Data Driven Approach: Advanced Analytics & Machine Learning in the Non Profit sector. Davide Camera - Excelle Srl
1. Data Driven Approach
Advanced Analytics & Machine Learning in the Non-Profit sector
DAVIDE CAMERA | MAY 18, 2018 | ROME
2. Davide Camera
Seventeen years of experience
in Data Science, in particular
on areas related to Business
Intelligence, CRM, Data
Mining, Customer Experience,
Campaign Management,
Customer Care, Loyalty
Program, Advanced Analytics
& Machine Learning.
Founder of Excelle
(www.excelle.it), consulting
company working in the Data
Science and Data Driven
Marketing field.
Currently CEO, teacher and
lecturer at Data Driven
Marketing conferences.
https://www.linkedin.com/in/davidecamera
3. Who we are
Excelle was founded in 2012. It is a consulting company working in the Data Intelligence
field. We are able to:
structure and manage both “smart data” and “big data” integrating conventional and digital
channels
understand data, identify strategies, develop and monitor marketing plans.
We implement strategies, action plans, algorithms, statistical and mathematical techniques to
increase «Profits» and «Marginal Revenues» for our Costumers
4. “Data Intelligence is about transforming data into information, information into
knowledge, and knowledge into value”
–quora.com
5. We develop Data Intelligence projects
The transformation of data into information, information into
knowledge and knowledge into value.
Data intelligence represent a facilitator of a culture and a
business strategy based on knowledge deriving from data
collection, reading, analysis and interpretation.
Data Intelligence becomes one of the main digital
transformation driver for companies.
The main company challenge is to gain competitive
advantage working on data.
Data Intelligence is...
6. Data Intelligence is not a tool or a technology
Big Data does not mean Data Intelligence
Smart & Big Data and Technology are some of the
fundamental elements for a business approach based on
Data Intelligence.
Analytics are not “data mixers” where solutions come out
magically
Without goals, paths and methodology the results
obtained are bias
Data Intelligence is
NOT...
We develop Data Intelligence projects
7. ELEMENTS
When we talk about Data Intelligence we immediately
think about technological components or mathematical /
statistical skills.
Actually, Data Intelligence consists of numerous
elements, including:
Strategic and operational objectives
Methodologies
Expertise
Technology
Data
Mindset
Creativity
Actions
Deploy and Delivery
A / B test
Measurement
...
8. BARRIERS AND
LEGENDS ABOUT
DATA
INTELLIGENCE
There are some barriers and legends about Data
Intelliugence:
Lack of internal skills
Data Accessibility
Data Reliability
Lack of technology
Company not big enough
Huge expenditure of time
The results do not reveal anything new compared to
the one already known in the company
9. Analytics Assessment
A timely diagnosis of the current
analytical situation and a future
roadmap with the benefits that
can be obtained if methodologies
and processes guided by data
analysis are adopted.
Training
“Accademia del Dato” was
founded to disclose
knowledge towards actors
interested in the magical
world of Data Science, a
central element for a
business culture based on
Data.
Training courses for
Managers and Data
Scientists. Each course can
be customized according to
the needs and objectives of
the customer.
Descriptive
Analytics Tables, graphs
and main KPI's for an initial
description and knowledge of
customers.
Reporting & Monitoring
Understand and make sense of
business based on past events.
Ex .:
reporting by product / customer /
sales network
directional dashboards
Etc.
Predictive Analytics &
Machine Learning
Advanced predictive models that
span the whole life cycle, from
acquisition to win back.
Forecasting and what-if models
Marketing Mix
Model
Analysis to understand:
What is the ROI of
marketing and
communication activities?
How much do they
contribute to sales?
What is ROI per channel?
What is the ROI per
customer?
How long do I reach the
break even point?
Etc.
Marketing (Engagement Plan, Loyalty & CRM Program, Fundraising, etc.)
Sales
Digital & Innovation
Comunication
Data Science +
Market Research
Our partnerships with
research institutes allow us
to strengthen the Data
Intelligence offer by
integrating it with evidence
coming from outside the
company (eg finding
strategic targets even
when information on
corporate DBs is missing).
Area & Service
13. Donors Journey
Acquisition Value Stimulation
Up sell. & Share of
Wallet
Churn
Prevention
Churn Retention
Dormienza
NEW Deactivation
Win Back
14. 01 Donors
Analysis 02 Campaign
Analysis
04 Win Back
03 Churn
Prevention &
Retention
05 Media Mix
Model
No Profit Analytics
15. Automated reporting & dashboard
Donors clustering & profiling
Machine Learning to prevent churn
Machine Learning to identify donors journey
Automated Forecasting based on past & future marketing campaign
Machine Learning to optimize touch point & channels
Machine Learning to identify donors to have interest to growth
donation value (up-selling)
Machine Learning to identify donors to have interest to do another
donation (cross-selling)
Media Mix Model to evaluate advertising investment
No Profit Analytics &
Machine Learning:
PROJECTS
17. INITIAL
CONSIDERATIONS
ON PREDICTIVE
MODELS
Each predictive model returns a result, but not all
results are useful. You can create models with 2
variables (es. gender, regions) but is this what you
need?
The Data Management & Data Preparation phases set
up the basis for a “winning” predictive model.
The context and trend analysis, the definition of the
Costumer Base (CB), the knowledge of the main
metrics about the Customer Base performance are
necessary conditions for developing a predictive
modelling.
The human component, the methodological paths, the
mathematical and statistical know-how, the sector and
business knowledge are indispensable and not yet
replaceable factors.
20. CHURN
PREDICTIVE
MODELS
The churn model’s aim is to determine, for each
Customers, the probability of abandonment/closure.
Churn analysis is strongly based on the definition of the
"abandonment / closure phenomenon" (in some cases it
may involve the analysis of total abandonment, in others
of the switch, of contractual downgrading, etc.)
In churn models, the variable to be predicted is
dichotomous (0: not abandonment 1: abandonment); the
Churn predictive model thus returns a probability
between 0 and 1 that the event takes place.
21. METHODOLOGY
CRISP-DM includes 6 not temporally-bounded
phases, which follow each other. It is always possible to
return to the previous phase or review the ideas on the
basis of analysis in view of the results at any point in the
process.
The process is not a linear process, rigidly
characterized by a beginning, a certain number of
predetermined steps and an end, but it is a cyclical
process, which becomes point of reference for possible
future projects.
22. DETERMINISTIC
ANALYSIS
Based on the project objective, preliminary context analysis
are performed.
The goal is to obtain useful insights for the development of a
predictive model and for improving the identification of CB
and target.
23. In the below dashboard the customers linked to the main regions in the «X» period.
% risk
Riskevolutiion
Livello rischio provincia in decrescita stabile in crescita % clienti
Minor
rischio TN 2,4% 0,4%
VR 2,9% 1,6%
PD 3,2% 1,5%
BO 3,4% 2,6%
AN 3,5% 1,5%
TO 3,7% 2,8%
MI 3,8% 5,4%
RM 3,9% 10,1%
rischio 41,6% 49,4% 11,0% 40,0%
Medio
rischio VA 4,0% 0,8%
CA 4,6% 1,5%
SA 4,8% 2,6%
VI 4,9% 1,5%
CE 5,0% 2,0%
NA 5,2% 4,6%
FI 5,3% 1,8%
rischio 37,7% 101,6% 33,8% 34,6%
Maggior
rischio RC 5,6% 0,3%
PI 5,6% 2,5%
GE 6,1% 6,1%
LU 6,3% 3,1%
MS 7,4% 2,0%
VE 8,1% 1,8%
Maggior rischio Tota 17,2% 50,8% 56,8% 24,2%
% Rischio
Risk Trend
24. Digital C&D - All Rights Reserved
t0 t2 t3 t4t1tm1……tm12 tm2
T-1 (past) T0 (present)
Historical Datasets Start Forecast Target
T+1 (future)
Target and Costumer Base analysis
26. ALGORITHM
CHOICE
There are many factors contributing to the algorithms
choice: not only mathematical/statistical but also business
factors.
For example, neural networks can be a good prediction
model, but they do not provide details on their elements
(which variables estimate it, the weight they have, etc.).
The use of regressions or decision trees provides information
on the elements (with different level of detail) employed in
predicting processing.
The final predictive model is created by testing different
algorithms (logistic regression, neural networks, decision
trees, etc.) in order to choose the most "performing"
model.
Different algorithms can be developed for different objectives.
27. Digital C&D - All Rights Reserved
A
Main Algorithms
• Logistic regression
• Decision trees
• Neural Networks
• Principal Component
Analysis
• Association Analysis
• Hierarchical clustering
• Clustering K-Means
• Support Vector Machine
• ….
B
Type
• Supervised Learning
• Unsupervised Learning
• Semi-supervised Learning
C
Evaluation
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
1,00
2,00
3,00
4,00
5,00
6,00
7,00
8,00
9,00
0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100%
Performance of training vs test
TRAINING
Cumulative LIFT
TEST
Cumulative LIFT
Available algorithms
29. The information were studied and optimized first in a univariate logic and then in a multivariate logic in
order to derive the maximum information and the maximum predictive capacity.
Database :
DB divided into Training &
Validation
DB TEST
Models
Logistic regression
Decision trees
Neural Networks
Model Choice
Logistics guaranteed:
Best absolute performance
Best performance in passing
from Training to Validation
Confirm Results
The DB TEST confirmed
the results obtained
during the development
phase.
Predictive model flow and choices
30. MAIN OUTPUT
Context and trend
Focus on segments and relationship with the
abandonment phenomenon
Report about the phenomenon with the main metrics
(eg seniority, customer value, CLTV, ...)
Measurement of quality and potential risk on new
donors
Risk factors
Declination of risk factors on company processes (eg
caring)
Probabilistic scoring
Management probabilistic traffic light for simulations
and operational management
31. The scoring produced by the churn model is the probability
assigned to a given customer to deactivate itself in a given
time frame.
On this score, the «management probabilistic traffic
light» is subsequently constructed
Probabilistic Scoring
32. % of Captured Phenomenon:
According to the score assigned to the client, it is now possible to identify among high-risk level clients those that guarantee
the best balance between true and false positive.
Once customers have been sorted by score, it is possible to divide the population into deciles and observe the % of actual
churner in order to assess at what level of risk to manage the client (trade off volumes - risk intercepted)
Management probabilistic traffic light definition
Management Probabilistic scoring
TAG Innovation School - all right reserved
33. Identified risk factors (area Customer Care)
Contracts details last 12
month
Contracts details last 6
month
Increase/Decrease the churn
rate
Information about
procedures of
withdrawal last 2 month
Information about
procedures of
withdrawal last 3 month
Call complaints
last 6 month
Call complaints
last 12 month
Call for being in
arrears
last 3 month
High-risk level
contacts
last month
Medium-risk level
contacts last month
Call for lack of
communication
last 2 month
Contacts for «transfer»
last 3 month
35. ANSWERS
OPERATIONAL OBJECTIVES
1. How much is the churn risk? The churn risk is about XX%
2. What affects the churn? customer engagement, donation
frequency, donation methods, previous insolvency,
donations for multiple "products" and more
3. How can I manage the churn abatement? Thanks to the
deliverables of the predictive model (management score,
management probabilistic traffic light, actions on single
abandonment factor)
4. Is a timely communication important to establish a loyal
relationship with donors? YES
36. 1 year after…
“double digit” churn reduction
Donation value increases several hundred thousand
euros
INCREASE FUNDRAISING
LOSING LESS DONORS