A Market Analytics Perspective of an Online/App Based
Recharge Company
Agenda• About the Firm - Model
• Objectives
• Data Analysis
• Inferences
• Recommendations
XYZ Ltd.
• A recharge portal
• Recharges largely for mobiles; data cards and DTH follow
• Offers discount coupons of popular food joints and
retailers for a small fee
Business Model
Commissions from telcos/DTH
providers for each recharge
done on the site
Small fee from customers when
they opt for coupons
Fees from coupon
merchants
Revenue Generation
Merchants account for the largest share of the company’s revenues (50%),
followed by customer fee for coupons (30%) while recharge commissions come last
(20%). The margins in the recharge business are low, lying in the range of 0.5% -
1%.
Recharge Process
Objectives• Identify points to increase revenue generation
• Evaluate performance of campaigns
The Data• Absence of customer – wise data (no transaction IDs); missing data
• Data cleaning – missing values filled in as 0s in some cases, and
‘Unknown’ in others (device type)
• Categorical variables (Source, device type) transformed into
continuous predictors
• Part of data converted into categorical variables – high rate of
‘Payment Success’ as a criteria
Techniques Used
Tool Predictor Variables Response Variable
Excel Pivots and Graphs N/A N/A
Naïve Bayes
Source campaign, Total Registrations, Coupon page visits,Coupon option, Payment attempt - Credit Card, Debit Card, Net Banking, HDFC ATM, Promo
Coupons Payment Success
Multiple Linear RegressionSource campaign, Installation, Total Registrations Coupon page visits, coupon
option, Payment attempt Payment Success
Data Analysis
Affiliate – 1 fares well on all elements of AIDA
Campaig
n - 1
Campaig
n - 2
Campaig
n - 3
Campaig
n - 4
Campaig
n - 5
Campaig
n - 6
Campaig
n - 7
Campaig
n - 8
Campaig
n - 9
Campaig
n - 10
Campaig
n - 11
Campaig
n - 12
Campaig
n - 13
Campaig
n - 14
Campaig
n - 15
Affiliate -
1
Partners
-10.00%
5.00%
10.00%
15.00%
20.00%
25.00%
7%6%
5%
10%
15%
3%
11%
14%
12%
21%
13%14%
12%
4% 5%
9%
16%
Campaign - wise Customer Revenue (Pay Success)
Analysis - Graphical
foodsto
re
lifesty
lecar
e
clothing
electr
onics gifts
travel
books kids
miscella
neous
enter
tainmen
t0
20000
40000
60000
80000
100000
120000
140000
160000 151023
34403
0 1818 2084 1 0 1790 0 0 1865368
Coupons Opted for by Customers
Analysis - Graphical
10.75%
57.20%
24.91%
0.02% 7.12%
Credit Card
Debit Card
Net Banking
HDFC Credit Card - ATM
Promo Coupons
Modes of Recharge Payment
56.37%32.35%
3.23%
1.27%
0.43%
0.84%0.06%
5.46%
Android PhoneAndroid TabiOS PhoneiOS TabDesktopWindows PhoneOthersUnknown
Devices Used
Analysis - Graphical
Naïve Bayes
• Post data cleaning, campaigns with pay success of 14% and
above were selected
• Excel random number generator - numbers between 1 and 10
were assigned to each row
• Rows with #5 were selected and used for categorical data
simulation
Analysis - Classification
• Since the data size was small, no pivots were used
• Data partitioning - Training data – 60%; Validation data – 40%
• Cut-off probability for success – 0.5
• 9 predictor variables
Analysis - Classification
Naive BayesSimulated data and Naïve Bayes Model
Classification Confusion Matrix Predicted Class
Actual Class 0 1
0 193 12
1 5 26
Analysis - ClassificationClassification Confusion Matrix
Predicted Class
Actual Class 0 1
0 125 12
1 2 18
Training Data Scoring Validation Data Scoring
Error rate: (5+12)/236 = 7.2%
Accuracy rate: (1 - .072) = 92.8%
Sensitivity: 26/(26+5) = 83.9%
Specificity: 193/(193+12) = 94.1%
False positive rate: 12/(12+26) = 31.6%
False negative rate: 5/(5+193) = 2.5%
Error rate: (2+12)/157 = 8.9%
Accuracy rate: (1 - .089) = 91.1%
Sensitivity: 18/(18+2) = 90%
Specificity: 193/(193+12) = 91.2%
False positive rate: 12/(12+18) = 40%
False negative rate: 2/(2+125) = 1.6%
Analysis - Predictive
Multiple Linear Regression
• Post data cleaning, ‘Transform Categorical Data’ was
applied on predictor variables Source and Device Type
• Data was partitioned into 60% training and 40% validation
• 5 predictor variables usedMLR
Multiple Linear Regression Model
Analysis - PredictiveInput
Variables Coefficient Std. Error t-Statistic P-Value
pay_attempt 0.9037 0.0032 278.2815 0
Source_red 0.1428 0.0467 3.0565 0.0023
Total_Registrations 0.0003 0.0006 0.4961 0.6199
Installation -0.0002 0.0006 -0.4294 0.6677
Coupon_page -0.1656 0.0017 -95.9321 0
Intercept -1.2571 0.5486 -2.2917 0.022
Residual DF 3062R² 0.9959Adjusted R² 0.9959Std. Error Estimate 14.3005RSS 626191
Due to their low p-values, we consider pay_attempt, source and
coupon_page as important predictor variables.
Pay attempt and Pay success are positively and highly correlated.
Inferences• Affiliate 1 seems to be a fairly good campaign in terms of
creating awareness, converting awareness to registrations
and subsequent steps, leading to successful payment
• The main pain point seems to occur at after the total visits
to coupon page. Not many customers end up attempting to
pay
• Food coupons account for approximately 77% of total
coupons selected by customers
• Debit card payments account for more than half the recharge
payments made
• The android OS based devices are more popularly associated
with the usage of the recharge app
Inferences
• The Naïve Bayes model can be used to classify users into
those who’s payments would successful and those who’s
wouldn’t, based on the payment modes
• The MLR can be used to best predict payment success.
Payment attempt is the best predictor of payment success
Inferences
Recommendations• Campaigns modelled on Affiliate – 1 could be conducted
• Vendors for food campaigns could be charged a higher
commission as the demand for these coupons is high
• Explore reasons as to why conversion rate from Coupon Page
visits to Payment Attempt is low – coupons of choice not
available, preferred payment mode not available, transaction
did not go through, etc.
• Make the app more user friendly for iOS and Windows phone
users to leverage these segments
• Using the Naïve Bayes model, specific campaigns can be
designed and emailed to the target users to boost customer
recharge revenue
Recommendations
• The MLR model can be used to predict the effectiveness of
future campaigns based on their similarities to the campaigns
presented in the data
Recommendations
Thank You