generalized the technique across countries, experiments, and long-term outcomes, and enabled our partners to use it quickly, easily, and appropriately.

We have realized many benefits from validating and implementing the technique for internal experimentation. By building the technique in a robust infrastructure, we have enhanced and standardized this methodology. Now, analyses of experiment results are consistently reliable, reproducible, and of high quality. We are just coming out of the implementation and validation exercise. Now, with our validated approach to estimate these long-term ATEs more rapidly and reliably, we’re excited for the next phase: using them in our live experiments.

The post Early-read models for shorter-duration experiments appeared first on Tala.

Inflation affects the Global Majority personally 

Our customers are acutely impacted by inflation as nearly 1 in 2 customers surveyed indicated they had no rise in pay in 2023. This makes it challenging for customers to keep up with daily expenses. In fact, 90% of respondents agree that inflation impacts their family budgets, and most note groceries as the most common pain point. Across the board, our customers around the world are feeling financial pressures.

Inflation has implications for both essential spending as well as leisure. Thirty-five percent of global respondents feel they rarely or never have anything left for leisure or enjoyment. However, this varies slightly by market. Filipino and Kenyan customers are cutting back on non-critical expenses to meet basic needs, more so than in Mexico. 

Customers find ways to cope via entrepreneurship

Despite the complications of inflation on customers, they’ve identified a variety of methods to cope with increasing costs and manage difficult times. Globally, 59% report starting business, side hustles, or part-time roles as a means to cope. Although a global behavior, we see slight variances. In Kenya and Mexico, there is a preference for starting a business or a side hustle (40% and 42%, respectively). In the Philippines, 34% report picking up a second or part-time job, and 36% report starting a business or side hustle. 

As customers develop other avenues for income, we also see signs of resilience. Across our markets, 48% occasionally experience stress related to inflation, but not constantly. In fact, 19% rarely experience stress. Reports of high financial stress were seen most notably in Kenya. There, 17% reported almost always feeling stress about their financial situation — nearly three times that of our other markets. 

The majority feel supported to navigate financial headwinds 

When asked how Tala has impacted their finances, we see high reports of peace of mind. Generally, our customers feel they have the right financial resources — the majority of respondents agree they have the right tools to manage their money how they want to, and 26% strongly agree. Customers worldwide have been able to handle financial shocks, manage household expenses, and leverage Tala loans to start and grow businesses. 

This report illuminates the remarkable resilience and resourcefulness demonstrated by our customers grappling with the challenges of inflation. As we navigate economic uncertainties, it is evident that the Global Majority are not just passive observers but proactive participants in shaping their financial destinies. From disciplined budgeting to embracing supplementary income streams, the diverse strategies employed by our customers underscore the dynamic nature of economic progress. Learning from these behavioral shifts is crucial in developing effective solutions to support individuals in maintaining their desired quality of life despite ongoing inflationary pressures. 

Notes on methodology

Tala conducted an online survey among 2,173 Tala customers living in Kenya, the Philippines, and Mexico to understand perceptions of inflation and how people are managing their financial lives accordingly. The sample is representative of our active customer populations with a 95% and higher confidence level across each sample set.

The post Resilience and creativity in the face of inflation appeared first on Tala.

I’ve been thinking a lot about the words we use and read every day. 

Words represent our worldview and impact how others perceive us and themselves. Words can lift people up or tear them down, motivate or deflate. The right words can inspire hope and drive action. This blog post is about finding the right words for the people we exist to serve.

The Global Majority

At Tala, we use the term Global Majority to describe the portion of the world’s population who, while having trillions of dollars in economic power, have historically been excluded from accessing financial services. In many cases, these people earn money in the formal and informal economy, but navigate their daily lives without access to savings, credit, bill payment tools, or the ability to affordably transfer money.

Industry experts often refer to this population as unbanked or underbanked, but, alone, these descriptors fall flat and ignore the power of this enormous group. At best, these words define over half of the world’s population by one increasingly unjustifiable barrier to participation in the global economy—access to financial services. At worst, they cast the majority as victims. We think this framing is wrong, and we want to flip it upside down.

The fact is, the Global Majority have immense potential and are more numerous than any economic subgroup on the planet. If their economic activity was concentrated in a single country, that country would boast one of the largest economies on Earth. This truth must be acknowledged and respected, a journey that starts by using the term Global Majority to actively shift the perception of this population away from what they don’t have toward what they do have—power.

By the numbers

The Global Majority do not resemble the media stereotype of lower-income people. They have jobs and spending power and survive with less government assistance than some of the world’s wealthiest people. 

The largest portion of the Global Majority is the population working and living above the poverty line but below middle-income. Many of them have multiple jobs and run small all-cash businesses. This includes those who might slide across the invisible poverty line monthly, weekly, or even daily, based on personal and macroeconomic circumstances of the country and community they live in. Pew Research Center refers to this income band, defined as living on $2.01-10.00 per day, as low income and estimates this population to be 52% of households globally or 4 billion people.

While Tala provides financial services to this population, we also know that many of our customers fall into Pew’s middle-income band, representing another 17% of the world’s population. And in some markets, we even see customers whom Pew would define as upper-middle-income, living on $20.01-$50.00/day.

Pew’s research and Tala’s customer data, taken together, say something shocking about the world we live in: 1) roughly 20% of the world’s population—those in the high and upper-middle-income brackets—likely constitute 100% of the customer base for legacy financial institutions, and 2) 60-70% of the world’s population lack access to the financial services they need—from savings, to credit, to remittances. This means that while they’re generating massive economic activity, the Global Majority are still unnecessarily disadvantaged when it comes to achieving basic financial wellbeing for themselves, let alone upward mobility. 

Tala has already proven that these old barriers can be smashed with innovative technology, creative thinking, and trust. Now, we’re working to expand access globally. And to any financial institution that still thinks the unbanked are unbankable for a reason, we have more than eight million customers who would like a word. 

Tala exists to unleash the economic power of the Global Majority.

We say this because words matter, and the right words can inspire hope and drive action. For us, that action is building best-in-class technology that democratizes access to vital financial services for hardworking, trustworthy people everywhere.

The post Who are the Global Majority? appeared first on Tala.

Data science is at the heart of Tala. Our combination of a data moat and state-of-the-art AI tooling enables us to offer credit to the underbanked in dynamic regulatory environments without necessarily relying on credit scores. A key market advantage is the ability to continuously innovate predictive and causal credit, fraud, and recovery machine learning models with the promise of lifting customer lifetime value and expanding financial inclusion. But gains from research efforts go unrealized unless the ML model time-to-market is equally agile. 

We are thrilled to announce that Tala has entered into its new era of continuous ML model delivery. This marks a significant leap forward in efficiency, responsiveness, and overall performance that ensures Tala will maintain its leading position as a credit product innovator for the global majority. 

This year, we rebuilt our batch model training, real-time feature and model serving, and acceptance testing infrastructure from the ground up to address our biggest pain points and make progress toward push-button automation. As we have rolled out the changes this month, we are seeing dramatic, double-digit lifts in speed and efficiency with historically low risk-event exposure. Less time wrangling machine learning code and infrastructure translates directly into more time improving models and building new ones to drive compounding business impact and, ultimately, value to our customers.

Streamlining deployment to enable continuous innovation

We focused on three areas of improvement to minimize manual toil for our developers and data scientists and to address bottlenecks in our model development lifecycle. 

1. Streamlined feature service deployments using a simple architecture that prioritizes automation, observability, debuggability, durability, and fast canary deployments using Flask, Kubernetes, and Kafka. 
2. Reproducible and scalable model trainers using Metaflow from Outerbounds that allow data scientists to own production-worthy, push-button model training code.
3. Automation and self-service at all opportunities. 

Reduced model training time

Under the new infrastructure, keeping models fresh and compliant is light work. In just a day or two we are able to: 

• remove model features that become unavailable
• add already acceptance-tested new model features
• refresh the model by simply training on newer data

We achieve speed and reliability by leveraging cached, precomputed features and adopting Metaflow to run model training jobs with horizontal parallelism. Metaflow also yields us a 5x boost in cost efficiency over our previous practices because it uses shared compute resources on demand instead of data scientists using dedicated, always-on computers.

Blazingly fast onboarding

Our new designs also significantly decrease the time it takes to initiate a training run in production for data scientists rotating across models and for new data scientists joining the team. By eliminating cross-team requests, leaning into self-service and automation, and prioritizing trainer durability, the time-to-trigger for these training runs is now just minutes. Putting it all together, total time for a new hire to train their first model has gone from multiple weeks or months to about a day. 

Sub-second credit scoring

Median scoring time for new borrowers has also sped up, primarily due to simplifying the ML architecture. We see sub-second typical feature generation and model scoring latency. End-to-end credit approval decisions for Tala customers, from application submission to decision, now take less than three seconds – a sizable improvement over our previous architecture.

Stack modernization 

We modernized the stack, including leveraging canary deployments in Kubernetes that autoscale up and auto-rollback when errors are detected, and we optimized bottlenecks in our deployment pipelines. Previously, our deployments were time-intensive and occasionally required painful manual reversions. Deployments now take less than two hours, and reversions are automated and fast to dramatically minimize exposure to our customers. While there is still room for improvement, these deployment times are substantially faster than before. Production hotfix time is now less than ten minutes, and feature service rollback is essentially instantaneous when we need it.

Strategic scalability: Configurable solutions for Tala’s growth

Each of these solutions is scalable to all current and future Tala markets primarily via configuration changes and shared code. This means we can speedily enter new markets with low variable costs, which is no small feat. While US lending hinges on making marginal improvements to FICO scores in a single market, bureau data within our multiple markets — if available at all — is inconsistent and does not cover the huge populations we at Tala are specifically trying to serve. Our ability to sustainably provide the global majority with access to credit hinges on ingesting novel data types and quickly incorporating them into our core ML models. Doing this, especially under varied and changing regulatory environments, requires a high level of agility. This is Tala’s differentiating core competency and a top priority when improving the ML infrastructure. 

These improvements are a testament to our pursuit of efficiency and excellence. It is progress with compounding benefits. We’ve increased our data scientists’ research and analysis time to upwards of 70% as they reduce time spent wrangling code and infrastructure, and our machine learning engineers can operate an ML infrastructure that they feel deeply invested in and can be extremely proud of, and that enables Tala to unlock the power of data to drive value for our customers.

The post Tala’s new era of agile machine learning model delivery appeared first on Tala.