About the talk
Over 600 million people in Africa do not have access to modern energy, which impacts a household’s health, education, quality of life, and financial wellbeing. Thanks to hardware and financing innovations—all-in-one solar home systems that don't need a grid connection paired with flexible, pay-as-you-go credit plans—there is a promising potential path to energy access and financial inclusion for those otherwise left behind. However, despite initial growth, off-grid solar needs to scale even more rapidly to meet the needs of their customers and help achieve the SDG7 target of universal energy access by 2030.
To achieve this will require data innovation, using the power of advanced analytics to better target customers and assess credit risk for households that are largely unbanked. Nithio is leading that charge, as an analytics and finance platform that utilizes applied machine learning to create alternative credit scoring and other data science-powered insights, including: where to find ideal off-grid customers using geospatial modeling; how to market products households need and want, using segmentation and prediction; and what their expected payment streams will be using time-series forecasting. Nithio’s B2B analytics products help both solar companies and their investors manage risk to expand and adapt in a growing, challenging market. We believe that implementing context-specific data science models, alongside financial and operational innovation, is key to building ecosystems for critical social goods such as sustainable energy access for all.
Madeleine Gleave is the Chief Data Scientist at Nithio, where she leads the data team in utilizing machine learning for unparalleled consumer risk analytics across Africa. Ms. Gleave was previously at Dharma.ai where she led implementation of end-to-end data management software for major iNGOs and companies. Prior to her work in social impact startups, Ms. Gleave also conducted policy research on data-driven humanitarian and development strategies at the International Rescue Committee, and on energy access and infrastructure investment at the Center for Global Development.View the profile
So without further Ado, welcome back from lunch everyone. Next up. We have Madeline Glee's. The chief. Is it a scientist at 5000 is an energy Finance platform. That provides analytics and Financial Solutions in Africa's distributed energy sector today. She's presenting her talk data, science is key to achieving energy access in Africa. I thank you, Sarah up for that introduction. So go ahead and get started. So. Here's my picture. Feel free to contact me on Twitter on LinkedIn or check out miss you as website. But yes, there is an introduction
of this is really helpful. I'm just as summarized again. We are a startup that is focused on the social impact space. So I don't know how well will compete with their kind of feel. Good of classifying lions by with their spots, but hopefully will also help demonstrate how they designed can be used for, in a social impact or important issues. That have long been, an unsolvable for lack of technology and enter the data resources. Pinocchio's focus is on expanding access to Modern energy, particularly focused on Africa where there are
over almost 600 million people who don't have access to Modern forms of energy of those. Do you have to pay for Energy Services, but they just don't have access to them. And there are products that do exist package solar off-grid solar system, but they're not stealing at the pace that are needed to reach Everyone by 20-30, which is when the Sanibel development goals has targeted for Universal energy access to provide credit risk, analytics Solutions and financing that help unlock that access app. Quick primer on off-grid solar for those who may not be familiar.
So it's a really interesting Innovation on the hardware side in the last decade or so that's enabled. A retail model or Energy Services that requires no great connection. I'm so sore home systems. There are a number of different brands that have emerged specifically come with a package of a solar panel and various lighting solutions and can also include an appliances that can be used to improve quality of life, such as televisions or radios. That also productive use of cultural process thing. So there are over four hundred and twenty million users of these
solar home systems worldwide. And one of the really important Innovations with them is a pay-as-you-go financing model, which makes them very affordable and I'll speak to you a little bit this unique and bedded pay-as-you-go model. We're actually the companies can turn off the system. Turn it off and on depending on what how much the customer has paid for those services on for that energy enables a very flexible payment schedule but also introduces some challenges and opportunities from a technology and from a data standpoint.
Philly solar home systems are package to scale and we know that there are millions of Africans that have the ability to pay for them at the affordable price. And yet, it's not stealing at the rate. That's needed to reach everyone who needs energy. There is a lot of investment that's going into the sector but it's not necessarily targeted in the most effective way and it's not the skill that's needed. And we really see that there are two main bottle. Next one is a lack of financing infrastructure. That's scalable which one if you also works on but really the team that I lead and, and the,
the products that we've developed is focused on the second Baldwin neck, which is that ninety percent of African. Do not have a credit score unlike in the United States. Where are our car? Many developed markets where you have like a FICO score or something similar. There's just very limited credit bureau coverage because most of these households have been on banks. Historically, they don't have Financial access and so their wrists of a financing and asset like a solar home systems. $9,700 is pretty significant for the company has and can provide a lot of uncertainty about how to spell.
So with these emerging trend of the market and this new technology products, and there's several really interesting as a consequence of a really interesting opportunities for data science to play a role in resolving that credit risk problem. So the first is that you're there at these companies, he's Retail Energy service companies that are essentially digital natives. They've emerged in the last few years in the last decade. I don't like many traditional financing players or even microfinance players or traditional retail service models in in Africa,
they have embraced and really from from their design, have digitized records that enable really interesting work with data. So it's often been a problem that there have been a lack of digitized financial transactions to build credit scoring model models off of. But because the pay-as-you-go models are tracking. And payments through mobile money or true true kind of regular digital payment platforms that are embedded Within These Energy System. It has that in a uniquely
digitized record that we can use to train models and better understand payment behaviors. They are also using surround platforms that has high integration potential for other data, science Solutions on and mobile applications. That field agents are using a point of sale to register customers on taking information and also to make decisions on the ground. And then the third is that they are. There iot enabled products by and large and have remote sensing data on both energy usage and locations which provides a really interesting additional data stream that
can be used. So that's on the side of kind of the customer level data. There's also been advancements on geospatial data, collection and methods, but providing unique opportunity for data science. And as this is the date of the special session. Kicking it off. I must spend a bit of time talking about this but we work with partners that I have developed proprietary methods to improve the geospatial data, that's available for the continent of Africa, has been pretty locking at least in a degree of granularity. So using harmonization of household surveys that have been conducted by
large-scale entities like the World Bank usaid the UN and then combining that with satellite imagery to produce, very granular, high-resolution roster, layers up to one square. Kilometer of estimates of socio-economic and environmental characteristics that are highly predictive of credit risk, inability to pay. I know they've been available through a variety of open-source, but also, proprietary platforms, which Nikia works with. But some of the challenges and solutions for off-grid energy for these
companies that are essentially and if he has clients were a B2B company. So they can be summarized as one of three mainstream one where to find ideal offered customers to how to market products that customers need and wants soap, kind of what's the right products fits or or financing fits for the type of customers that they're talking to and Third, how to manage the cash flows from this really unique and flexible pay-as-you-go financing asset. Essentially their loans that are being offered to these really remote and ofttimes, very low income household.
And because that the pay-as-you-go model has no concrete payment term for payment trajectory is very flexible. It's it's it's literally just when the customers want to use the service, those payments dreams can be very on. Predictable and hard to translate into traditional models of risk. 2006 to use data science to solve this tree and kind of pain classified into three use cases or product. One targeting too boring and three valuation. I'm so leveraging that high-resolution geospatial data to
prioritize areas with Target customers developing kind of credit scoring tools that can be used to underwrite or monitor portfolio. Performance and third valuation, which is really used on that financing side to link up with investors, help them, understand the value and the cash flow that can be expected from portfolios of the account. So the way that material Works, kind of our analytics engine on so summarizing 40 steps than, and this is formulated for Carnival mortech non-technical audience. So if I do
simplified words, but I'm starting with the sources of data we work with, like I said, and we have the the CRM data that are coming from the, the the company's distribution company that we work with. And those peers spatial data layer is, which we have access to thousands of individual, and geospatial leader. Arrives, household characteristics. That capture things that are Central near kind of traditional criteria. That would be used in underwriting. Such as the number of assets that our own economic activity, in the area at cetera. So then we use a variety of models
to process that data starting with segmentation and clustering using unsupervised learning to translate the transaction data into something that can be more appropriately modeled or predicted and then building prediction models to actually look at, you know, with a more limited perspective of data on how can we predict new customers and revealed at the key drivers for their repayment outcome and then kind of more from an analytics perspective. What does that mean in terms of financial forecasting and the relative risk downside risk of that
portfolio? So this is an example of one of the geospatial data layers that we work with the percent of household with finished floor. So this is a roster or representation of a roster layer of each of these pixels represents the different encoded value on. So the the this is what the percentage of households in one square kilometer Bridge. That's the resolution of the raster, the percent that have finished Warren. So that's a pretty good indicator of relative will status than an asset.. How cold has we? Have a stack
of these rosters that we use. You can see some of them below but this help solve for one of the key challenges of working in the office space or Bill are using data science for this type of Youth case and not sin in assembling a training set, typically companies that are trying to do. Data-driven decision-making, would have to collect this data on their own using field surveys, which is extremely costly. To do it on a kind of a wide enough perspective. Many of these companies are working across multiple markets would be extremely expensive and and many of them are
startups themselves, but don't really have that type of capacity. What are the other challenges is a lack of unique identifiers or granularity across state is that? So even if you did have multiple household surveys, how to compare them, how to spell harmonize them and make them consistent. The advancements. That we have really enabled us to have that very high degree of of granularity and understanding the social and economic characteristics that really a neighborhood level and augmenting, the data that might be collected through the company's existing
registration process. Thinking that about how do we label this data? So the core question, if you think about what is it? What's the expected, cash? Flows are gorgeous, pected. Performance of a given type of account is actually quite complex in the pay-as-you-go structure. So there's no set term or, you know, they're sort of unexpected term of how long I'm a customer, should repay. And there's an expected amount that they should repay each day or each week, but technically, because
it's pay-as-you-go, that is not a hard-and-fast rule in terms of considering them into fault. If they miss one of those payments each company that we work with dogs have their own definition of default. And ultimately enough payments are about that system will be repossessed. But there is a lot of inconsistency of what the definition is in front of machine learning perspective, you know, it's difficult. It's a very arbitrary process if you were just to take Shut off at a certain point and and say, you know how they just posted by 1X times, the normal repayment term kind of the, the, the
expected value or should it be sometime later on the week? Use an approach that actually looks at the trajectory of the repayments on and use unsupervised learning to Cluster those normalize first. And then cluster the receivables timeline. And what we say is a really consistent group into three main types of repayment Behavior. So we have the fast repairs That Bass. Please follow the nominal contract expectations. So, they pretty much pay off in time. Although a little bit delayed often. Then there's the moderate repairs, which
are still engage customers. They haven't turned, but they may have extended, maybe because of a growing season if there are farmer and that's where they're getting most of their income from or they have They're using it as a substitution for the grid, that might be very unreliable. So there's a lot of factors that could lead them into that and then we have the the full repairs which are those that are the highest risk and really never listen to Flatline in their payments. So this helps normalize across a number of different factors and translates into a label that we can
then. Do you spell prediction modeling? So we have, and we've done this in a number of countries. We focused a lot in Kenya and in Nigeria since that's where this solar home systems model, has really taken off the most. And so when we have done, this means we kind of use those combination of features. The payment state of the individual account data, that was collected through the registration process. And then this augments the neighborhood geospatial profile of characteristics to be able to predict the outcome out of very kind of early
point in the account history, sometimes, even before the account is even originated, we can do it, just based off of the geospatial characteristics and the relative risk their. So using this, we've been able to improve the production, Precision by two to three times first as if we were just assumed the kind of Tire Distribution of default. And so this is one case where you do, stay some variation. It's easiest to Britax. Extreme the fast repairs that are kind of your best customers and the slow repairs. That are your worst customers a little bit harder to project on
because there are a variety of different factors influencing out for the middle. But again, this is a significant improvement over what the company's current Baseline is which is essentially, you're going in blind, doing very little underwriting. If any of their customers. So, what does this look like in action, but I wanted to highlight a given. The focus of this session are targeting work, which is the most kind of geospatial e oriented. And if there's an example of how we layer those different geospatial characteristics and the prediction model that the
segmentation a labeling for the, the transaction history to arrive at kind of a priority station of the answer. That first question. I talked about him or where are the ideal customers for these types of solar home systems product. So this one, we're looking at a company that would be marketing. Fuller, TV products on, which has been extremely popular, especially in the covid Europe where we've had a lot of household really needing that connectivity for entertainment in the other spending more time at home during lockdown, in the lot of these places, but also for information and
getting an accurate updates on the situation across the So here looking at a couple of different factors that would qualify or region for kind of being a good area for solar homes with some TVs. So high in a population density of a high enough number of kind of top consumer household, which is reflecting kind of their asset ownership and education levels. And then looking at areas that have a low rate of existing TV. Because, you know, we want to avoid existing products that situation. But then, getting more granular, we can look at
differentiating, the man's by repayment. Refund basic solar home systems on so we can do that prediction model kind of looking at where their densities of a predicted probability of high repayment or kind of that that Vin Diesel full alarm system and leering on where their identities of regular TV viewer. So that there is actually demand for MTD products. Maybe they're going to neighbor's house or two. A bar to watch TV, but they don't have one there themselves. So using that we can create a prioritization of communities. And again, each of these
squares is representing a square kilometer. And so those can be grouped into sales territory that companies can can send you know, their agents to go prospects and to kind of improve their efficiency of operations. So, this is a little bit different of a perspective on here, looking at a productive use products. So things that can generate economic benefits for the household on, which is it really important priority. And especially, if we think about
you, how can we identify products and areas where the connections to, I'll go to electricity to having energy. Cannot only can basically pay for itself, right on my Pick, N improve their income generation capacity. It can become even more affordable. So obviously if the drivers here, we're looking at a solar water pumping product, which is a little bit higher value and I'll be asleep really pertinent to small, holder Farmers to can you know, it's a big improvement over manual manual irrigation techniques
or not having irrigation. So looking at things like I really do levels and their ideal precipitation levels for agricultural culture. Looking at from, from survey data. We have got that from satellite imagery. From survey data. We can understand what they do. This thing water sources that are used by Farmers again, then we can later on. Credit projection of where we expect to see high credit worthiness and high repayment rate. And then looking at kind of filtering by population, density to filter out, kind of dense, urban areas to really focus on those
agricultural hubs. But this is a project that we did to help on actually government entities identify areas where it would be important to provide one of these programs that support company to roll out for water pumping. In this case, in Nigeria, and identify can of that high market suitability for solar water pumps. And then finally refused Products near we're very exposed Gaston energy access but really the core of the methodology can be extended to other sectors
Beyond energy in a really anything that's related to asset financing because we have this wealth of geospatial data. That's broadly applicable. I can also be very specific to specific types of profiles of clients. I'm so we worked with and organization called 1-acre funds. They are pretty prominent Farmers support organization to provide financing for farming. Inputs such as fertilizer, their speed on what you need to do finances in advance, but then farmers can use their proceeds of the Harvest to pay back. So they expanded into into a new
place. I'm in Nigeria and they wanted to better understand what the the composition of their clients were in this pilot and the repayment pass that they could expect so that they could structure. Financing terms more effectively. So using our 7th Asian approach, we were able to identify four groups, two of which were representing kind of core Target customers, those that were either setting repairs and we found through the correlations with her geospatial data that they were likely to be salaried. And therefore, had kind of an incremental income stream
versus those that were truly kind of farming, dependents and needed to wait for the Harvest funds. And that was highly correlated with the types of crops. We saw that were grown in their areas to use the second of three screen their customers. I'm in different areas and set them on these different financing past, that best fit their income streams. I'm here also to able to use an indexing approach, a combination of the most productive roster layer. So I did identify the likelihood of ability to pay and establish a market size. For i o a s expanded into new areas within
the state in Nigeria, what they should expect. Kind of the fair market 5 to be and and then actually prioritize specific communities to go into do a couple of thoughts of a future applications would also considerations again, thinking back as mysterious position as a social impact, start up and using their Define for good on the way, you know, we think about how do we use this. This boring application for entities Beyond just kind of the the the company is kind of portfolio Management on which is really important one. But thinking about, how can
this be used to tug? Really Target concession with suspense? No subsidies or incentive programs financing or grants available through the government or through Foundation, to really Target. Those most vulnerable households, the still repair that, you know, might be in areas that are subject to climate change or have all their kind of poverty. Indicators that are really high so that we're not just kind of Terry pitching off the top, the best kind of commercial customers, but ensuring everyone has access to electricity. After working towards a really granular application of building it into
point-of-sale credit scoring for the field sales teams, which ties into I think the second set of questions of you know, how are we ensuring that these data science tools reduce inequality and not reinforce it. So again, thinking about that. How do we connect the right type of product is Affordable to all types of household, and connects the kind of Grant subsidies and assistants. That does exist to those that need it. Most. You in these Emerging Markets with increasing digital access, as they've introduced me to these new types of
digitized tools. How do we ensure that that data is being used responsibly and ethically? Especially as there are kind of involving data privacy regulations that might be not as well and horses in more developed markets as soon as you've taken a role in you. No answer. That the data that we were using, it's fully anonymized to the extent possible and kind of advocating for that data privacy and protection. I think that we are. I hope I am on time, but I want to make sure there. Is there any questions
and give a quick caveat. We are actually hiring for position. So if anyone is interested in learning more about what we do would recommend going to a website or reaching out to me or hiring for data scientist position and would love to chat with you.
Buy this talk
Buy this video
Our other topics
With ConferenceCast.tv, you get access to our library of the world's best conference talks.