Analytics Consulting IIM Nagpur Slides Used 6 Mar 22

Download as pdf or txt
Download as pdf or txt
You are on page 1of 45

Analytics Consulting

Subhashis Nath
What if the customer hasn’t called
you for a problem?

Typical CXO problem today:

I have a lot of data but I feel I am not making the most of it.
If my competitors make better use of data, they will have an advantage
My competitors are not the old ones, there are new tech competitors
One of the leaders in CPG adopting
advanced analytics and later..
How do analytics consulting projects
start?

As a downstream from business consulting

As a business problem where the client is clear


that he wants it solved using analytics/ AI

When the client calls you and asks for the art of
the possible leveraging the data they have
Steps in analytics consulting are
similar to business consulting

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

Create Convince Monitor/


Manage Course
business and get Pilot program
change correct
case consensus manage
The unknown steps differ

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

Why do
they What is
What have they Get to Root cause
believe Facts / broken,
done about it? the real analysis
AI can Perceptions discover
pain
solve it
Can we
Who Resear
What is the have a
Firm or needs ch on
Has no one cost – internal ballpark
individual? it analytic
faced this? and external ? estimat
done? solutio
e
ns Hypo-
How Do we need to thesize
What is the
difficult Data buy data? Is the Multiple Short and
business
will this map data in line solutions/ Long term
benefit
be creation with problem capabilitie
stated s
Collaborate with the internal data
science team

 Internal teams typically do not scale


 Data science skills are expensive
 Individual organizations have relatively small central teams
 Core and peak – they cannot plan for peak loads
 They still mostly deliver use cases, not analytics products
 They do a numbers of POCs but scaling them in another ball game

They really know the data


- It’s a jungle in there
- Teams say they spend 75% of their times getting the data found, explored and prepared
- Data hunters may already have relationships with providers
- They know how to handle some of the business and IT stakeholders
- To sustain and maintain and review models, the internal DS team will be needed
Analytics was always used in most
industries

Supply chain Supply chain (multi level


Supply chain (uber high
(forecasting, route optimization, real time,
accuracy forecasting, risk
logistics) – Manugistics, simulation for network
scoring, early warning)
i2, Capslogstics, cplex optimization)

Merchandising
Merchandising (store
(recommendation
clustering, allocations, Merchandising (pricing
engines, pricing
markdowns) – SAS, SPSS, and promotion)
automation, assortment
JDA, Retek, Demandtec
automation, RGM)

Marketing
Marketing (customer
Marketing (more firms (Reinforcement learning
analysis, segmentation,
doing it and propensity in promotions, customer
campaigns, media mix) –
and churn) journey, multi-touch
SAS, SPSS, Unica
attribution

Open source Open source


Traditional
use initial use now
When a CXO asks you: what do you
think led to the AI revolution?
What does an analytics consultant
bring to the organization – it’s a
specialization

An analytics consulting
project should never be a
Strategy project alone.

It is always a strategy and


architecture project.
Business, AI and architecture tracks -
intertwined

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

Hypo- Explore data If hypo-thesis Document Make Suggest


Estimate
thesize & proven, use case capability archi-
complexity
solutions architecture finalize (solution) map tecture

Hypo- Run Document Make Suggest


Use industry Estimate
thesize experiment- approach capability possibilities with
templates complexity
solutions al models and model map the models
Data availabilities assessment

 Industry templates – retail will have POS data, telecom will always have
billing data. i.e. some data may be assumed by industry
 Data quality assessment cannot be done as part of consulting. It has to be
very detailed for it to be useful.
 Data assessment is done at a level of metadata and in some cases, sampling
 The firm is not an island. External data exploration is essential. Any
competitor data especially.
 Factor in on-going technical projects.
 Given pressure to deliver quick wins, try to provide solutions that assume
the fundamental pieces rather than waiting on other projects
 Educate client that every modelling exercise needs its own EDA. It cannot be
done for every potential solution except at a very high cost.
Model feasibility assessment

 Most firms using industry experience know which models and solutions
should work in a certain industry
 Check data assessment report so that at a high level there are no surprises
 In case of missing data aspects, try and use surrogates or derived data (e.g.
kids items = family, feminine cosmetics = female, shaving = male, pet
products = pet owner)
 Outline capabilities needed such as real time processing, video feed
processing, IoT middleware, CDP, Programmatic buying, etc.
 Look for more models needing the same capability in the same list of
solutions. Else the capability having to be built for one model is expensive
 In FS/HC, will it be feasible to build models within regulatory standards
 Research on the industry and identify feasibility issues in similar situations
and any innovative solutions used
Estimating business benefits

• Forecasting for a retail store


• Every percentage point of sales reduction in error (e.g. error reduces from 20% to 19% of sales), it
reduces out of stock by 1% of sales – typically 8% becomes 7%
• It reduces shrink by 0.125% of sales – typically 2% becomes 1.875%
• It reduces inventory carrying cost by 0.5% of previous value
• On average, a grocery store level forecasting initiative reduces error by 4%

• Pricing for a retail store


• Typically a price optimization exercise
helps categories increase margins anywhere
from 0.25% of sales to 4% of sales.
E.g. a margin of 20% improves to 20.25% or
upto 24%

• Campaign analytics, segmentation and


recommendation
• Typically improves response rates
anywhere from 1.1x to 3x

 In case of innovation, point out expected


benefit areas and discuss with business
leaders as to extent of benefit
For specific insights, calculation is
based on firm’s revenue

If churn is 10% instead of 20%


for 50% of those who are
churning at 20% today =

5% more customers sticking


with us in that segment.

If that segment today has 5


Million customers and CLTV
is $60

Benefit = 5*0.05*60 = 15M

 Business benefits have to assume that insights are acted upon


Create a score for business benefit, complexity of
architecture and complexity of analytical models

Business Business Tech Tech comp- Model Model Net


benefit $/ % benefit complexity lexity score complexity complexity complexity
score score score
Forecasting 4% revenue 8 9-12 months 9 Feature 9 9
data buying engineering very
cost 100K high. Academic
partnership
preferred
Pricing 2% margin 5 9 months 8 Open source but 10 9
comp survey some products may
25k need IP products

Segment- 3x marketing 3 3 months 5 Open source, 4/9 4.5


ation ROI complex if FLOC
needed
Early warning 2% revenue 5 5 months data 7 Feature 9 8
10% penalties buying cost 50k engineering very
high. Specific event
dictionaries to be
created
Brand Intangible 6 2 months 5 Open source. Pre- 3 4
sentiment 2.2% profit data buying 50k existing
dictionaries
The benefit-complexity map

Fore-
casting

Brand
sentim
ent
Early
Benefit warn- Pricing
ing

Segme
ntation

Complexity
Stop press… what about
dependencies

 What if we said that pricing and forecasting are related and that if pricing is
done, complexity of forecasting becomes half
 Or if pricing is done, benefit from forecasting becomes 50% more
 And if price elasticity analysis is done, it helps segmentation to be better
 Pricing can move up from being the last to become the second thing to take up
 It may even make sense to take up forecasting and pricing together
Corporate appetite

 Cash flow
 Motivation
 Belief
 Stakeholders
 Decision risk – brand of the consultant

Can we break pricing into price elasticity, simulation, optimization. Can we model them in the
same roadmap? Is elasticity the low hanging fruit since it helps forecasting, segmentation and
can also help decisions though not in an optimized manner.
The changed benefit-complexity map

Fore-
casting

Brand
sentim
ent

Price Early
Benefit Price optimiz warn-
elastic- ation ing
Price
simulat ity
ion
Segme
ntation

Complexity
Our coffee firm

Reduce cafe Improve wifi, Improve


Increase café Create quick Increase Improve
expansion or offer printer loyalty
footprint in changing margin on marketing,
new opening services, benefits for
office areas, flavours – smaller sizes, target start
in upscale quiet corners regulars –
universities, e.g. pumpkin for different ups, friend
residential – become an create loyalty
malls latte flavours groups
areas office card wallets

Conduct Conduct social Explore savings


Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)
Complexities from an analytics firm

Conduct Conduct social Explore savings


Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)

Cannibalization Forecast Segmentation Cost analysis Brand sentiment Impact of Impact of


marketing on marketing on
sales sales
3 months build 9 months (20k/m) 3 months build 1 month ($10k) build 2 months build - $10k 9 months build 9 months (20k/m)
Each month – $20k build. 3 months 1 Each month $10k Visibility to cost per month 3 months 1 sub- build. 3 months 1
category. External elements External data cost - brand. subbrand. External
data cost - $100k p.a. movement $10k a year $20k a month data cost - $100k a
year
Run + Build CPU Run + Build CPU Run + Build CPU Run + Build CPU Run + Build CPU Run + Build CPU Run + Build CPU
4 VM D2 4 VM D2 2 VM D2 1 VM D2 2 VM D2 4 VM D2 4 VM D2
20000 S4 HDD HDD and storage can HDD and storage 2000 S4 HDD 20000 S4 HDD HDD and storage HDD and storage of
100M storage be re-used since it’s can be re-used 10M storage 50M storage can be re-used brand sentiment can
transactions per the same since it’s the same transactions per transactions per similar to the one be re-used + 5000 S4
month transactions transactions month month for forecast. HDD 50M
transactions/m
2 year history 2 year history 2 year history 12 month history 3 months history 2 year history 1 year history
Bandwidth Canada Bandwidth Canada Bandwidth Canada Bandwidth Canada Canada central only Canada central Canada Central only
Central to East Central to East Central to East Central to East 1000 only
10000 GB per 10000 GB per 10000 GB per GB per month Use Azure pricing calculator
Benefits discussion

Conduct Conduct social Explore savings


Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)

Core problem Will pay for re- Core to strategy Will pay for re- Pods are doing Will pay for Will pay for
make make well – brand is ok remake remake but
digital spend is
low compared to
traditional
8 7 8 7 3 7 5
Complexity scores

Conduct Conduct social Explore savings


Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)

Cannibalization Forecast Segmentation Cost analysis Brand sentiment Impact of Impact of


marketing on marketing on
sales sales
6 10 5 2 4 8 9
The benefit-complexity map

Segme Cannibal
ntation ization

Cost Fore-
analysis MMM casting

Benefit
MTA

Brand
sentim
ent

Complexity
Benefits if we break some initiatives

Conduct Conduct social Explore savings


Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)

Core problem Will pay for re- Core to strategy Will pay for re- Pods are doing Will pay for Will pay for
make make well – brand is ok remake remake but
digital spend is
low compared to
traditional
8 7 8 7 3 7 5

Forecast for MMM for


higher wastage product and geo
category - 6 that is high
spend - 6
Complexity scores if we break some
initiatives
Conduct Conduct social Explore savings
Forecast at Conduct loyalty Conduct a cost Conduct MTA
cannibalization media analysis by leveraging
store day level data analysis to analysis to analysis if digital
analysis on-going for digital
to reduce milk identify regulars identify marketing is
amongst outlets brand sentiment marketing vs.
and food and in café time potential cost already
in residential by interest traditional
wastage spenders rationalization happening
areas group (MMM)

Cannibalization Forecast Segmentation Cost analysis Brand sentiment Impact of Impact of


marketing on marketing on
sales sales
6 10 5 2 4 8 9

Forecast for higher MMM for


wastage category - 7 product and geo
(less build time, that is high
leverages spend – 5
transaction and (reduced build
storage already time, leverages
needed for transactions and
segmentation storage already
needed for
segmentation)
The changed benefit-complexity map

Segme Cannibal
ntation ization

Cost
analysis
Fore-
MMM casting
selected selected
Benefit
MTA

Brand
sentim
ent

Complexity
Dependency on foundations

 Platforms and people – NLP is a capability, Video analytics, Optimization


is a capability, and so on
 Building capability is the way initiatives progress
 Once you have an NLP COE, you would expect that you try to leverage
the same by finding more COE opportunities
 While creating roadmaps if we believe the firm will do one NLP project
and then will move on, that may be invalid
 Look at capabilities for all ideas that are listed as potential initiatives
Capabilities and initiatives.. Our milk
case

Explore Production
Route
Pilot forecasting Consumer promotion optimization –
optimization Market analysis Move marketing
supply – since its analysis of optimization using optimum
introduced for of margin budgets online
not in control consumption of possible e.g. if water content in
routes with growth trends, after due
and shows a dairy, e.g. high cheese is made cheese – to
especially low RGM analysis analysis (MMM)
trend protein yogurt in high supply improve shelf
orders
times life and cost

Capabilities Route Supply forecast RGM Consumer Promo MMM H2O


optimization analytics optimization optimization

Optimization * * *
Real time * *
Forecasting * *
Nielsen skill * * *
Consumer *
analysis
Elasticity * *
Marketing *
analysis
IoT *
Predictive maint *
Deriving the capability roadmap

Consumer
Forecasting Optimization Real time Elasticity Nielsen skill
Analysis

Capabilities Route Supply forecast RGM Consumer Promo MMM H2O


optimization analytics optimization optimization

Optimization * * *
Real time * *
Forecasting * *
Nielsen skill * * *
Consumer * *
analysis
Elasticity * *
MMM/MTA *
IoT *
Predictive maint *
Toolkit to identify initiatives

How analytics/ AI is applied in firms

1. Diagnostic
2. Predictive
3. Expert
(e.g. patterns in past and multi-source data)
4. Cognitive
5. Anomaly detection
6. Clustering or grouping
7. Decision Granularity or scale
8. Propensity
9. Auto-learning
(e.g. an ad engine that changes ads based on customer reaction)
10. Optimization algorithms
A logistics case

 European logistics provider


 Pratt, who leads operations, comes to you for help
 Unable to meet time commitments, late deliveries
 Customers are committed delivery times based on an online estimator that customers can use
 He doesn’t know what is the issue but has data
 Loads come in peaks and troughs, peak time is tough to meet
 Sorting issues, wrong bar code pasted on package etc.
 Problem is more in some regions and less in others
 Storage depots are running at capacity
 Typically, delays are seen more in large projects
 They have a depot in every state
 More depots in north-east and west.
Structured approach

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

What has Pratt done to tackle the problem so far?


Why is
do they
What have they Why did those efforts fail?
think AI
done about it?
can
solve
Who While delivery times are not met, what is happening to firm
Firm or needs
individual? it sales?
done?
Is what we find, aligned to what one might expect?
How
difficult Business
will this
be?
benefit
Immediate pain and also long term impact. Difficulty rating of
some initial thoughts, like robotic sorting using image analytics.
Structured approach

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

The failure of current attempts – do they indicate


anything about more problems? Get to
Facts /
Is delivery time window a risk the real
pain
Perceptions Market data past and current,
accepted by leadership what is happening to competition?
Has no one Resear How are competitors handling it?
faced this? ch
Metadata available?
Data
EDA? map
Data to clarify
veracity of
Capacity utilization of trucks creatio
n
symptom
Structured approach

List Frame and Plan


Discuss Understand symptoms/ Diagnose reframe the solution
pain points problem roadmap

List components of the problem


What is there in the data that helps isolate the problem? What is
Root cause
broken,
Rule out components not supported by data discover analysis

Identify further data needed to confirm/ rule out components Can we


What is the have a
cost – internal ballpark
Having componentized the problem, ask again if it is feasible and external ? estimat
e
to solve it. Hypo-
thesize
Cost? Multiple Short and
solutions/ Long term
capabilitie
s
Hypotheses generation – in collaboration with client, now that problems may
be clearer and in parts.
How do we resolve it from the root? How do we resolve immediate pain?
Breaking it down

Delivery time
SLA not met

Over
Forecasting
committed
problem
against reality

Based on a
Strategy Fluctuation Our models
tool?

Unchanged Provide +/- Classify types Different


Tool not Do more real
Tool bug For more sales from previous Inertia leeway in of package types of Re-check
known well time
KPIs commitments and tell delay customers

Analyse the “estimator”.


How are delivery times
calculated?
Just like customer journeys, products
have journeys, isolate the problem
Compare estimated time and actual time at
stages
Toolkit

Build a diagnostic tool that continuously identifies


bottlenecks
How analytics/ AI is applied in firms Improve forecast of delivery time

Use all data available including population, fleet,


1. Diagnostic package size, locations etc. to identify types plagued
2. Predictive with delays
3. Expert Use image analysis for auto sorting
(e.g. patterns in past and multi-source data)
4. Cognitive Find anomalies in journey of packages that were
inordinately highly delayed
5. Anomaly detection
Segment partners or locations based on parameters,
6. Clustering or grouping with target variable as efficiency
7. Decision Granularity or scale Make estimator work using package level and zip-code
8. Propensity level data and not average across a city or region

9. Auto-learning Leverage 3 and build propensity models to show


inclination for delay for a package route combination
(e.g. an ad engine that changes ads based on customer reaction)
Not applicable right now
10. Optimization algorithms
Explore route optimization and network optimization
Network optimization

Major

Incremental
Initiatives
Business Business Tech Tech comp- Model Model Net
benefit $/ % benefit complexity lexity score complexity complexity complexity
score score score
Forecasting Likely to 7 9-12 months 9 Feature 9 (8 if 8.5
(closer to real reduce data buying engineering very estimator
time than delivery flaws cost 100K high. Academic already done)
estimator) by 10% partnership
preferred
Propensity - 5% delivery 4 3 months 3 Open source. Pre- 3 3
light impact existing
dictionaries
Propensity - 10% delivery 7 3 months 5 Open source, 7 (5 if light 5
expert impact Ext data 50k, complex and data already done)
overlap w/ and compute heavy
forecast data
Making 40% impact 8 5 months data 7 Feature 9 8
estimator on buying cost 75k engineering very
granular expectation But overlap w/ high. Specific event
setting forecast data dictionaries to be
created. Will lead to
forecasting.
Network 15% impact on 9 9-12 months 10 Open source but 9 9.5
optimization deliveries data buying also may need IP
150k products
The benefit-complexity map

Network
optimiza
tion
Estimat
or

Propensi
ty - Fore-
expert casting

Benefit
Propen
sity -
light

Complexity
After current problem – top 15 AI use
cases in logistics firms
Deriving the capability roadmap

Propensity Pattern Forecasting Real time Optimization Anomaly Cognitive

Capabilities Forecasting Propensity - Propensity - Estimator Network Extreme delay Auto-sorting


light expert optimization cause

Optimization * *
Real time * *
Forecasting * *
Pattern * *
recognition
Anomaly *
detection
Propensity * *
Cognitive vision *
Thank You!

You might also like