EFL has transitioned our blog series to our new website and insights can be found here


The new face of EFL

Over the last 7+ years, EFL has undergone dramatic internal changes: from idea to impact, from over 2.5 hour paper surveys by hand to an automated Android app and API, and from a few people to over 30 across 12 countries. At the time of our first website, we had collected just over 2,500 hand-written paper applications.

Now EFL collects as many in a single week, streaming in 24 hours a day from more than 20 lending partners in a dozen countries. 

And we're now happy to say that our external image more closely represents our progress, vision, and product like never before, through the hard work of our internal and external marketing/branding teams*.


Our new logo is meant to visually capture the essence of what we’re doing at EFL:

  •  Expanding access to finance and opportunity for lenders
  •  Identifying new pockets of risk across the spectrum
  •  Bringing focus and organization to the lending process

Expanding circles depict the burgeoning markets that lenders seek to serve, flecks of red in a blue spectrum highlight new dimensions of risk and opportunity, and links in each circle convey EFL's focused and objective approach to credit assessment.

Though we've loved our red identity through the years, the team has really taken to our new blue persona.


In April we began transforming EFL’s brand and visual identity into a website with the goals of being: SIMPLE, INTUITIVE, and ACTIONABLE.'
  • SIMPLE: Because what EFL is innovative by nature, our prospective clients have, by definition, often not been introduced to our technology and approach. A website is a one-way conversation, and making sure visitors leave with a clear idea of what EFL does, and how it can help their customers and company is crucial. 
  • INTUITIVE: Much like the work of our product team on the User Experience of those who take the EFL application, (read more on our blog HERE,) we wanted it to be clear where to go for what’s needed, have a flow of information that’s a narrative, and actually FUN to learn about. 
  • ACTIONABLE: We not only want visitors to enjoy viewing the website, we also want them to be able to interact with us, and to come away equipped with resources they can digest offline, and share with their colleagues.

To do so, we’ve introduced a few new features that you’ll find different from our prior website:

  • CASE STUDIES: Thanks to the tireless efforts and successes of our partners, and our account managers around the world, we have been able to place an emphasis on REAL RESULTS. 
  • HOW IT WORKS (aka “EFL101”): In person, we spend most of our time explaining exactly what EFL is, what we do, and how it works. We want visitors to the website to come away with a very clear understanding of why EFL is relevant to their business, and to answer questions up-front. 
  • MAPS, MEDIA AND MORE: Rich and varied content to see beyond just our product, from the people here at EFL (aka the Elfs) and the individuals our tool helps around the world.
Through your support and enthusiasm, EFL will continue to grow. It took us 4 years to administer 100,000 surveys; just 4 months later we’ve grown that tally by over 50%.  

At EFL, we enable better lending decisions through better information. EFL helps people earn credit; thanks for reading, and for helping us spread the word!

*Addis Creson for branding and KA+A for web design


Demystifying Modeling

Understanding Predictive Power through the Lens of Power Failure

Predictive statistical modeling is a critical, but often omitted component of the “big data” discussion. Myriad stories tout the revolutionary consequences of the massive amounts of data being generated, but what techniques are being used to transform that data into meaningful insights, and should those techniques all be considered equal?

At EFL, our success hinges upon our ability to generate new, highly predictive data, but also on our ability to use powerful statistical models to derive value from this new data. We predict loan repayment from individuals’ answers to psychometric, and other non-traditional questions, enabling financial institutions to make better informed credit decisions. And while we recognize that not everyone loves data quite as much as we do, we think it’s important to better understand the various approaches to interpreting it.

The Basics

Predictive modeling is the process by which statistical models are created to predict the probability of an outcome. Models are like statistical factories - they generate an output from a series of inputs. The better the inputs supplied and the model selected, the closer the output is to reality. But models get more interesting when you allow the inputs to interact with each other, and when you layer models to capture new insights. Let’s take a closer look with a specific example: power outages.

Predictive Power and Power Failures

Power outages are a massive impediment to growth around the world, and particularly in developing countries. The World Bank estimates that in Tanzania, for example, 18% of annual private sector sales are lost due to electrical outages each year. As an example, let’s construct a model to help businesses anticipate power failures by thinking about what might be relevant in predicting a power failure:

Next, we need to think about the different ways that we can interact these variables to predict an outage. A few common methods are logistic regression, CART, random forest, and ensembling. 

Logistic Regression

One of the most widely used techniques in statistical modeling, logistic regressions predict the likelihood of a binary outcome by evaluating its relationship to a series of independent variables. Consider, for example, the relationship between a single variable, RainFall, and the event in question, a power outage.  A logistic regression like the one below would tell you how likely a power outage is given each centimeter of rainfall. In this case, the two variables are positively correlated: the more it rains, the more likely the power is to fail. By incorporating additional variables, like TimeOfDay, OutageBefore, and NaturalDisaster, a regression will become more predictive.
Though logistic regressions are widely used to evaluate basic relationships between binary variables and outcomes, they also possess important weaknesses. Specifically, they are unable to automatically uncover interactions between variables. While the regression would be able to predict if a power outage is more likely to happen in the morning or at night, or during a rainy or dry season, it would not naturally predict if power is more likely to fail during the morning in the dry season or at night during the rainy season. Such interactions must be foreseen and manually entered before running a regression. For a more dynamic methodology, we must look at Classification and Regression Trees. 

CART (Classification and Regression Trees)

Whereas logistic regressions evaluate many variables simultaneously, CART analysis evaluates variables incrementally through a decision tree. These models start with one major variable, and then branch out to additional variables subsequently. The order of these decisions is decided by their impact on the outcome: in this example, the occurrence of a natural disaster was the variable most indicative of power failure, so NaturalDisaster is placed first.

CART analysis builds upon logistic regression modeling by evaluating interactions between variables. To use the previous example, this means CART analysis would be able to capture the difference between a morning in the dry season and an evening in the rainy season, introducing a new level of nuance into the modelling process.

However, CART analysis suffers from a systemic “over-fitting” problem - the generated model can be overly tailored to the data it was modeled on. In other words, a CART model may perform well when evaluating past power failures, but perform poorly when predicting future failures. To overcome this challenge of over-fitting, data scientists often turn to Random Forest analysis, which mitigates the risk of over-fitting by introducing randomization into the modelling process.

Random Forest

While a CART methodology generates an output from a single tree, Random Forest modeling evaluates the outputs of many trees. A subset of variables and a randomized subset of data are used to generate multiple tree models, resulting in the creation of a collection of unique decision trees, or a Random Forest. From there, one can infer the predictive power of particular variables by counting their frequency and location on each of the trees.

Random Forest analysis is better able to cope with the problem of over-fitting, and thus boost predictive power, but it also requires significantly more time to execute. With large datasets, the numerous random trees can take a great deal of time to generate, thereby slowing down the modeling process.

Ensemble Methods

A useful approach from the field of machine learning is Ensembling, which allows us to use multiple models’ predictions to obtain better prediction performance: it takes the outputs of each of the models above and evaluates them as variables in a new model. Then, through any of the various techniques observed earlier, weights can be assigned to each and a new final output can be generated.

Ensembling is valuable not only because it enhances predictability, but also because it can provide a kind of security check. Ensembling can rely on many different models, so if one model fails the others can save the ensemble model’s output, thus better protecting your analysis.

Ensemble models are often uninterpretable, however, in that they don’t illustrate the relationship between independent variables like Rainfall and an outcome like a power outage. Because variables are interacted over many different models, it becomes impossible to parse out their individual impact from the model as a whole. So while ensembling is excellent at predicting outcomes from variables, it is weak at uncovering insights about the variables themselves.

Moving Forward

Not all modelling techniques are created equal. Rather, they each possess important strengths and weaknesses. Logistic Regressions, CART, Random Forests and Ensembling, along with many other techniques, have distinct advantages and disadvantages corresponding with different types and amounts of data. Power outages provide just one application, but you could build a model to predict just about anything. Modelling is both an art and a science, and while understanding the granular nuances of different statistical techniques could take years, understanding the fundamental differences between models provides a valuable window into our data driven world.


Maintaining Integrity Across Four Continents

An EFL application is administered every 5 minutes somewhere in the world.
So how do we ensure that the data we collect is reliable?

EFL’s psychometric credit scoring technology measures credit risk in markets where traditional credit data is scarce, helping financial institutions in the developing world identify and invest in borrowers whose lack of information had precluded them from financial access.  As EFL has grown from a research initiative at Harvard to a global company operating across 4 continents, we’ve developed new ways not only to model and analyze survey data from our partners around the world, but also to ensure that the survey data they collect are reliable depictions of prospective borrowers.

Understanding the Challenge of Data Reliability

Unlike most credit scoring methodologies, EFL actively gathers data to measure risk. Because we don’t rely on existing, backward-looking data, our technology can be used with anyone, anywhere.  Indeed, the universal accessibility of our tool is a big part of what differentiates EFL from other scoring solutions, but it also makes ensuring the reliability of the data we collect all the more important.

To put this in context, EFL’s psychometric credit scoring methodology comprises three steps.

When EFL was just beginning, we were able to manually track data coming in to make sure survey responses were reliable. For example, it was easy enough to check if applicants were taking time to thoughtfully answer questions just by taking a look at the time it took them to complete each survey.

But as EFL’s volumes and geographical footprint have grown, manually ensuring the reliability of survey data has become infeasible, and we’ve been pushed to find new ways of evaluating and monitoring survey data automatically. The sophisticated and nuanced analytical tools that we’ve built can not only help EFL ensure the reliability of our data, they can also help our partner banks better understand their operations on the ground.

A Scientific Approach to Data Reliability

To build a toolkit that would allow us to accurately and automatically identify surveys with unreliable data, EFL initiated a 4 phase R&D effort:
1.       Industry Research and Persona Development: We began by developing a set of scenarios in which a loan applicant might be motivated to answer questions differently than he or she would if taking the survey free of outside influence. For example, though EFL’s questions have no clear correct answer, some loan officers will coach loan applicants to answer questions in particular patterns that they believe to be advantageous.

2.       Develop Hypotheses: With these scenarios in mind,  EFL designed hypotheses to identify reliability issues within survey data. Using the scenario above, for example, applicants that have been coached by a loan officer will display similar patterns as other applicants by the same loan officer.

3.       Test in the Field: Next, we worked with a select group of partners, their loan officers and our field staff to test these hypotheses. We investigated preliminary findings and refined our analytical methodology accordingly.

4.       Ongoing Partner Feedback:  Finally, we developed a feedback loop through which we could gather and analyze field insights on an ongoing basis. Thanks to the steadfast commitment of our partners, we continue to regularly review survey reliability issues and stay one step ahead of challenges to data integrity.

Measuring Data Reliability in Real Time

Utilizing this approach, EFL has been able to build and hone sophisticated analytical tools to identify data reliability issues in real time. As soon as an EFL application meets our servers, we automatically flag applications that appear unreliable using a variety of survey and metadata variables. Importantly, EFL delivers these insights by analyzing not only an applicants’ answers, but also the way that applicants answer questions.

EFL uses these variables to flag EFL applications, ensuring that applications with unreliable data do not lead to loan decisions, and that loan officers in charge of those administering those applications are identified as soon as possible. Furthermore, EFL aggregates this data over time to show the evolution of data reliability and to pinpoint problem loan officers and branches.

Providing Value beyond EFL Scoring

EFL’s sophisticated data reliability tools have benefited partners not only by ensuring that the bank’s loan decisions are well informed, but also by providing partner banks with an unprecedented level of operational insight. A propensity to attempt to cheat on the EFL application, our partner banks have observed, is a strong indication of suspicious activity in other areas of a bank’s operations. By integrating EFL’s scoring, therefore, lenders have gained a new understanding of how staff are performing and which loan officers are less trustworthy than others.

In most cases, the vast majority of data reliability instances occur under the supervision of just a few loan officers. Of more than 160 loan officers in one South Asian microfinance institution, for example, a single loan officer was responsible for more than 20% of all flagged EFL surveys, and 5 loan officers were responsible for more than 60%.

Data insights like these allow lending institutions to identify suspicious loan officers, and to investigate them with a level of understanding that was previously impossible. In some instances, investigations into data reliability cases have uncovered larger problems at the branch level, problems that would have remained uninvestigated without EFL’s insights.  As many lenders face high staff turnover and have limited ability to quantitatively monitor the performance of field staff on a day-to-day basis, our partner institutions have wholeheartedly welcomed and endorsed the EFL data reliability tools.

Part Two
In part two of this series, will take a closer look at the exact mechanisms EFL uses to maintain data reliability across our geographical footprint, and how they have helped improve the lending efforts of our partner institutions.
Stay tuned for part two!

Kyle Meade
Product Manager

*Global data reliability averages based on partners using the EFL score for decision-making in 2013 and 2014


A User Experience for the Emerging World: Part 1

“People Ignore Design that Ignores People” – Frank Chimero, Author of “The Shape of Design”

Creating a positive User Experience, or UX, is a critical challenge for any organization, and in many ways it is especially vital in the emerging world. In places where basic technologies like mobile phones have only recently taken hold, financial institutions must find new and creative ways to make users feel at ease, engaged, and empowered.  

This challenge is particularly important in the banking industry, where increasing competition between a growing number of financial services providers is creating more options for customers. A positive experience can not only set a financial institution apart in the early stages of a banking relationship, it can also build customer loyalty and attract new clients. 

One study from the Grameen Foundation, for example, found that basic shortcomings in user interface design have dramatically hindered the uptake of mobile financial service offerings in the developing world. Simple changes in navigation, syntax and language, the report concludes, could significantly expand the reach and user base of those services across the globe. In markets where word of mouth is the most trusted endorsement, user experience can go a long, long way. 

An Loan Officer administers the EFL questionnaire in Jakarta, Indonesia's Capital City

“Design is not just what it looks like and feels like. Design is how it works.” – Steve Jobs

EFL's credit scoring technology enables partner institutions to lend to segments of the market previously considered "un-bankable." We understand that a positive user experience can propel the efficacy of our product, and that a negative one can just as easily undermine it. We've seen first hand that financial institutions in emerging markets can be intimidating to the entrepreneurs they aim to serve, and that's why over the past two years EFL has invested heavily in building a credit application that is both engaging and challenging, enjoyable and rigorous. We've spent countless hours in the field, asking questions, visiting bank branches, and working to understand every nuance of how entrepreneurs interact with our technology around the world. 

We knew from the beginning that the experience of the EFL survey needed to center around the end user. In our case, end users usually are not tech savvy smart phone owners but rather emerging market entrepreneurs, often with enormous business potential but relatively low degrees of technological literacy. That means that the EFL survey needs not only to be inviting and un-intimidating to those unaccustomed to tablets and computer software, but also to introduce unfamiliar exercises in clear and concise ways while keeping users engaged and motivated throughout the survey.

We approached this challenge with a three pillar strategy: priming, guiding, and communicating.

PrimingMotivating applicants to answer questions to the best of their ability begins with clear instructions and seamless transitions. By using positive priming techniques throughout the survey, EFL ensures applicants remain energized and enthusiastic throughout the application.

GuidingMany applicants have limited experience operating tablets and PCs and have never answered psychometric questions before, which can make the survey experience seem intimidating and stressful at first glance. By making the survey simple and easy to interact with, EFL ensures applicants will remain motivated.

Communicating- Making applicants comfortable with an unfamiliar application requires communication. By building clear signals into the survey to direct applicants, and also creating opportunities for applicants to report pain points, EFL ensures not only that the user experience is engaging, but also that it keeps improving. 

"If you can't measure it, you can't manage it" - Peter Drucker, Management Consultant and Author of “Managing Results”

User experience hypotheses are just hypotheses until they have been tried and tested in the field. So to prototype and collect feedback on our UX design innovations, we developed a beta stream of the EFL application and made it available to select partners in Peru, Indonesia, India, and Haiti. New features are rolled out to partner financial institutions in these regions quarterly, and feedback is collected directly from field staff through EFL’s Account Managers using a UEQ survey (User Experience Questionnaire). EFL aggregates and analyzes these thousands of data points from across the globe to identify which features are most effective at increasing user engagement and improving overall user experience. Successful features are subsequently rolled out to all EFL partners, while less successful features are withdrawn. This approach ensures a well-rounded and un-biased audience with which to test new UX designs, as well as a systematic method of continually improving EFL’s user experience. 

A loan applicant views newly released graphics in the EFL survey
In Part II of this UX blog series, we'll talk more about how specific design efforts are changing the way people interact with the EFL survey and transforming the way small business owners apply for working capital loans. How has segmenting exercises into smaller, more manageable pieces improved applicant engagement? How have animations enhanced short term memory assessment? How have graphical instructions improved applicants' ability to recognize patterns? 

Stay tuned for Part II!

Kyle Meade
Product Manager