Why Star Ratings Matter for Financial Inclusion

February 2016

Main findings

2

• E-commerce sales data, not typically available for credit scoring, can enrich existing scoring models and improve their predictive power.

• Also, e-commerce data far removed from strictly financial characteristics, such as star ratings, can also be used as a predictor in credit scoring models.

• This is good news for the financially excluded in two ways, with further pilot testing needed:

1. Visibility: First time credit assessment for customers with no formal financial history can be based on this alternative data.

2. Enrichment: Thin financial data can be enriched for unbanked segments to better assess their risk.

• Almost 40% of the active sellers under analysis do not have an open credit in the market, hinting at the huge opportunities of unbanked markets.

MercadoLibre in numbers

3

• MercadoLibre is the largest e-commerce ecosystem in Latin America and offers a wide range of services to sellers and buyers including marketplace, payments, advertising and e-building solutions.

• MercadoLibre operates in 13 countries including Argentina, Brazil, Chile, Colombia, Mexico, Peru and Venezuela. Based on unique visitors and page views, MercadoLibre is the market leader in the countries where it has a major presence.

• 2015 full year numbers:

• 144.6 million registered users by the end of the year.

• 128.4 million items sold through MercadoLibre, 26.8% growth over previous year, resulting in a gross merchandise volume of $7.2 billion.

• 80.4 million total payment transactions, 73.7% increase over previous year, for total payment volume of $5.2 billion.

• Consolidated net revenues of $651.8 million,17.1% increase over previous year.

Methodology

4

• Bi-variate analysis to test whether selected e-commerce variables rank sellers by credit score.

• E-commerce variables were chosen from a pool of around 1,500 internal variables. Credit bureau data were retrieved for a segment of 12,794 active sellers in Argentina.

• Argentine credit bureaus have almost universal reach given the high penetration of unique IDs, though roughly 50% have very “thin” information.

• The Equifax/Veraz Argentina score used a range from 1 (most risky) to 999 (least risky). A derived ordinal variable was created with five categories, in which each category represents a range of 200.

• Only showing statistically significant results 0.01 ANOVA (Brown-Forsythe and Welch statistics).

• Definition of “unbanked”: Sellers without an active credit line in the market for the last 5 years, taken from Equifax/BCRA.

1: Alternative data with good predictive power can be used to assess the creditworthiness of customers with no financial information, turning them from invisible to visible.

2: Imperfect correlation when considering more granular segments presents an opportunity to enrich existing models with these newly acquired data points. This is particularly true for credit scorecards built using limited information and strictly financial data. Perfect correlation would not add substantive information to existing models.

Practical implications: Correlation and granularity

5

The case for pilot testing is supported by the following:

Tim

e si

nce

regi

stra

tion

(mon

ths)

Tim

e si

nce

regi

stra

tion

(mon

ths)

All sellers

Risk score category

Unbanked

Risk score category

Variables 1: Seller maturity, months operating with MercadoLibre

• Longer duration of engagement yields a higher (less risky) credit score, true also for unbanked sellers.

6

Tim

e si

nce

regi

stra

tion

(mon

ths)

Risk score category Risk score category Risk score category

Tim

e si

nce

first

pub

licat

ion

(mon

ths)

Unbanked

Tim

e si

nce

regi

stra

tion

(mon

ths)

Risk score category Risk score category Risk score categoryTim

e si

nce

first

pub

licat

ion

(mon

ths)

All sellers

Tim

e si

nce

first

sal

e (m

onth

s)Ti

me

sinc

e fir

st s

ale

(mon

ths)

Variables 2: Star ratings or feedback from buyers

7

• Sellers with a higher percentage of positive feedback tend to have higher credit scores. The opposite happens when feedback is negative.

Risk score category Risk score category Risk score category

Risk score category Risk score category Risk score category

Unbanked

All sellers

% p

ositi

ve fe

edba

ck (l

ast 1

2 m

o)%

pos

itive

feed

back

(las

t 12

mo)

% n

egat

ive

feed

back

(las

t 12

mo)

% n

egat

ive

feed

back

(las

t 12

mo)

% n

egat

ive

feed

back

(per

mo)

% n

egat

ive

feed

back

(per

mo)

Recency since last publication (in months)

Activity (for sellers with at least one sale in the past 12 months)

Variables 3: Recency and activity

8

• With recency of publication, longer time since last publication corresponds to a lower credit score (more risk).

• The greater the activity in the past 12 months the higher (less risky) the credit score.

All sellers

All sellers Unbanked

Unbanked

Risk score category Risk score category

Risk score category Risk score category

Rec

ency

sinc

e la

st p

ublic

atio

n (m

o)

Rec

ency

sinc

e la

st p

ublic

atio

n (m

o)

Num

ber o

f mon

ths

of a

ctiv

e se

lling

In p

ast y

ear

Num

ber o

f mon

ths

of a

ctiv

e se

lling

In p

ast y

ear

Variables 4: Sales growth, last 3 months vs. last 12 months

• Sellers who tend to increase their sales levels showed higher (less risky) credit scores.

9

Last 3 mo. against last 12 mo. of sales

Risk score category Risk score category Risk score category

Risk score category Risk score category Risk score category

All sellers

% s

elle

rs w

ith in

crea

sed

sale

s

% s

elle

rs w

ith in

crea

sed

sale

s

% s

elle

rs w

ith in

crea

sed

sale

s

% s

elle

rs w

ith in

crea

sed

sale

s

% s

elle

rs w

ith in

crea

sed

sale

s

% s

elle

rs w

ith in

crea

sed

sale

sUnbanked

Last 6 mo. against last 12 mo. of sales

Last 3 mo. against last 6 mo. of sales

Last 3 mo. against last 12 mo. of sales

Last 6 mo. against last 12 mo. of sales

Last 3 mo. against last 6 mo. of sales

Assessing the potential unbanked market

10

The average credit score is 715 for the banked population and 608 for the unbanked.

• Almost 40% of active sellers are unbanked.

• The unbanked are probably underrepresented here compared to other segments

Two opportunities:1. Visibility of no-hits through

alternative credit scoring (in market other than Argentina, where coverage is quasi/universal).

2. Enrichment of existing models to find new creditworthy customers.

Risk score scale Low RiskHigh Risk

Advancing financial inclusion to improve the lives of the poor

www.cgap.org