Digital Signals & Access to Finance in Kenya

Landscaping study on the
potential to gain insights on
Kenyans’ access to finance
through analysis of big data
USAID DEVELOPMENT CREDIT
AUTHORITY & UN GLOBAL PULSE
PROJECT REPORT, SEPTEMBER 2013

1

Executive Summary
In emerging markets, eight out of ten small businesses cannot access the loans they need to
grow. USAID’s Development Credit Authority (DCA) uses risk-sharing agreements to mobilize
local private capital to fill this financing gap. The goal of this collaboration between UN Global
Pulse and USAID is to explore how big data could support the work of USAID’s Development
Credit Authority. Kenya has become an established tech leader in Africa in recent years –
generating greater volumes of digital data as a result. The goal of this study is to explore what
new sources of digital data, and methods for analysis, could be helpful in answering the
question: “What barriers to accessing loans do small businesses in Kenya face?” Accordingly,
this report paints a picture of the big data landscape in Kenya, shows preliminary findings, and
lays the groundwork for further investigation.
The study began by looking at baseline data for ICT and social media usage rates in Kenya. By
interviewing a sample of DCA’s clients – farmers from the region around Nakuru and Nairobi –
Global Pulse created a taxonomy of colloquial language used in relation to accessing loans and
finance. Global Pulse’s analysts then used this taxonomy to mine available data from Twitter
and Google search trends, to understand online conversations and information-seeking behavior.
It’s clear that in Kenya’s current digital landscape, the quantity of relevant social media data
restricts its utility. For example, the study found that in July 2013 there was an average of only
about 10 tweets related to “loans” on Twitter, per day. The relatively small numbers of relevant
digital signals available indicate that in the short-term, social data can likely only provide
supplementary insight about barriers to finance in Kenya. For instance, online analysis could be
useful for revealing early themes or trends, to inform the topics of focus groups to validate.
Despite the relatively small numbers of contextually relevant tweets available in 2013, there is
an emergence of a Kenya-specific Twitter culture. In particular, Twitter is being used to seek,
access and share information about loans, especially mobile loans. The introduction of a savings
and small loans service called M-Shwari to M-Pesa customers in Kenya has already sent ripples
through Kenyan social media.
The study also conducted some initial tests with Google to understand finance related
information-seeking behavior in Kenya. Events that are broadly shared (i.e. the beginning of a
school semester or elections) are clearly visible in the data. This could mean that other systemic,
or anomalous, events will be visible as well. However, the analytical tools made available by
Google for Kenya data are limited - currently it is only possible to know the relative volume of
searches, rather than access real numbers.
Finally, in addition to this analysis, the study looked at the digital footprint of the Development
Credit Authority, its clients and priority sectors, and offers a menu of what other relevant sources
of digital data might exist in Kenya, and how the data could potentially be accessed.
This landscaping study provides insight into the challenges and opportunities of developing
methodologies in a region where use of social media has not yet reached saturation point. For
example, setting up monitoring systems at this early stage could avoid the need for reverse
engineering. Online discussion and search behavior around access to finance is on the rise in
Kenya; and in order to gain early insight or warning of emerging trends, further work is required
to turn the possibilities for big or small data analysis into meaningful means of operation.

2

TABLE OF CONTENTS
4

Introduction

6

Background on Digital Landscape in Kenya

10

Trends in Social Media

28

Google Trends

32

Finding the Digital Footprint

36

Recommendations and Next Steps

ANNEX 1: Survey Questionnaire

Acknowledgements:
The lead researcher and author of this report was Eva Kaplan (eva.kaplan@gmail.com). It was
prepared with the support of the team at the United Nations Global Pulse
(www.unglobalpulse.org), the USAID Development Credit Authority (DCA), and USAID-Kenya.
Additional acknowledgements are due to Anoush Tatevossian, René Clausen Nielsen, and Robert
Kirkpatrick at the UN Global Pulse; Stephanie Grosser at DCA; and Linda Kagota at USAIDKenya. Thank you also to Esther Ajambo for implementing the survey and Clarise Aduma from
Kenya Commercial Bank for generously working with us to find clients to survey. The study was
made possible through a Learning Improvement grant through USAID Training Support Fund.

3

INTRODUCTION
The goal of this study was to better understand how big data can support the work of the
Development Credit Authority at USAID. This is landscaping study and as such is intended to
point to areas that might be appropriate for further investigation. As an initial foray, the analysis
was confined to Kenya, and is targeted specifically to the question “What barriers to accessing
loans do small businesses in Kenya face?” It is important to note that the purpose of this
exercise is not primarily to answer that question, but rather to determine the feasibility of
answering such a question using new sources of digital data. The purpose of this report is to
paint a picture of the big data landscape in Kenya, show some preliminary findings, and lay the
groundwork for further investigation by highlighting possibilities and challenges.
The report has two primary goals:
Describe the digital landscape of Kenya: Section I of the report lays out baseline data of ICT
and social media use in Kenya with a particular emphasis on understanding gaps in baseline
information.
Describe big data in relation to loans in Kenya: Section II and III dig in to available data—
namely Twitter data and Google search trends—to better understand online conversations
and information seeking behavior. Section IV takes the opposite approach, looking not at
what data is currently available but rather at the digital footprint of DCA clients and priority
sectors. Section IV first summarizes information collected from DCA clients themselves, and
then uses two case studies—agriculture services and bank websites— to look at where else
relevant digital signals might exist in Kenya.

What is Big Data?
“Big data” is typically defined as terabytes or petabytes of data that is generated as people go
about their daily lives in the digital age. Big Data is characterized by the “3 Vs:” greater volume,
more variety, and a higher rate of velocity. This includes information seeking behavior (i.e.
Google searches), social data (i.e. Twitter or Facebook data), mobile phone data, and digital
services (i.e. Amazon purchasing or mobile banking).
Big Data is both the information that is passively generated as by-products of people’s everyday
use of technologies and the information people willingly communicate about themselves on the
web. Even when individuals do not have direct access to mobile phones or other technologies,
they may still be passively emitting information as they go about their daily lives (e.g., when they
make purchases, access services, or when they interact with better-connected members of the
community). These digital trails, or “smoke signals” can reveal changes in trends in our
collective preferences, struggles and overall well-being. Big Data for Development refers to
Global Pulse’s effort to determine how big data can be used to better inform the policy and
planning of development programs.1

1

For reference, UN Global Pulse’s introductory guide, “Big Data for Development: A Primer,” is available online and
for download at: http://unglobalpulse.org/bigdataprimer

4

In general, sources of Big Data that can be useful to development are those that can be analyzed
to gain insight into to human well-being and change. A preliminary categorization of sources may
reflect: WHAT PEOPLE SAY (i.e., international and local online news sources, publicly accessible
blogs, forum posts, comments and public social media content, online advertising, e-commerce
sites and websites created by local retailers that list prices and inventory) or WHAT PEOPLE DO
(i.e., passively collected transactional data from the use of digital services such as financial
services (including purchases, money transfers, savings and loan repayments), communications
services (such as anonymized records of mobile phone usage patterns) or information services
(such as anonymized records of search queries).2
In some cases, the volume of data that could be analyzed to augment development planning may
not be technically be “big” (for example, many of the digital services listed in the agriculture
case study in Section IV do not generate massive amounts of data, relatively speaking). In these
cases it might be more appropriate to use the label “new data” or “digital data.” However, the
term Big Data has evolved to be a catch-all phrase describing the new age of digital data, and
the innovations in technologies and methodologies now available to collect, store, and make
sense of it all.

Methodology & Purpose of Landscaping Study
In order to fill in gaps in our understanding of the big data landscape in Kenya, the study was
undertaken in different phases, which are reflected in the format of the report.
First, the report summarizes who in Kenya is generating digital data, looking specifically at what
is known about the penetration rates and demographics of mobile, Internet and social media use
in the country. In addition, to find out how DCA’s target population (underserved entrepreneurs)
generates digital data we completed a non-representative survey of 10 DCA clients. All the
clients were from Kenya Commercial Bank, and all of them were farmers. They were all from a
similar region of Kenya, between Nakuru and Nairobi. The clients were asked about (1) words
they use to talk about loans in informal and formal chatter; (2) how they use mobile phones,
computers and social media; and (3) how they find information about their business, their
existing loans, and new loans.
Second, the report explores finance-related data generated through use of the Internet. This
analysis primarily relies on Crimson Hexagon’s proprietary social media analysis platform, called
Forsight™. We also make use of the publically available tools on Google Trends.
Finally, in the last section we provide examples of other big data methodologies that might be
useful to DCA, exploring what other types of digital data exist.
This report is intended to inspire new thinking in how DCA can use new sources of digital data to
inform its work. Big data may help researchers ask the same types of questions they have
traditionally researched in other ways—for example, by better understanding people’s behaviors
based on observation. However, it may also facilitate entirely new ways of thinking about issues.
For example, big data is often particularly well suited to understanding networks or information
2

Excerpted from “Big Data for Development: A Primer” (June 2013) http://unglobalpulse.org/bigdataprimer (page 3)

5

seeking behavior. In addition, it is worth noting that while some big data may be well-suited to
track individual behavior, Global Pulse recommends that at this stage responsible uses of big
data analytics for global development should target aggregate trends rather than individuals to
protect the privacy of individuals3.
Being imaginative about big data is thus important, in two key ways—first, being imaginative
about where to “look” for relevant data sources; and second, being imaginative about how to use
the information. While big data might not provide a silver bullet answer to key questions faced by
DCA, it might be useful in ways that were not anticipated.

BACKGROUND ON KENYA’S DIGITAL LANDSCAPE
Over the past 5 years, Kenya has emerged as a tech leader in Africa. The rise of mobile phone
use in Kenya is a well-told story, and the mobile money service M-PESA has changed the way
many entrepreneurs globally think about providing services, from finance to health to energy
provision.
In order to understand the implications of the Kenyan digital revolution for big data analytics, it’s
important to understand in greater detail the demographics behind these trends. However,
despite the great deal of interest in Kenya’s digital landscape, the exact make-up of who is
accessing services is not very well understood. The purpose of this section is to explore what is
known and not known about the demographics behind Kenya’s digital revolution.
Mobile and internet penetration
According the Communications Commisson of Kenya, as of December 2012, mobile phone
coverage in Kenya had reached a total of 30.7 million subscriptions, or 78.0% of the adult
population. This represents a growth of 19.6% percent since 2009, when 49.1% of the adult
population had a mobile phone. As of September 2012, it was estimated that only 7% of these
phones were smart phones, but this sector is projected to grow rapidly, with Safaricom’s launch
of a new smartphone in January 2013 selling out in less than two weeks after launch. Data for
internet accessed through the mobile networks can be purchased using the same pre-paid card
that is used for mobile phones.
Internet, by contrast, stood at 9.4 million subscriptions as of December 2012. While this
number clearly lags significantly in comparison with mobile penetration, it represents a growth of
75.1% of internet subscriptions over the same period the previous year. The vast majority of
subscribers access the internet via mobile networks, representing 99% of total internet
subscriptions. This includes both those who access the internet directly on their phones, and
those who use mobile plug-ins to computers. Including non-subscribers, it is estimated 16.2
million Kenyans, or 41.1% of the population, were accessing internet by December 2012.4

3

For more information on UN Global Pulse’s privacy and data protection principles and recommendations, please visit:
http://unglobalpulse.org/privacy-and-data-protection
4
Communications Commission of Kenya, Quarterly Sector Statistics Report: First Quarter of the Financial Year
2012/13, http://www.cck.go.ke/resc/downloads/SECTOR_STATISTICS_REPORT_Q1_12-13.pdf

6

Social Media
Data on social media is less readily available. In a survey done in 2009 by AudienceScapes5,
20% of those surveyed reported having used internet in the past year. Of those who had used
internet, 71% reported using some sort of social media, and 50% reported using social media at
least once a week.
A study by Portland Communications6 in 2011 which analyzed 3 months of tweets geolocated to
sub Saharan Africa (representing 11.5 million tweets) showed Kenya as the second most prolific
African country on Twitter, after South Africa. Africa-wide trends show that 81% of these
“tweeps” are using Twitter to engage in social conversations. Sourcing information about news
was also highly ranked. Indeed, across the continent, many Twitter users reported that Twitter
has become a primary source of information for them, particularly for international news.
According to Facebook, as of May 2013, there were 1,934,980 users in Kenya. These represent
those with Facebook accounts who have identified themselves as living in Kenya.
Mobile Services
Globally, Kenya stands out as a hub of innovative mobile services. These services are highly
relevant to big data analytics because they leave digital trails. As people interact with mobile
services, they are generating valuable data on anything from saving rates to food prices to health
problems. This data is held privately by service providers.
M-PESA—or mobile money—is the most salient example of Keyna’s mobile services. M-PESA
allows people to transfer money via their mobile phones. A plethora of financial services have
also been built on top of the M-PESA platform. These have largely failed to have broad appeal.
However, Safaricom’s new paperless banking service M-Shwari is experiencing rapid growth.
Launched in November of 2012, Safaricom reported that M-Shwari had 1.6 million customers
after just 2.5 months. M-Shwari allows customers to save money and collect interest. It also
allows them to access emergency loans.
In addition to M-PESA related services, other types of information is transmitted via mobile
phone. For example, information systems such as mFarm or iCow allow people to share
information on their farming practices over the phone.
As mentioned, this data is not publically available. However, Section IV will explore the potential
of analyzing the data generated by these services, as well as what would be required to access
them.
Demographics
If big data is to be used for designing policy or programs—for example, for understanding unmet
financing needs—knowledge of who is represented in the data would be paramount. While
information on ICT penetration is published quarterly, the demographics of who is using ICT is

5

AudienceScapes, Internet Access and Use in Kenya, copyright Intermedia 2010, found at
http://audiencescapes.org/country-profiles/kenya/media-and-communication-overview/internet/internet-322
6
Portland Communications, How Africa Tweets, July 2012, found at http://www.portland-communications.com/wpcontent/uploads/2013/05/Twitter_in_Africa_PPT.pdf

7

published in an ad-hoc way, typically by researchers doing a one-off survey. Uptake of new
technology is Kenya is happening faster than this type of research is apparently keeping up with.
Some impressive attempts to better understand the demographics of mobile use have been
made. In 2009, a Financial Sector Deepening Kenya (FSDK) survey in 646 communities in
Kenya sought to give shape to mobile trends, gathering demographic data on phone ownership
and phone usage and sharing. In general, counties with higher urban populations reported higher
phone ownership.
The survey found that there was some level of mobile phone ownership in every income bracket,
although as might be expected there was a higher concentration of ownership at higher income
levels. It was also found that women are more likely to be phone sharers than men. Among those
who had never used a phone, the vast majority (81%) were women.
Except for those under 17 and those over 60, the survey found comparable rates of phone
ownership across age brackets.7
While it appears that in the past year or so growth in mobile phone access may be leveling off,
growth in other ICT sectors—for example smart phones— continues to expand rapidly. Thus, it is
unclear how relevant evidence from 2009 is for today’s digital landscape.

Figure 1: Phone ownership across Kenya (source: FSDK)

7

Wesolowski A, Eagle N, Noor AM, Snow RW, Buckee CO (2012) Heterogeneous Mobile Phone Ownership and Usage
Patterns in Kenya. PLoS ONE 7(4): e35319. Doi:10.1371/journal.pone.0035319

8

Demographic information on Twitter usage is not clear and much less readily available. Again,
ad-hoc studies for particular purposes may shed light on Twitter demographics. For example,
iHub Research is seeking to uncover trends on who was tweeting during the 2013 elections. The
study, once completed, is likely to be illustrative; however it does not represent a comprehensive
assessment of demographic trends of Twitter users.
Relevance for big data analytics
All data has bias, and big data is no exception. However, without up-to-date demographic
information, it is not possible to understand biases in the data. There might be some areas of
investigation where having demographic trends is not critical—for example, bank-specific
information like interest rates or different types of loan instruments. However, for some analytical
questions, gathering baseline information on DCA clientele may be important.
The survey completed for the purposes of this report is instructive. While it only covered 10
clients, all farmers from the same bank, it does offer some clues as to the digital footprint of
DCA clients. The table below describes how many respondents use digital tools.
Owns a
mobile
phone

Owns a
computer

10

3

Has
accessed
internet on
mobile
phone
4

Has
accessed
internet

Has
Facebook
Account

6

2

Has
discussed
loan or
business on
Facebook
0

Has a
Twitter
Account

Uses
Mobile
Banking

0

8

Table 1: Number of DCA clients who reported using digital tools.
While it is impossible to extrapolate broader trends from this data, this indicates that social
media may be limited in its ability to represent these clients.

9

TRENDS IN SOCIAL MEDIA
Through an existing partnership with the company Crimson Hexagon, Global Pulse is able to use
Forsight™, a social media monitoring and analysis platform, which allows access the full
“Twitter firehose” (that is, all Tweets being published today, in real-time), and archival Twitter
data going back to 2009. With the ForSight™ platform, it is possible to build “monitors,” which
filter the billions of Twitter posts according to particular specifications.
For this analysis, we built monitors to do the following:
•
•
•
•

Estimate the country of origin of each Twitter post, and narrow the analysis to Tweets
originating from Kenya
Filter out Twitter posts with content/keywords related to the specified issue area (in this
case, loans)
Sort the posts according to specified sub-categories (for example “business loan,”
“personal loan,” etc.)
In some cases, we also did “sentiment analysis,” sorting posts according to positive and
negative sentiments.

In this section, we describe the step-by-step process of building the monitors, along with
analyzing the results. The process of building a monitor involves first developing a list of
keywords to filter out relevant Twitter posts. Then, subcategories are defined and refined based
on available data. The last step is to “train” the monitor to automatically categorize based upon
the specifications. This involves hand-sorting roughly 20 posts per sub-category as model
examples. The machine will then automatically categorize the remainder of the posts (and
capture new Tweets as they are published in real-time) once it is trained. To provide a clear
demonstration of this process, the process of building and analyzing the first monitor is
described in depth.
To analyze the data, this report largely relies on the tools available in ForSight™. After the
monitors are built, there are several types of investigations:
•
•
•

Changes in how Twitter is used over time.
Categorization of general Twitter chatter, including narrowing down to investigate
particular tweets.
Exploration of anomalous events, represented by sudden spikes in the volume of relevant
Twitter data.

10

Throughout this section, it is important to keep in mind the initial question: What barriers to
accessing loans do small businesses in Kenya face? However, the point of this exercise is not
exactly to answer the research question, but rather to describe the type of conversation related to
finance that are taking place on Twitter—to investigate whether, and where, information relevant
to this question exists in social media.
To this end, the starting point for analysis is broad, and seeks to scan social media in different
ways, looking for data clues. This question is approached from 3 vantage points, using three
distinct monitors to filter Tweets related to each of: (1) General Loans; (2) Loans by Sector; and
(3) Banks. In each area, the general goal is to understand what types of loan conversations
Kenyans are having on Twitter. In order to do that, we used the survey to better understand (1)
the words people use to discuss loans; and (2) the types of information people share in the
public sphere.
When survey participants were asked who they discuss business loans with, most people said
they discuss loans with loan officers and families. While some of those surveyed cited friends as
a source of information about their business, several people explicitly stated that they do not
discuss loans with friends due to issues of trust/privacy. If it is indeed the case that people feel
that discussing loans is a private endeavor, this will impact what we can expect to find in Twitter.
Having said that, Twitter is a dynamic space, and usership in Kenya is growing. What we have
discovered in this initial foray is likely to change over time as digital cultural norms evolve.
Vantage One: General Trends in Loan Chatter
The first monitor looks at general loan chatter. The goal was to get information on all loans
except for university loans8.
Building the taxonomy
A taxonomy is the list of keywords that is used to extract relevant posts from the full Twitter
dataset. The taxonomy has two goals: (1) to include relevant posts and (2) to exclude irrelevant
posts. When building the taxonomy, it is important not only to have a list of keywords that the
post should include, but also to exclude particular words that result in irrelevant posts.
Step one: survey bank clients to identify relevant keywords
The first step for building this monitor was to survey 10 bank clients to get clues about the
language that they use to discuss banks. This was done by asking the clients to recall specific
conversations with bank officers, friends and colleagues. The field officer prompted the clients
to recreate these conversations as she took notes on the keywords that were used. These words
formed the backbone of the initial taxonomy.
Step two: test and refine taxonomy
The initial requirements were that the post had to include one of the following: loans, loan,
mkopo, wakopo, financing, finance, or credit AND one of any of the keywords that DCA clients
mentioned in the survey.
8

Please note this only includes loans from Kenya’s university student loan authority, HELB. Loans from private banks
for education are not explicitly filtered out.

11

This did not have the desired results—many of the mined posts were not related to loans. Upon
further investigation, the taxonomy was revised in two ways.
First, words related to finance and credit create a lot of “noise” (or, irrelevant results) in the
data, without returning relevant posts. Most financing chatter refers to government or corporate
finance:

Barclays Bank - Kenya, signs $13m financing deal with Marriot Drilling...
http://t.co/rk1sLbyJqu
Credit in particular creates a lot of noise. Most credit-related tweets refer to phone credit:

Yes @Nettoh_ken Yes. You can take home the credit, give to a friend, donate, purchase
bundles. Whatever dya want 2do @NakuruRFC @wanyore
An alternative definition of credit was also frequently used:

Politics is about seeing an opportunity, use the masses to achieve and taking the credit
for it.
@shisilo @shizphotography Thank you we'll credit and tag your photo today.
Where the word credit does appear in a relevant manner, it typically also includes loan. The
words related to financing and credit were thus removed from the taxonomy.
The second challenge was that the keywords generated from the surveys did not return many
results. To refine the taxonomy, we took out the second set of keywords to see all of the posts
related to loans without any additional specifications. We extracted additional keywords from the
resulting posts that were relevant. This resulted in several additional words, but the words that
made a significant difference were fairly simple—first was “bank” and second was “mshwari,”
the mobile loan service. We also excluded common keywords that resulted in irrelevant posts, in
particular those related to sports and student loans.

Step Three: Pre-viewing results and further refinement
At this point, the data was fairly clean, and there was enough data to train the monitor
(described below). Once the categories were set and the monitor was run, another challenge
arose. Because the daily volume of relevant posts is low, one joke or retweet can create a lot of
noise in the data. For example, one Twitter user posted the same post 40 times over the course
of a day, creating a dramatic spike in the Twitter data for that date. This user was thus excluded
from the data. The following tweets in particular were shared widely, creating an outsized imprint
in the data.
I want the bank to do two things for me.....give me a loan and leave me alone
#STOLEN Hooker walks into an Equity branch... "nataka loan ya kupanua biashara."

12

It is important to note that if the overall number of tweets were higher, even widespread retweeting would not throw the results. On the flip side, because of the low overall volume, it’s
possible to easily find these disruptive tweets and exclude them.

Final taxonomy
For the purposes of this report, the final taxonomy is below. If USAID and Global Pulse continue
to use this monitor, it is likely that further rounds of refinement will be required.
(loan OR loans OR mkopo OR wakopo) AND ("Top up" OR "Payback period" OR
installments OR expansion OR mpesa OR mbesa OR financing OR "business
financing" OR biashara OR dairy OR msoto OR red OR doh OR qualify OR stocking
OR application OR maximum OR duration OR interests OR delay OR security OR
"land title" OR deed OR "deposit dates" OR tembelea OR "fixed deposit receipts" OR
secured OR "calculated interest" OR interest OR guarantees OR guarantor OR lawyer
OR Agricultural OR agriculture OR development OR application OR procedures OR
payback OR improvement OR n’gombe OR wakora OR repay OR balance OR
"agreement letter" OR period OR clear OR siri OR security OR sambaza OR
defaulted OR "cooperative society" OR Faulu OR credit OR Agrovets OR mfugo OR
zidisha OR "penalty charges" OR penalty OR Emergency OR "ketes temiship" OR
inflation OR expectations OR capital OR terms OR payment OR "nilitemelea banki"
OR farm OR status OR assets OR asset OR mshwari OR land OR animal OR animals
OR "long term" OR "short term" OR "mini statement" OR "mini statements" OR
ministatements OR "shamba shape ups" OR "fixed accounts" OR mshwari OR
zidisha OR bank OR banki) AND -helb AND -@MweuDeh AND -hooker AND @helbpage AND -Hooker AND -@HELBpage AND –“car-jacker” AND -Chelsea AND
–Manchester
Defining the categories
Step one: detailed sorting
In the first round of establishing the categories, we started with 6 categories and two subcategories:
I want a loan
-Business
-Personal
I have a loan, negative
-Business
-Personal
I have a loan, positive
-Business
-Personal
I have a loan, neutral
-Business
-Personal
Information seeking
Information provision
Step two: train the monitor and refine
After setting the categories, the next step is to train the monitors. This means sorting posts by
hand into the relevant categories. This “trains” the machine to automatically categorize the rest
of the posts.

13

When we tried to sort the posts into these categories, it became clear that the categories were too
specific, and that most of them did not have enough data to use for training the monitor. We first
removed the distinction between business and personal, because, using this taxonomy, most of
the loan related tweets refer generally to loans, and do not distinguish between business and
personal loans.
Next, there was not enough data to clearly distinguish between “I want a loan” and “I have a
loan.” For example, the loan below does not relate to either, and rather is a general comment on
M- shwari:

Ati Mshwari can only allow you loan equivalent to what you have in your account, why do
you need the loan in the first place,is u withdraw
Final categorization
The final categorization is therefore rather general:
General Loan, positive
General Loan, negative
Supplying information about loan
Seeking information about loans
Jokes and news items were excluded.
While this monitor does not allow for specific categorizations, for the general loan monitor it was
important to cast as wide a net as possible on chatter related to loans. This monitor is useful for
tracking the nature of chatter related to loans on Twitter, and to observe how it changes over
time.

Results
Figure 2 below represents the volume of tweets from January 1, 2012 and August 25, 2013. As
is clear, there is a steady rise in overall chatter over this time. Those posts are also becoming
more relevant, meaning they are less related to things like international news (ie. World Bank
loans to countries). On January 10, 2012 there were 2 relevant posts of 10 total posts. On
January 10, 2013, there were 19 relevant posts of 28 overall posts.
Despite this growth, the volume is still very low. In July of 2013, there was an average of 10.7
posts per day.

Figure 2: Volume of tweets, generally about loans, from January 1, 2012 and August 25, 2013

14

Categorizing general chatter
It is clear that much of the growth in chatter about loans is related to M-shwari. Figure 3 below
represents the tweets driving conversation in the month of July 2012 (before M-shwari was
launched) and July 2013 (after M-shwari was launched). The inner circle represents the top
words that co-occur with “loan” over the specified period. The outer circle displays the top words
co-occuring with those words, in effect automatically creating a set of sub-categories. For
example, “bank” was the top co-occurring word in July of 2012, with popular discussions
breaking down further into 5 sub-categories, including “school fees” and “loan to start.” By
contrast, the top topic in July 2013 is “Mshwari loan,” with five sub-categories. The overall
number of relevant tweets in July 2012 was 82, while in 2013 it was 336.

Figure 3: Key trends driving Twitter chatter from July 1- July 31, 2012 (left) and July 1-July 31,
2013 (right).
More specifically, growth is even more pronounced after Safaricom launched its customer care
service “Okoa Fresha,” which includes direct engagement over Twitter. The service, launched in
early February 2013, created several days of increased twitter activity in relation to loans.

Figure 4: A rise in chatter about loans upon the February 2013 launch of Safaricom’s customer
care service, which includes direct engagement with customers over Twitter.
What is more, the general chatter about loans changed based on the new services. As can be
seen in Figure 4 above, prior to the service being launched, most conversation on twitter was
general commentating on loans, while after the service was launch there is a rise in information
seeking and information provision. Much of this growth is not due to Safaricom’s informational
posts, but rather to retweets of Safaricom’s informational posts.

15

Figure 5: Categorization of general loan tweets between January 1, 2013 and March 14, 2013.
The type of chatter, and by extension the type of analysis thus changes with the launch of Mshwari, and even more so with the launch of Okoa Fresha.
That said, of course the data is not exclusively related to M-shwari. The clustering below shows
the breakdown of general chatter between March 28, 2013 and May 13, 2014. This is after the
launch of Okoa Fresha, and represents a period with no dramatic spikes.

Figure 6: Main words driving Twitter loan chatter in Kenya between March 28 and May 13,
2013.

16

This shows that the nature of general chatter changes over time—in this case due to the
relatively large Twitter presence of Safaricom.
Investigating anomalous events
In addition to the prolonged increase in Twitter volume explored above, there are two dramatic
spikes in this period.

Figure 7: Anomalous behavior in Twitter. The date circled in blue is November 27, 2012. The
date circled in red is December 22, 2012
The first spike, on November 27, is the launch of M-shwari. The figure below represents the
nature of the conversation on November 27-28. As can be seen, almost all of the conversation is
about Safaricom and M-shwari, with a subset discussing M-shwari interest rates.

Figure 8: Clustering of relevant content from Twitter on November 27-28, 2012.
The second spike, on December 22, 2012, came with the announcement by vice- Presidential
Candidate William Ruto that if elected, the government would offer an interest free loan to
women and youth. As is clear from the break down below, many of the tweets were neutral in
nature, related to sharing the news. However, it is also clear the announcement was met with
some skepticism—in addition to “gullible” and “silly season,” one popular retweet was “Ruto
anafikiria sisi ni wajinga” or, in English, “Ruto thinks that we are fools.”

17

Figure 9: Clustering of relevant content from Twitter on December 22, 2012.
Investigating particular tweets
With the low overall volume of tweets, it is possible to go through the sorted tweets of relevance.
Some tweets do shed light on barriers to accessing loans:

I need a bizness loan..interest is double
Understanding the positives and negatives surrounding bank loans help people make the
wisest decisions when they need loans.
The youth and women have no skill nor security to secure the loans that they need to start
enterprises #OrangeNDC
However, there are not enough of these tweets to discern clear trends, and without demographic
information if they represent the voices of underserved entrepreneurs.

Vantage Two: Loans by Sector
This portion of the analysis was focused on setting up monitors related to various sectors. In the
first attempt, we concentrated on the sectors listed in DCA’s Market Assessment for Kenya9.
While this led to a large number of overall tweets, very few of them were relevant. We therefore
focused on tweets specifically about finance and loans. For this monitor, we left “financing” in
the taxonomy to capture the chatter about financing in particular sectors, although they might
not be related to individual loans. The rest of the taxonomy was combination of (1) keywords
extracted from the surveys; (2) sectors identified as priority areas of DCA; and (3) sector-specific
words surfaced from the posts themselves. Therefore, this taxonomy only surfaces posts which
are related to a specific type of loan.

9

USAID, DCA Opportunity Overview: Market Assessment and Recommendation for USAID/Kenya, November 9, 2012.

18

(loan OR financing OR mshwari OR mkopo) AND (water OR maji OR energy OR
agriculture OR n'gombe OR business OR biashara OR bizness OR personal OR
emergency OR cow OR dairy OR health OR crops OR resistant OR maize OR
infrastructure OR equity OR home OR "short term" OR afya OR land OR animal OR
animals OR farm OR shamba OR machine OR property OR machine OR auto OR
vehicle OR car OR gari OR bank OR SME OR Enterprise OR stima) AND -"KIKUYU
MIND'" AND -"take that to the bank and get a loan"
Our first attempt at categorization was to parse between and within sectors. For the business
category, we first tried to sort according to the DCA’s Market Assessment for Kenya priority areas:
Agriculture (Dairy, Drought-resistant crops, Horticulture, Maize value chain)
Clean Energy
Health
Water
Infrastructure
However, it was clear that this would not work because of the overlap between personal and
business loan keywords. For example, chatter on Twitter related to energy (or “stima” in
Kiswahili) are both related to the energy sector and personal loans, but none that came up when
training the monitors were related to specific energy enterprises:

#WorldBank backs #Chinese firm’s loan for energy plants in #Kenya
http://t.co/L8MkQd458n
@KenyaPower whats the procedure of acquiring stima loan. What are the requirements?
The second attempt took a broader scope:
Personal
Emergency
Student
Home/ electricity
General
Business
Agriculture
Industry
General
Government
International Loans to government
Local loans to government
Government financing to sectors
However, the low volume of data made it impossible to define the categories with this level of
specificity. The next attempt was to create sub-categories in each broad grouping of “Negative”
and “Positive” and “Neutral.” However, most of the tweets were neither negative nor positive. In
the end, the broader categories were used on their own:
Personal
Business
Government

19

Results
From January 1, 2012 to August 25, 2013, there have been 5,317 relevant posts, representing
an average of 9.2 posts per day. There has been a growth in Twitter similar to the one seen in the
first monitor, with a sustained growth after the launch of M-shwari. This growth is driven by
chatter in business and personal loans, as opposed to government loans. The anomalous events
in this graph are very similar to those described in the general loan category. The chatter,
however, differs.

Figure 10: Volume of Business, Personal and Government related loan data from January 1,
2012 to August 26, 2013.
Categorizing General Chatter
It is possible to see trends between sectors. Figure 11 shows the nature of the conversation
between March 26, 2013 and August 26, 2013.

20

Figure 11: Twitter chatter between March 26, 2013 and August 26, 2013 breaking down loan
by sector. This represents 597 posts from 2,190 relevant posts.
Although the low level of data made it impossible to create specific sub-categories with each
sector, it is possible to look at the actual Twitter data to see how relevant it is to assessing
barriers to accessing loans. Here is a sampling of business related tweets.

Breaking the loan data down this way does allow for a deeper understanding of Twitter data that
might relate to barriers faced by small businesses seeking to access loans, but the low volume of
data prohibits monitoring those trends over time.

Vantage Three: Banks
The final vantage point was to look at how people discuss the banks themselves.
The first step was to look at how much data there is on each bank in Twitter. We first did an
overall search for bank names listed as partner banks in the DCA Market Assessment for Kenya.10
The two with the highest amount of search data were Equity and Kenya Commercial Bank.
For these monitors, we relied heavily on sentiment analysis. Sentiment analysis relies on an
algorithm that ascertains language clues to measure the overall “feeling” of the tweet. In
addition to the predefined algorithm, the algorithm also employs a machine learning technique.
The user sorts a subset of tweets by hand, in the process defined above.
10

We looked for the following banks: Bank of Africa Kenya, Equity Bank, Faulu Kenya Deposit, K-Rep Bank,
Oikocredit; Kenya Commercial Bank, FINA Bank, Acumen Fund, Kenya Women Finance Trust, Micro Africa Limited
Kenya, and SMEP Deposit, taken from USAID, DCA Opportunity Overview: Market Assessment and Recommendation
for USAID/Kenya, November 9, 2012.

21

Equity Bank
Equity Bank consistently has a small amount of Twitter data. The graphic below represents the
volume of Equity Bank tweets from Jan 1, 2012 to August 25, 2013. Over this period, there is a
clear growth in chatter about Equity Bank, both in terms of general daily chatter and in the
frequency of spikes in conversation.

Figure 12: Volume of posts related to Equity Bank from January 1, 2012 to August 25, 2013.
The volume of tweets ranges between zero and a peak of 385.
Sentiment analysis was run on this monitor, but because the level of data for general chatter is
so low—most of the data is from news and events—we were not confident in the results. A
further refinement of this monitor might include decisions regarding exclusion of certain news
items to better focus on peoples’ discussions of the operations of the bank. For the purposes of
this report, all news items were included to provide a broad overview of the type of Equity Bank
related chatter.

There are two drivers of peaks during this period—one is news related to Equity, accounting for
most of the peaks. This news might be related to a new service Equity is launching, or business
news about Equity. The latter includes the largest peak on October 4, 2012, circled in red
above. This is a news peak, Equity appointed Julius Kinpng’etich as its new COO.
Major events can also be seen in this data. The second major peak, which occurs in a cluster
between July 8, 2012 and July 13, 2012 (circled in blue above) is related to a widespread
systems failure in Equity’s ATM services. This event can be seen very clearly in Twitter. Figure
13 illustrates this conversation.

22

Figure 13: Break-down of tweets related to Equity Bank during the Bank’s three-day ATM failure,
date range July 8, 2013 – July 13, 2013.
General chatter
For Equity Bank, the level of general chatter—ie. posts that are not driven by specific events or
the news—is too low to analyze with any confidence in trending behavior.
Kenya Commercial Bank (KCB)
The overall volume and trends in KCB twitter data is represented in Figure 14 below. KCB data
in Twitter somewhat more interesting than Equity data for two primary reasons.
First, KCB has more tweets overall than Equity Bank. Over the same time period, KCB had
13,297 relevant posts, compared to Equity’s 7,944. Second, KCB appears to have a more robust
Twitter strategy than Equity. They post information about bank products, and they also engage
somewhat with customers via Twitter,11 although as can be seen below the instance of customers
reaching out with questions is low. This volume made it possible to do more detailed analysis of
the content of the posts. Figure 14 shows the proportion of posts that are negative, positive,
neutral/ news related, or information related.

11

Equity Bank often advises Twitter users to contact a local bank when they ask questions; KCB does appear to answer
them. Other companies are even more directly engaged with customers via Twitter—while this report does not include
a monitor of Safaricom, they are very engaged with their customer base on Twitter.

23

Figure 14: Volume and sentiment of all tweets related to Kenya Commercial Bank from January
1- August 25, 2013.
One challenge working with the KCB data is that KCB sponsors sports team and sporting events.
This creates a lot of noise in the data, with many KCB references being recaps of games. While
most of these posts can be removed from the data by excluding posts with words like “football,”
this still accounts for the large number of irrelevant posts—less than half of the posts linked to
KCB are relevant. It also accounts for most of the dramatics spikes seen in Figure 14. On game
days, up to 95% of the tweets are irrelevant to function of the bank.
Figure 14 represents a period of time when the majority of posts (those shaded in blue) were
related to the bank. This period shows both a spike representing 215 relevant posts on the peak
day of February 28 (circled in blue). The rest of the period is more representative of “normal”
posting behavior, and appears to be general chatter.

Figure 14: Detail of February 26, 2013 to March 23, 2013.

24

The spike on February 28, 2013, circled in blue in Figure 14 above, is the day KCB released its
investor report. Not only is this event clearly visible in Twitter, it’s also possible to break down
the parts of the report the Twitterverse found the most relevant. This particular event may not be
relevant to the initial research question; however, it is worth exploring what we can see in Twitter
as an example of how one can analyze anomalous events.
Figure 15 groups the primary categories of posts on that day. Here, it is possible to see the
elements of the report that people found interesting, including chatter about the CEO and about
the 14% growth of the bank.

Figure 15: Twitter reacts to KCB press release on February 28, 2013
Categorizing General Chatter
KCB was mentioned much less in Twitter on the dates circled in green in Figure 14—from March
6, 2013 to March 20, 2013. However, because the Twitter trends on those days were more
“normal”, they could potentially have baseline information on more typical trends in Twitter.
These dates include a total of 472 relevant posts, with a peak day of 83 relevant posts on March
9 and a low day of 12 relevant posts on March 10. One cluster—with the word “Masomoloan” is
about education loans.

25

Figure 16: Word cloud showing popular words in loan related tweets from March 6, 2013 to
March 20, 2013.
Another, much smaller, trend that can be seen in this word cloud is “smeloans,” short hand for
small to medium size enterprises. Because small to medium sized enterprises are relevant to the
research question, we investigated that trend further by looking at the actual data. Upon further
investigation, it’s clear that this trend is related to a KCB/Safaricom announcement on March 6
and is largely driven by re-tweets.

26

This time period was chosen to show “general chatter”—however, the data is still driven by news
and events. In essence, it is less a demonstration of general chatter, and more a series of
smaller, news-driven peaks.
Sentiment analysis and the banks: Can we compare?
In looking at these bank-related monitors, a natural question arises of how we can compare the
data between banks. This would require developing the methodology further, including
understanding how the banks’ social media strategies impact related posts.

KEY OPPORTUNITIES
•
•
•
•
•
•
•

•

Twitter use is growing rapidly, and the general monitor shows the emergence of a Kenya
specific Twitter culture.
In particular, Twitter is being used to seek, access and share information about loans,
especially mobile loans.
Twitter is also being used to share and comment on the news related to personal and
business loans. Monitoring the news, as well as how the news is being perceived, might
be of interest.
Much of the data is related to M-shwari and other Safaricom related topics. A future
iteration of the monitors could focus solely on non-traditional banking or exclude Mshwari data for deeper insight on traditional loans.
Monitoring Tweets with a taxonomy of keywords related to sectors does a better job of
discerning trends that might relate specifically to entrepreneurs.
Redeveloping the sector-specific taxonomy such that it excludes personal and
government loans might be worth exploring.
While there is not enough data to train the monitors into further subcategories, the low
number of posts makes it possible to filter the tweets and actually glance through them,
making more qualitative methods viable. This could serve to bring in new ideas, similar
to the role of some focus groups.
For banks that have a robust social media strategy, monitoring the specific Twitter
handle of the bank could provide insight into (1) products and services available at the
bank, and, to a lesser extent, (2) information seeking behavior.

KEY CHALLENGES
•
•
•
•
•

Across monitors, the overall volume of chatter is low, so small changes in Twitter
behavior, for example due to a popular retweet or the behavior of one Twitter user, can
create spikes in overall number of conversations.
There is very little general chatter about loans (i.e., chatter is largely driven by news/
events or information seeking).
There is little demographic data available, so even where there may be tweets that are
directly related to barriers to accessing loans, it is difficult to know whether they
represent underserved entrepreneurs.
As seen by the launch of M-shwari, Twitter is a dynamic space and the analysis that is
possible using Twitter will change as the way that people use it changes. It is important
to keep an eye on these changes.
There is a lot of noise in the data. For KCB, this noise includes sports chatter. For both
banks, this includes news item related to the overall business of the bank, not
necessarily directly related to bank services.

27

GOOGLE TRENDS
Information seeking behavior can shed light on people’s activities in a profound way. In the
surveys we did, clients cited a wide variety of information needs, from new farming inputs to
different loan products to definitions of keywords like “guarantee.” Google represents an
enormous repository of data about information seeking behavior. While Google does not provide
direct public access to the search data, they do provide some free online tools to analyze the
data. This data is interesting in part because, although we have no individual demographic
information about the user, it represents data that is not shared publically, ie. it might illustrate
personal needs and interests that people do not wish to share publically.
In Kenya, the only available tool is Google Trends “Hot Searches.”12 With Google Trends, one can
track search terms over time. As an analytical tool, Hot Searches is somewhat limited13. First,
Google Trends only shows relative volumes of searches, it does not show real numbers. Second,
while it is possible to create a graph that aggregates synonyms in one trend line, it is not possible
to create subcategories within one search term—this is explored in greater detail in the examples
below.
Another limitation is that it is difficult to use Google Trends to “learn”—typically in big data
analytics taxonomies will be developed both by working with domain experts (in this case DCA
clients) and by identifying keywords in the data itself. Scope for this type of iteration is limited in
Google Trends—only one word that commonly occurs with the initial keyword is given. This both
limits discovery and the ability to contextualize the trends.
This section explores what is possible to look at using Google Trends, starting with the keywords
generated from the survey. It then moves on to look at loans generally, banks, and sectors.
In all of the experiments below, the time period is August 21, 2012 to August 21, 2013.

Taxonomy from surveys
In the surveys, people noted that they looked up definitions of banking terms. However, trends in
searching for this data is very erratic, suggesting that the overall volume is low, allowing small
changes in volume to dramatic spikes.
One of the survey respondents reported searching for the word “guarantee.” Because there are
several variations of that word that could be used, the red trend line is an aggregation of
“guarantee” and “guarantor.” Even with both terms, the search volume is too low to even register
on the graph.

12

Hot Searches can be found at the website: http://www.google.com/trends/explore#cmpt=q
Google Correlate, while not currently available in Kenya, is somewhat more useful than Hot Searches. Google
Correlate uses time series data sets—for example the unemployment rate—to detect search volumes that follow the
same trends.
13

28

Figure 17: Google searches for finance- terms in Kenya, August 21, 2012 to August 21, 2013.
In addition, when compared with searches for loans, it appears as if the volume of searching for
loans is much higher than for definitions.

Figure 18: Google search trends, comparing loan searches (excluding HELB) and “interest rates”
in Kenya, August 21, 2012 to August 21, 2013.
Loans
There is no straightforward way to create sub-categories within overall loan searches. In Google
Trends there are two ways to approximate this. First is to specify a full search phrase in quotes,
for example “business loan” or “personal loan.” When we did this, no data was returned. Second
is to exclude words from the search, for example “Loan -student.”
Google Trends shows the top co-occurring word with “loan,” which is HELB, the student loan
authority. The following graph shows loans with and without the words “HELB” and “student.”
This is essentially the extent to which it is possible to specify what types of loans people search
for with the information available. It is important also to underline a subtle distinction— the red
trend line below merely excludes those searches that include the words “HELB” and “student”—
it is quite possible that people searching for simply the word “loan” are looking for student loans.

29

Figure 19: Google search trends, comparing “loan” and “loan,” excluding “HELB” and “student”
in Kenya, August 21, 2012 to August 21, 2013.
From the above graph, it’s clear that general loan searches are fairly stable, with one spike in
early August, 2013. Given that this spike occurs roughly one month before the start of school, it
is possible that this is indeed related to student loans.
Banks
I also compared searches related to DCA partner banks. Here, despite the limited amount of
information we have about this data, there are potential insights we can extract. One is
anomalous behavior. In both of the graphs below, there is not much to suggest a spectacular
event, with the exception of the drop on March 3 across all issue areas. This was Kenya’s
presidential election.
The other is overall interest between search terms. It’s clear, for example, that more people
search for Equity Bank and KCB than Faulu and Bank of Africa, perhaps related to the size of the
banks.

Figure 20: Google search trends, comparing four banks in Kenya, August 21, 2012 to August
21, 2013.

30

Sectors
Finally, we compared the sectors listed in DCA’s Market Assessment for Kenya. As mentioned
above, it is not possible to look at how searches for loans intersect with sectors other than
student loans. The drop across sectors is again on March 4, the day of Kenya’s national election.
It’s clear that between sectors, people search for health more than any other sector. However,
this could be because there is more information about health online that is directly relevant to
people’s lives—compared, for example, with “infrastructure” which might be considered
important but less relevant in terms of immediate information needs. In this case, defining
infrastructure more precisely—for example “roads” or “municipal water” could potentially
change this picture.

Figure 21: Google search trends, comparing sectors in Kenya, including “Health + Afya,”
“agriculture + drought resistant + crop + horticulture,” “water,” “energy + stima + electricity,”
and “infrastructure” in Kenya, August 21, 2012 to August 21, 2013.

!
KEY OPPORTUNITIES
•
•

The initial tests with Google can provide baseline information to monitor for anomalies in
search trends in the future.
Events that are broadly shared—ie. the beginning of a school semester or the Kenyan
elections—are clearly visible in the data. This could mean that other systemic events will be
visible as well.

KEY CHALLENGES
•
•
•

The analytical tools made available by Google for Kenya data are limited.
Currently it is only possible to know the relative volume of searches, it is not possible to
analyze real numbers.
Student loans seem to be driving the loan searches, even when words like “HELB” and
“student” are excluded.

31

FINDING THE DIGITAL FOOTPRINT

Sections II and III starts from the available data to look for information that might be useful.
Section IV takes the opposite approach: looking at where business owners (1) discuss and (2)
seek business and loan related information. The section begins by discussing the results of the
survey, which asked people where they get information about loans and their business. It then
provides two case studies on where other digital data might lie.

Survey results
In the survey of 10 DCA clients, we asked where people got information about their business,
existing loans, and new loans. Respondents all reported getting information from multiple
information streams, listed in Table 2 below. People reported that, while interacting in public
spaces is a primary way to access information about business, they are much less likely to
discuss existing or new loans in public spaces. Across the board, those who mentioned that they
use the Internet to access information mentioned that their children as helped them. While
mobile phones were not cited as primary source of information about business or about new
loans, most respondents used mobile phones to manage their existing loans. This includes
paying, receiving alerts and other communications from loan officers via SMS, receiving ministatements and deposit information.

In
person

Media

Digital

Business Information

Loan information

Colleagues in markets,
buyers who come to the
farm, friends and family,
trade shows, Agrovets/
stores, Cooperative society

Family, visit banks

TV programs ("Shamba
Shake-ups"), Radio,
Newspapers, Magazines

Newspapers, pamphlets,
books (especially the
dictionary to look up
technical terms), Notice
boards

E-banking, Wikipedia,
Mobile phones (calling
colleagues,
communication with staff)

Bank website, Google,
M-shwari, SMS
(ministatements), Call
banks

Table 2: Summary of survey results regarding where people self-reported getting information
about their business, existing loans, and new loans

32

Case studies in digital data

Case Study 1: Agriculture
Kenya’s dynamic tech sector is developing new services across a number of sectors. These
services are creating a great deal of “data exhaust,” but the data is often either proprietary or it
is not collected in a way that makes it directly useable. This section looks at programs that may
create data exhaust, using the case study of agriculture. It’s likely that not all of these programs
will succeed; however, for those that grow, the data will likely be a valuable asset.
The key will be to have access to that data in order to conduct any analysis. Depending on the
nature of the service, accessing the data will require different efforts. One approach might be to
establish partnership agreements to access to the data. In some cases, the service itself is an
aggregation of data, and a simple subscription will allow access to the data. The final suggested
way forward is a big data methodology called “scraping.” For data that is made available on-line,
a tool can be built which will automatically “scrape” the data from a website. Scrapers are
typically designed such that the data is pulled down from the web and is made available in a predesignated format. Data from different websites can thus be integrated in such a way that it can
be analyzed together. The table below lists programs related to agriculture that generate data
exhaust, as well as what would be required to access that data.

33

Program

Data Description

What's required

Agrimanagr

Real time data all along the value chain, from
crop health to distribution to staff
management. Designed for medium to large
agribusiness.
Real time data on dairy credit within
cooperative credit structures.

Partnership

General agricultural market data on producers,
banks, agro-dealers and buyers. This
integrates other data, like weather data and
subscribers can access the aggregated forms
of data.
Sector information on cows, around things like
cow health, calf management, disease
management, milking methods and feeding.

Partnership or subscription

Dairy Sacco App
DrumNet

iCow

Partnership

For all data: partnership with
Safaricom; alternatively, the
basic messaging could be
aggregated by signing up for
the service.
Partnership

Kilimo Salama

Weather data, basic micro-insurance data.

Mfarm

Market information on agriculture, price
information and supply and demand.

Partnership or scraper

Mkilimo

Data on what information farmers are seeking
(Farmer help line)

Partnership

National Farmers
Information Service

Aggregates information streams from
government sources, markets, and weather
data, including national livestock marketing
information system.
Data on livestock producers and traders.

Scraper, partnership

Partnership

Regional Agri Trade
Network

Social data-- Next2 connects people and
organizations with similar concerns or
interests over mobile phone, a text-based
social network.
Agriculture market information for East Africa,
includes trade data.

SokoniSMS

Price information

Subscription

Sokopepe

Market information on agriculture, facilitates
mobile transactions

Partnership

Soko Shambani

Potato supply and demand data, using SMS
short term, integrated with Twitter.

Partnership

Ukulima.net

Aggregation site to integrate data from the
web, mobile and specialized apps. (In
development)

Partnership

National Livestock
Marketing Information
System
Next2

Partnership

Scraper, partnership

34

Table 3: List of digital services related to agriculture in Kenya.
Case Study 2: Banks
Of course, the treasure trove of relevant big data would be to access Safaricom, M-PESA and Mshwari data. However, aside from the mobile banking data, banks put a great deal of information
online which could be useful. Depending on what data could be relevant, and how often the data
is changed, building scrapers to collect this data from websites could be another approach.
Some banks, for example Barclays Bank below, post interest rates, but this is not consistent
across banks. However, most banks provide information about the types of loan products they
provide. A scraper could help track changes in loan products and other bank information across
banks.
Two examples of online data that could be scraped are below, from the websites of Barclays
Bank, Kenya and Equity Bank.

Figure 23: Barclays Bank interest rates from the Barclays Kenya website
(http://www.barclays.com/africa/kenya/commercial/assetfinance.php), accessed on August 20,
2013.

Figure 24: Screen shot of Equity Bank Agricultural Loans from the Equity Bank Kenya website
(http://www.equitybank.co.ke/index.php/loans/agriculture-loans), accessed on August 20, 2013.

35

RECOMMENDATIONS AND NEXT
STEPS
How do big data projects typically work?
Big Data projects work best when they involve several iterations and collaboration between
topical domain experts (who understand both the context, and the programmatic information
gaps or needs), and data scientists/analysts. This landscaping study was intended to lay out the
landscape of finance-related big data in Kenya. The next step is for DCA to map out their own
data needs in relation to these possibilities. At that stage, creative brainstorming with the Global
Pulse team might be productive in connecting the opportunities with the needs.
To assist with discussions, some potential next steps are below. These are intended to inform a
larger conversation on whether further research could support the work of USAID.
Laying the Foundation: Better Baseline data
The survey about DCA clients’ use of digital services, limited though it was, yielded insights that
proved valuable to this report. To better understand how to work with new sources of digital data
in Kenya, expanding the survey to more DCA clients or other underserved entrepreneurs could (1)
provide more contextual information to inform the approach, including how underserved
entrepreneurs use digital services; and (2) provide further ethnographic data on how and where
people discuss finance or related issues. It might be the case that DCA would choose to continue
working with the banks to target the surveys. It might also be valuable to explore other targeting
options, for example by working with trade associations or cooperatives. Looking at where clients
reported accessing information about their businesses (see Table 3 above) might be a good place
to start thinking through targeting mechanisms. Another option would be to experiment with new
ways of targeting, for example using Twitter itself, or perhaps Facebook ads or sending out opt-in
mobile phone surveys.
Looking for low-hanging fruit in social data
While it’s clear that social media is on the rise in Kenya, it is also clear that for the purposes of
informing research on financial inclusion, it is probably not saturated enough yet. If DCA is
interested in using Twitter data to inform its work now, it might be a good idea to use this
methodology in a country that DCA works, which has a stronger social media culture. In some
emerging markets, Twitter has become fairly mainstream. For example, it has been widely
reported Jakarta is the top-tweeting city in the world—and the 6th most prolific city is another
Indonesian city, Bandung. Indonesia, India and Mexico are the 5th, 6th and 7th most prolific
countries, respectively14.
Looking ahead: staying in front of digital data in Kenya
That said, there are areas that might be worth exploring more to anticipate the rising relevance of
this data in Kenya. This is the case both for social media and other digital data.

14

Lipman, Victor, “The World’s Most Active Twitter City? You won’t guess it,” Forbes, December 30, 2012. Accessed
at http://www.forbes.com/sites/victorlipman/2012/12/30/the-worlds-most-active-twitter-city-you-wont-guess-it/

36

Social Media
For social media data, looking ahead could involve keeping an eye on the evolution of Twitter
culture to understand how it could be informative for future research. By looking at the
“general loan” monitor, it’s possible to see the emergence of a Kenya-specific Twitter
culture. At this stage, the Twitter culture includes:
Information seeking behavior: Twitter is being used for customer care. For banks with
robust social media strategies, in particular Safaricom’s mobile banking, customers are
asking questions and receiving answers. This might be a good opportunity for further
investigation to understanding people’s concerns.
M-shwari and non-traditional banking: Close to half of the chatter related to loans is
related to M-shwari. In addition, 8 out of the 10 peri-urban or rural farmers that we
surveyed mentioned that they use mobile banking. Monitoring trends in Twitter might be
a good way to understand how this type of lending intersects with business finance.
News, events and institutions: Across all categories, news and events drive the
conversation. Keeping an eye on these trends could provide information about new
financial products. Sentiment analysis could be particularly useful to show how Kenyans
perceive news, events, and institutions like banks or even USAID.

Institutional Innovation
Laying the groundwork for future work might also include thinking about DCA or USAID’s
current capacities and practices, including:
Partnerships As shown in the agriculture case study in Section IV, new mobile services
are being developed in Kenya all the time. All of these services create data—which is not
publically available. However, USAID funds some of these initiatives and should be
thinking about data from the beginning of project development. In addition, it might be
possible to access such data through partnership agreements, although the data would
have to be properly anonymized.
New Capacities A great deal of information is readily available on-line, but collecting it by
hand doesn’t make sense. Building the capacity to automate collecting that data through
scrapers might be able to contribute to market analysis, in particular around things like
prices, products, and/or interest rates. Deciding whether to go in this direction will
require identifying websites that hold information of interest of DCA and assessing how
often the information is updated. These scrapers need some supervision—for example the
scraper would need to be updated if a website is redesigned—so building a scraper is not
a one-time expense. Therefore the next steps here are to (1) assess the information online that could be of use to DCA; (2) assess the requirements of the scraper to automate
the data collection; and (3) assess the maintenance requirements of the scraper.
Combining new and old: “ground truth” data
New, digital data sources need not be considered on their own. There are several important
reasons to bring together digital data with more traditional data sources.
Establishing relationships between on-line and off-line behavior

37

Evaluating digital trends in relation to off-line trends could provide more insight into the
meaning of observed digital behavior. As a hypothetical example, it might be possible to find
a correlation between loan application rates and Twitter chatter around loans. Once a pattern
is established, detecting similar behavior in Twitter going forward could provide clues about
loan seeking behavior. (However, because the level of general chatter about loans is so low
in Kenya, for now this approach might not be very useful in that particular country.)
The same could be true of trends in mobile services, if such data can be accessed.
Correlating anonymized mobile data with statistical data-sets or dedicated surveys could lay
the groundwork for understanding, for example, what changes in M-shwari loan patterns
mean in term of people’s lived experiences.
Integrating digital data with existing data sources
Just like any data source or information stream, digital data only paints part of the picture.
When thinking through a big data research agenda, it might be worth considering how
existing traditional data can support the work. Conversely, when thinking through a more
traditional research agenda, it’s worth considering how big data might provide additional
insights.
Conclusion
From this landscaping study, it is clear that certain big data methodologies could be leveraged to
enhance USAID’s programmatic capacity to get early insight or warning of emerging trends, keep
an eye on the “pulse” of a given issue area, and get more real-time feedback on programs as they
are being rolled out. However, it is equally clear that further work is required to turn the
possibilities for big data analysis into meaningful means of operation in practice.
To do so, creativity and imagination is key—and it is worth thinking both “big” and “small.” At
some stage, big data could contribute to projects that we previously have struggled to undertake,
for example understanding complexity in markets; or gaining new insights into social networks
and financing. Big data might also be able to support relatively simple information needs, for
example capturing the pulse on sentiment around news and events.
This report seeks to provide the information required to understand the opportunities for big data
analytics around financial inclusion in Kenya. Determining which line of inquiry makes the most
sense moving forward is a matter of matching the needs and opportunities, as well as exploring
the internal capacity required to employ new methodologies.

38

ANNEX 1:
ID Number ________
Bank Questionnaire
Introduction
Hello, my name is Esther Ajambo. I’m doing a survey in conjunction with the United Nations
Global Pulse initiative and USAID. We are trying to understand how people get information
about loans and we have some questions for you. As someone who receives a loan, you were
recommended by KCB bank (?). The survey will take between 30-45 minutes. Are you available
to do the survey now?
If yes, then begin survey.
If no, then say
Is there another time that you are available to do the survey? I will call you at a later time.
I’m going to start by asking a few background questions.
1.

How old are you? _____________

2.

Esther please fill out without asking. Are you male/ female? M/ F

3.

What town and county do you live in? ________________

4.

What is your highest level of formal education? (circle one)
(1) no formal education
(2) primary school
(3) secondary school
(4) university
(5) graduate school

5.

What type of business do you do? ____________________________

6.

Do you employ anyone?
a. If yes, how many people?

Y/N
__________

Loan Profile
Now I’m going to ask you some questions about loans that you have applied for or received in
the past two years.
7.

In the past 2 years, how many loans for your business have you applied for? _______

39

8.

In the past 2 years, how many loans for your business have you received?
__________

9.

In the past 2 years, which banks did you apply for loans from? (list)

10.

In the past 2 years, which banks did you receive loans from? (list)

11.

Please think about all the loans you applied for in the last two years. How did you
find out about the loan?
a. (List, make sure their number of answers matches the number they said they
applied for)

We are looking for words that people use to talk about loans—we’re interested in the actual
words, the languages that you use, and the conversations that you have. Please try and
think of the key words that you use when you talk to people about loans. For example, you
might say to a loan officer “I have a question about whether I can top-up my loan?” And
the loan officer might answer giving you a new interest rate, a payment schedule, or other
things. We would be interested in the keywords in this discussion, like “top-up,” “payment
schedule” etc.
12.

Can you remember the last time you spoke to a bank official or a loan officer about a
loan—like a friend, or a family member, or a colleague? Y/N

13.

What is that person’s job title/ relationship to you? __________________________

14.

What type of loan did you talk about? _______________________________

15.
1
2
3
16.

What language or languages did you use to talk about the loan to this person?
English
Kiswahili
Local language, please list _______________________
What words did you use to talk about loans to this person? Please list all the words
they use. You might have to explain this, asking them a little bit about the
conversation that you had with them.
_____________________________________________________________________
_____________________________________________________________________
_____________________________________________________________________

40

Esther, let them answer, and then prompt them.
a. Did you use any technical words, like the type of loan?
List technical words:

b. Did you use any slang words—for example in a dialect?
List slang:

17.

Can you remember the last person who was not a bank official or loan officer that
you spoke about loans with? Y/N

18.

What is that person’s relationship to you? ________________________________

19.

What type of loan did you talk about? ___________________________________

20.

What language or languages did you use to talk about the loan to this person?
1 English
2 Kiswahili
3 Local language, please list _______________________

21.

Including all languages that you spoke to this person in, what words did you use to
talk about loans to this person?

Esther, let them answer, and then prompt them.
a. Did you use any technical words? For example, type of loan?
List technical words:

b. Are there any slang words that you use to talk about loans—for example in sheng
or another dialect?
List slang:

41

22.

Have you ever discussed loans on (1) Facebook?
(2) Twitter?

Y/N
Y/N

23.

Have you every searched for loan information on Google?
i. What search terms did you use?

Y/N

Digital footprint
I’m going to ask you some questions about your access to technology.
24.
25.
26.

Do you have a mobile phone?
Y/N
Do you have a computer?
Y/N
Have you ever accessed internet? Y/N
If yes, then ask. . .
- Mobile phone
Y/N
- Cyber Café
Y/N
- At home on a computer
Y/N
- Friend’s computer
Y/N
- School computer
Y/N
- Work computer
Y/N
- I don’t use the internet
Y/N
- Other, please describe _________________________________________

27.

How often do you use facebook?
(1) More than once per day
(2) 2-3 times per week
(3) Once per week
(4) 2-3 times per month
(5) Once or less per month
(6) I don’t use facebook

28.

Do you ever talk about your business on facebook?

29.

Do you ever talk about loans on facebook?

30.

Do you know what Twitter is? Y/N

31.

Do you have a Twitter account? Y/N

32.

Do you use banking services on your mobile phone? Y/N
a. If yes, please list all banking services you use on your mobile phone.

Access to information

42

33.

How do you access information about your business? This includes price
information, market opportunities, and general queries. Please list all the places you
get information for your business.

Esther, use these as prompts—
i. Do you talk to friends or family to get information that you use in your
business?
Y/N
What type of information?
____________________________________________________________
____________________________________________________________
_______________________________________________________
ii. Do you access information on about your business on your mobile phone?
____________________________________________________________
____________________________________________________________
____________________________________________________________
____________________________________________________________
Please list all programs that you use on a mobile phone for business
related information
____________________________________________________________
____________________________________________________________
_______________________________________________________
iii. Do you access information about your business in the newspapers,
pamphlets, or books? Please describe the type of information and the
media.
____________________________________________________________
____________________________________________________________
iv. Do you ever seek business related information on the internet? Y/N
Please list the sites you would use on the internet.
_____________________________________________________________
_____________________________________________________________
_____________________________________________________________
v. Other, please describe.

34.

How do you access information about your existing loans?

43

i. Do you discuss your loans with friends or family? Y/N
____________________________________________________________
____________________________________________________________
_______________________________________________________
ii. Do you access information about your loans on a mobile phone?
Please describe what information about your loans you get on a mobile
phone.
____________________________________________________________
____________________________________________________________
____________________________________________________________
____________________________________________________________
Please list all programs that you use on a mobile phone to access
information about your loans
____________________________________________________________
____________________________________________________________
_______________________________________________________
iii. Do you physically go to the bank?
Y/N
Please describe what information you get by physically going to the bank.
____________________________________________________________
____________________________________________________________
____________________________________________________________
iv. Do you use the internet? Please describe what information you get via the
internet.
____________________________________________________________
____________________________________________________________
___________________________________________________________
v. Other, please describe
35.

If you were interested in getting a new loan, and you wanted information on what
types of loans were available where would you look for that information?

i. Would you talk to friends or family to get information about new loans?
Y/N
1. What type of information?
____________________________________________________________
____________________________________________________________
_______________________________________________________
ii. Would you access information on about new loans on your mobile phone?
Y/N

44

____________________________________________________________
____________________________________________________________
____________________________________________________________
____________________________________________________________
Please list all programs that you would use on a mobile phone to get
information about new loans
____________________________________________________________
____________________________________________________________
_______________________________________________________
iii. Would you access information about new loans in the newspaper,
pamphlets, or books? Please describe what information and the type of
media.
____________________________________________________________
____________________________________________________________
iv. Would you ever seek information about new loans on the internet? Y/N
Please list the sites you would use on the internet.
____________________________________________________________
____________________________________________________________
____________________________________________________________
v. Would you physically go to the bank? Y/N
What type of information would you seek at the bank?
____________________________________________________________
____________________________________________________________
____________________________________________________________
vi. Other, please describe

45

Digital Signals & Access to Finance in Kenya

More Related Content

What's hot (20)

Viewers also liked (13)

Similar to Digital Signals & Access to Finance in Kenya (20)

More from UN Global Pulse (20)

Recently uploaded (20)

Digital Signals & Access to Finance in Kenya