You are on page 1of 34

Analytics

Application:
Sales and
Marketing
Analytics in Practice
Module 3: Session 1
Moumita Sarker
Analytics Application:
What we intend you learn
• Good overview of areas of analytics application in business

• What is the business objective that the analytics solution driving?

• What are the steps / broad framework of the solution?

• What are the techniques that you need to be equipped of to solve or appreciate the solution?

• How do you action the solution? So what happens after we have built this solution?
Which stage in business we would be
covering today?
Generate Leads

Sales
Generate Channel
Right Enablers
Consideration/ pricing
awareness of your
product and interest
via marketing through Right
Promotion
various channels
Efficient
Marketing
Spend
Allocation

Acquire a Customer Once a customer nurture to repeat


Analytics Applications in each of these
modules
1. Measuring the Marketing Channel-ROI

2. Pricing

3. Promotion Optimization

4. Leads Conversion and Nurturing

5. Sales Channel Enablers


Measuring Marketing Channel
Efficiencies
Can you name
channels
where
businesses are
marketing in?
Measurement of Channel ROI: 2 broad areas
• Marketing Mix Modelling

• Single/Multi Touch Attribution


Questions we want Answers to through MMM
ATTRIBUTION
What is the impact of my historical marketing activities on a KPI?
MODEL

If I execute a particular marketing plan, what would be the impact? FORECAST

What is the most optimal way to spend my budget given business constraints? OPTIMIZE
MMM Approach
Tradition Media (TV,
OOH, Radio, Press)

Digital Media- SEM, ROI, Brand paint


Social, GDN For understanding the
impact of past spends
Distribution
Equation
Variable Modelling connecting Saturation Curves
Sales force activities transformations algorithm inputs to sell To understand the impact of
out incremental spends on a
Pricing and Sales channel to sell out
Promotions
Adstock – An Ad shows diminishing impact as
time progresses
Optimized marketing plan
Competitor Actions Optimized distribution of spends
across channels subject to
Input Data Lag – Effect of an Ad is seen after some time constraints (budget, timing etc.)
All possible factors influencing sell
out including different ATL and BTL Shape – The transformation captures this
spends. This can also include non-linear relationship between spend
spends for other models to magnitude and marginal impact
capture Halo effect.
How do we estimate the incremental impact
of an activity?
• MMM utilizes historic movement in revenue to estimate incremental revenue due to marketing
activities
• Multivariate regression analysis allows the simultaneous measurement of multiple drivers of
sales performance, including seasonality, marketing and trade activities, etc.

Incremental sales driven by mass


media can be disaggregated from
90 16 100 base sales using regression
80 14 90 analysis
80

Mass Media Investment ($M)


70
12
Monthly Revenue ($M)

Monthly Revenue ($M)


70
60
10
60
50
8 50
40
6 40
30
30
4
20
20
10 2
10
0 0
0

Jul

Jul
Jan
Feb

Jun

Jan
Feb

Jun
Oct

Oct
Mar

Mar
May

Nov
Dec

May

Nov
Dec
Sep

Sep
Aug

Aug
Apr

Apr
l

l
ar

ar
n

n
ay

ay

v
Ju

Ju

p
Ja

No

Ja

No
Se

Se
M

M
M

Time Time

10
Modeling Process and Rigor
Model Rigor

Performance
Feature Functional
Target Variable Time Duration Model Level Metrics
Selection Transformation
selection
•Relevant
Kitchen Sink •Target
SalesVariables
Revenue •Variables
Adstockare Exploratory data Data points at Statistical tests
variables
Approach
and Vs
areSales
taken,Bills
and •sub-divided
Lag and analyses is taken at required and analyses is
•attributes
Top Bottom
are modelling •analysed
Shape Effect
as per conducted. Training granularity conducted for all
selected
approach
for exercise is built dependence with Period and the models
•model
Spend Vs (Usually
around them
model (one
Mostanother
of the and Validation period Regional Level • R square and
construction
Impression • Off-take
with bills as a independently
Marketing are concluded Product Category Adjusted R-
based on insights
• Salesvariable
target Vol Variables exhibits Level Square
from analyses. National Level
• Top to • Sales Value
performs better) S• shape
Adstock • MAPE
relationship with Month Level
Bottom • Search Intent • Lag Seasonality • Durbin-
approach • Brand Track the Sales)Effect
• Shape Watson
• Walk in Most Variables • P-Value
• Visitors exhibits S shape • Weights for
• Registrations relationship with each metrics
the Sales
Nuances of MMM: Ad-stock and Shape
Outputs: Channel ROI
3,000.00

SEM (Google)
2,500.00
4%
YouTube
2,000.00
Outdoor
Contribution

25% 26%
Facebook 1,500.00

YouTube
TV 1,000.00

18%
500.00
Base 31%
Facebook

- 12%
Jul-17 Sep-17 Nov-17 Jan-18 Mar-18 May-18 Jul-18 Sep-18 Nov-18 Jan-19 Mar-19 May-19 Jul-19 Sep-19
Months SEM (Google)

40% 39% TV
Optimal Range: Channel Sensitivity Curves
SEM is the best marketing channel to invest; the ROI for SEM is about 3 times that of TV (second best to invest)
Change in Vol SU w.r.t SEM Impressions
14
Cumulative Vol SU per Tsd Imp

Peak: 1.1
12
Current: 1.8
10 Min: 0.7 Mil

6 Max: 5 Mil

4 Optimum Reach
Optimal Range
2 Target Reach
-
SI/Off-Take >=6 Mil
0.0 0.7 1.4 2.1 2.8 3.5 4.2 4.9 5.6 6.3
Mil Impressions

Current GRP (Last 6 months, (Oct 19- Mar 20))


Multi-Touch Attribution:
Driving Marketing Efficiency through better attribution
• Advanced data science techniques for accurate attribution
• Driving efficiency through optimal allocation across media

• Media Optimization • Multi-Touch Attribution

Data driven credit allocation to marketing


Estimation of incremental impact of marketing
channels along the consumer journey
activities
Digital Marketing Optimization using MTA
Capture Session Level Data Data Driven Attribution Drive ROI by Efficient Reallocation

Markov Chain,
Shapely Value
Linear Data-
Time-decay driven
Heuristic MTA
Based
Single Touch MTA
Attribution Markov Chain
C C
Convert
1 2

C C
2 3

C C
Convert
2 1
Last Touch,
First Touch
Session Level Data

Shapley Value
Pricing
Differential Pricing
Centralized Pricing and De-Centralized Pricing
Approach: Pricing Dimensions
Given past evidences, a product
Price
observed to be price sensitive should
Elasticity
see lesser price escalations
40%
Similar products that
competition has already priced
higher versus us, have more
opportunity for escalation
Price Top sellers have an
Gap versus Traffic opportunity to observe
Competition Revision Driver higher escalation versus
20% Framework 10%
niche products

COGS
Products with higher
Share
COGS share, necessitate
higher escalation 30%
Computing Price elasticity
• Compute time period for analysis: How long will the price changes have an effect
on sales before business turns back to normal?
• Calculate change in sales in pre and post period of price change to get item-wise
elasticity scores (%Change in demand / %change in price)
Sales drop seen in the
Time at which last immediate 4 weeks
price change was and then tends
done towards Business As
Usual in the next 3-4
weeks.

So, 8 weeks chosen for


analysis - pre and post
price revision date to
compute elasticity.
Scoring: Final composite score and sample data
Composite
Items Comp Premium Elasticity COGS Sales Mix Items
Score

TA CHEESY FRIES 0% 100% 27% 8%


TA CHEESY FRIES 49%
BEEFBALLS 0% 2% 100% 16%
BEEFBALLS 33%
6P DRUMLETS 94% 100% 100% 31%
6P DRUMLETS 92%

Competitor 10P DRUMLETS 100% 0% 100% 24%


Elasticity COGS Sales Mix 10P DRUMLETS 52%
Premium
GARLIC BREAD 100% 0% 18% 100%
GARLIC BREAD 35%
30% 40% 20% 10%
6P WINGS 94% 83% 85% 17%
6P WINGS 79%
10P WINGS 100% 100% 100% 11%
10P WINGS 91%
6PCS BREADSTICK 56% 100% 0% 0%
Weights for the variables 6PCS BREADSTICK 51%
4P CKN TENDER 45% 13% 59% 6%
4P CKN TENDER 32%
Wingstreet 0% 100% 100% 5%
Wingstreet 70%

Normalized scores at
Final composite score
item level
Recommendations: Deciding on Pricing caps and
suggested price changes
Defining Pricing Caps Revised prices are finally rounded up
Minimum (Lower to appropriate values basis additional
0% Flexible depending on the
Limit) business inputs
Maximum (Upper overall impact on %COGS
6% obtained by modelling
Limit)
Composite %Suggested
Items Items
Score Price change

TA CHEESY FRIES 49% TA CHEESY FRIES 3%

BEEFBALLS 33% BEEFBALLS 2%

6P DRUMLETS 92% 6P DRUMLETS 6%

10P DRUMLETS 52% 10P DRUMLETS 3%


Maximum (Upper
GARLIC BREAD 35% 6% GARLIC BREAD 2%
Limit)
6P WINGS 79% 6P WINGS 5%

10P WINGS 91% 10P WINGS 5%

6PCS BREADSTICK 51% 6PCS BREADSTICK 3%

4P CKN TENDER 32% 4P CKN TENDER 2%

Wingstreet 70% Wingstreet 4%


Promotion Optimization
The Objective
Identification of Promotions Scheme / Campaigns Effectiveness to get
Retail Sales Lift
What worked ?– Scheme

What worked Where/Whom? – Dealer, Location

What worked Where and for What? – Product

What worked, where, for whom and When? – Time / Seasonality

• Objective is to evaluate the real sales impact of promotions segregating the impact of other factors
• Identify how promotion schemes fared across dealers, products and time
Promotion Feature Creation and Segmentation

Population Selection

Approach
Sales Decomposition Methodology

Modelling and Recommendation


I. Promotion Segmentation
Promotions to be classified into Schemes basis their characteristics. Say they would be called
Promotion Segments (combination at which impact will be measured)

Promotion Features

Promo Types Promo Intensity Promo Duration Call to Action

Melas
Immediate
High Weekend Offers
Roadshows Delayed
Buy Backs, Upgrades Short Duration, Urgent Staggered
Low
Outdoor Activities Promo Call Outs
Full Month
Image Based
Financial Schemes Various Intensity Bands
Special Occasion Off , Vs. Exchange

Incentive to Dealers Service Based


II. Population Selection for Analysis
Of the available data, relevant population is filtered out and used for this exercise to reduce the noise
present in the data. This is done at 3 levels:

1
• Only dealers with a threshold value of Revenue and Scheme Pay-out are considered.
Dealer • The selected dealer population is then segmented using dealer characteristics

2
• Only products with a threshold value of Revenue and Scheme Pay-out are considered
Product
• The selected model codes are then segmented using dealer characteristics

3
Dealer X Product • “Dealer X Product X Month” combinations with revenue below a defined threshold are
X Month removed from analysis
III. Promotion Variation in a level: Approaches

Can take a call on some


In absence of variability promo experiments to
Check Promotion Feature Check Promotion Feature
within a level, the cluster
Variation at the Variation at the modelling create variability in those
promo recommendations
modelling level level cases, majority will come
can be used
from recommendations

If Cases of Levels with No Cluster Dealers based on


Variation is high (say Location Potential,
>5%) then we need to do Dealer Characteristic, If Cases of Levels with No No additional step of
an additional step of Product Mix, Promotion Variation is low clustering is required
clustering Mix etc.
IV. Sales Decomposition Methodology
• For calculating ROI, impact of promotions needs to be
Impact of segregated from other factors influencing sales
Promotions
• This is achieved through decomposing sales into its various
Product components
Seasonality
Lifecycle • This is done for each dealer segment X product segment
combination and for each promotion segment

Total Sales = Base Sales (SB)+ Seasonality impact (SS) +


Competitor
Base Sales Sales Actions
impact of promotion (SP)

Total Base Seasonal Promotion


Dealer Product Period(i) Sales Sales Sales Sales Payout Promotion
(S) (SB) (SS) (SP)
Sales
D1 P1 Feb 3,070 2,150 +320 600 80 ROI =
D1 P1 Mar 2,150 -110 60 20
Payout
2,100
D1 P1 Apr 2,000 2,150 -230 80 30
V. Modelling
• Sales decomposition to be done separately for each of the Dealer Segment X Product Segment
combinations.

• Within each combination, modelling dataset consisted of records at a dealer X product X month
level.
Sales Decomposition
Methodology

Techniques considered
– CHAID trees, Linear Various lags of Presence of scheme vs Using past sales as a Monthly vs Quarterly
Regression, Time Series promotions / pay-outs Actual pay-out value predicting variable seasonality
Modelling

The final model to be chosen basis fit, business sense and accuracy
VI. Model Selection and Structure
Final solution is a set of equations developed which predicts incremental sales basis
1. Last year sales
2. Quarterly seasonality indicators
3. Presence / absence of various promotion segment

Total Sales = Intercept + (α * Last year sales) Base Sales

+ (β1 * Q1 Flag) + (β2 * Q2 Flag) + (β3 * Q3 Flag) + (β4 * Q4 Flag) Seasonality

+ (γ4 * Scheme Segment 4 Flag) + (γ5 * Scheme Segment 5 Flag)


Promotional Sales

Promotion ROI Promotion Sales γ4


for Scheme
= =
Pay-out Scheme 4 Pay-out
Segment 4
VII. Model Output – Overall (representative only)

y xyz
x Dealers z Months
Products Combos

# Combos # Combos Pay-out Pay-out Inc. Sales Inc. Sales / Inc. Sales/
(with (INR K) (% Revenue) (INR K) Pay-out Revenue
Payout)
Scheme 4 51.1K 24.4K 8,346 2.5% 42,086 5.04 12.7%
Scheme 5 51.1K 5.2 4,990 1.5% 12,938 2.59 3.9%

• Scheme 4 was available in half the cases and has produced incremental sales of 5 X Pay-out
• Scheme 5 was available in only 10% cases and has produced incremental sales of 2.6 X Pay-out

The above analysis could be replicated at a Region, Dealer & other Segments & various timeframes
within the population to identify winning scheme combos
Leads Conversion: Overview
To improve Leads Conversion
Leads Prioritization + Agent Mapping

2. CONVERSION 3. LEAD 4. AGENT SCORING 5. LEAD – AGENT 6. ROLL OUT AND


1. LEADS Database
PROPENSITY MODEL PRIORITIZATION (TME + RM) MAPPING HANDOVER

Client sends leads data All leads scored with Leads prioritized on the - All agents (TMEs Different leads are Template for Impact
in a specified format: propensity to convert basis of score. and RMs) scored mapped to Agent Monitoring. Process
with performance segments for ongoing usage.
- Source | Date - ML or statistical - Scored lists fed to score.
sourced| Lname | methods Dialers on the basis - Agent segments - For example high - Dashboard and
Mobile | Email| - Validation of of score created to be used performing agents reports
Location | models on holdout - Control groups held for allocation are given medium - Report
UTM_Parameters data out which are not as propensity leads conversions,
- Scoring at individual per scores for maximum lift population
record level - Testing in place to stability
- Can guide how check for best
many to target impact mix

FROM: Equal Priority to all TO: Prioritization based on Propensity scores

You might also like