1505 Statistical Thinking course extract

Copyright © 2012 Red Olive Ltd, All Rights Reserved. 11
Analytics Academy – “Statistical
thinking”
(Client: a household-name media company)
Information and
Data Management

Contents
1st Day
Introduction
The research process:
CRISP-DM
Analysis
Reporting vs. modelling
Is there an effect?
Is there a single cause?
Forecasting
Could there be more than
one cause?
2nd day
Working together
The Data Academy
Sharing results
What to show
How to show it
Next steps
A further project

Introduction:
Getting your data to speak

Customer Acquisition & Retention
Marketing Efficiency & Advertising Revenue
Cost to Serve & Profitability
Promotion & Pricing Optimisation
Demand Forecasting
Fraud Detection
Many business challenges
4Copyright © 2012 Red Olive Ltd, All Rights Reserved.

Do you know what you are looking for?
The business need for the analysis is formulated and confirmed.
The question that the stakeholder needs an answer for is articulated.
Steering away from analysing ‘right answer to the wrong question’.
Do you know what you will do with the answers you find?
The desired outputs from the analysis are shaped in detail to ensure
that the analysis produces outputs in a format that is fit-for-purpose.
Actual outputs can be easily integrated into the stakeholders’ target
documents, systems or processes.
Do you have a way to evaluate success?
Can you measure the current situation in terms of money, time or
units?
Do you have a way of tracking the results of your work in the same
units?
Before jumping into the data...
Copyright © 2012 Red Olive Ltd, All Rights Reserved.

Gather ideas from people in your business about the
cause –> effect relationships.
Gather impressions about the different classes or
types of events.
Consider both positive and negative outcomes.
Translate these ideas/impressions into data
What would data have to look like to detect the effects and
trends people believe in?
Translate business objective into analysis goal…
Getting your data to speak

The research process:
CRISP-DM

CRISP-DM process
Business data
for analytics
1 Develop
business
understanding
2 Develop data
understanding
3 Prepare data
4 Develop
model
5 Evaluate
results
6 Deploy live
model
Key:
Data set
Process stage
Flow between
stages

CRISP-DM
Business understanding
Determine objectives
Establish use cases
Summarise current situation
Determine project goals
Map business goals to data
problem
Estimate current value so that
ROI can be calculated
Create project plan
Data understanding
Collect initial data
Document the real meaning
of each data field
Capture baseline SQL
Explore the data
Check data quality

Analysis:
Reporting vs. Modelling

vs.
What was the rate of net growth? Why did we have higher/lower rate?
Information based on user-directed queries
(hypothesis testing)
Knowledge based on finding unknown
relationships (hypothesis generation)
Historical Analysis Predictive Analysis
Monitors performance measures Determines performance measures
Reactive Proactive
Reporting vs. Modelling
ModellingReporting

Data Input Target Output Algorithm Goal Results
Find Most
Important
Inputs?
Easy to
interprete/
visualise?
Numeric and/or Symbolic Symbolic C5.0 Predicting / Profiling Decision Tree or Rule Set with prediction and confidence Yes Yes
Numeric and/or Symbolic Numeric or Symbolic C&RT Predicting / Profiling Decision Tree or Rule Set with prediction and confidence Yes Yes
Numeric Numeric Linear Regression Predicting / Forecasting Equation for prediction with coefficients Yes No
Numeric and/or Symbolic Symbolic Logistic Regression Predicting / Probability Equation for prediction of probability and associated coefficients No1
No
Numeric and/or Symbolic Numeric or Symbolic Neural Network Predicting / Probability Prediction and relative importance of input neurons No2
No
Numeric and/or Symbolic None Kohonen Map Clustering / Segmentation Cluster Membership and deviation No Yes
Numeric None K-Means Clustering / Segmentation Cluster Membership with cluster centers No Yes
Numeric and/or Symbolic None Two-Step Clustering / Segmentation Cluster Membership with cluster centers No Yes
Symbolic Symbolic Apriori4, 5
Association Detection Association rules with confidence Yes Yes
Numeric and/or Symbolic Time to event Kaplan-Meier Strategic Planning Survivor / Hazard Curve No Yes
Numeric and/or Symbolic Time to event Cox Regression Tactical Interventions Survivor / Hazard Curve No Yes
Sometimes we put the data into the model and see what happens. Other times we
manipulate the inputs (or the outputs) in some way so as to give the algorithm more
information to work with.
By combining multiple techniques, we can often gain better insight into the nature of
potential solutions to a business problem and hopefully lead us to a more useful result.
Since more than one approach may be used to address a single business problem, the
same data may be used to address a wide range of applications. It will depend on which
model you choose, how you manipulate the data in the file, and which input or target
variables you choose.
Map analysis goal to
modelling technique

Analysis:
Basic statistical terms

Level of measurement
Level of measurement Summary
statistic
Visualization
Categorical or Nominal Mode Bar chart, pareto chart
Ranked or Ordinal Median,
percentile
Bar chart
Numeric or Scale Mean or
average
Histogram, line graph,
bubble chart
0
5
10
15
20
25
30
35
40
45
50
1st Qtr 2nd Qtr 3rd Qtr 4th Qtr 0
5
10
15
20
25
30
35
40
45
50

Inserting functions
Click in a cell
Go the Insert menu
Choose Functions…
Select a category
Click on a function
and look at the
brief help
(first letter search
works)
Click OK to paste
Click
Help on this function
for more information
and a worked example

Numbers that describe
a distribution
Statistic Function Definition
Mode =MODE The most common value
Percent =COUNT What proportion of the
cases are in this group?
COUNT in the group
divided by total COUNT.
Percentile =PERCENTILE
=PERCENTILERANK
How far down the list of
an ordered set are you?
Median =MEDIAN The middle value of an
ordered set. The 50th
percentile.
Mean =AVERAGE Add all the values and
divide by the count

Analysis:
Is there an effect?

Discussion of crosstabs
A method to test if two variables have
a non-random relationship
Also called chi-square analysis for the name of the
statistic that is calculated
Χ2 or X2

Discussion of crosstabs
Is there a relationship
between the section you
are reading on the website
and whether or not you are
motivated to subscribe?
Or are the numbers just due
to the normal visit pattern
on the site?
Data:
Subscribe y/n on this visit to
this section
Section
YES NO TOTAL
HOME
NEWS
SPORT
FINANCE
COMMENT
BLOGS
CULTURE
TRAVEL
LIFESTYLE
FASHION
TECH
Offers
TOTAL

Crosstabs example
Actual counts
Calculated %

Analysis:
Is there a single cause?

When we seek to predict something, what we are
really saying is that we have in mind what the cause
is and we are trying to predict how likely the effect
is.
Modelling techniques do not make predictions on
their own. Analysts structure the data input so that
the model can use it in a cause and effect way.
Thus, it is important to make sure that all of the
inputs into a model precede the output in time. You
can’t put the effect before the cause.
22
Predictive modelling

Try to get to models early in the process, even before
you think you are ready.
Models can tell you things about the data that you
can’t see “just by looking”
Build lots of models.
Throw away the ones that you are done with
Refine models based on what you learn at each
iteration.
Algorithms (within their limitations) are objective
Interpret the results, then make
them better

Could there be
more than one cause?
A topic for another course…
Advanced analytics
Structure the data into
before and after
Pick a target
Test multiple input
hypotheses at once
Forecasting:
ARIMA allows for
including multiple time
series inputs
Special events
Weather
Economic trends
Multivariate propensity
Discover different
predictive segments
Works best with
predicting Y/N actions
rather than values

Sharing results:
What to show

Your job is to inspire
Your job is not to convince or teach
Lead with the important and interesting findings
Explain in general terms
Leave the details at the Data Academy
Inspire them

Business
Objectives
Analysis
Results
Business
Terms
Modelling &
Evaluation
(Accuracy &
Significance)
Meaningful
Relevant
Actionable
Quantified
Translating analysis results into
business terms
27Copyright © 2012 Red Olive Ltd, All Rights Reserved.
Analysis
Goals

How you did it
How long it took
Statistical methods
What to leave in
the Data Academy
Problems you had
Caveats related to data
Dirty data
Analysis
Goals
Analysis
Results
Modelling &
Evaluation
(Accuracy & Significance)

Sharing results:
How to show it

Can everyone read it
Is someone color blind?
Does someone have corrective
lenses?
Will it print in black and white?
Test print
Black text on dark colors,
including red, will not print.
Use white text instead
Wrong Better
Better
0
10
20
30
40
50
60
70
80
90
100
1st Qtr 2nd Qtr 3rd Qtr 4th Qtr
East
West
North
Not as Nice
0
10
20
30
40
50
60
70
80
90
100
1st Qtr 2nd Qtr 3rd Qtr 4th Qtr
East
West
North
Design – Colour

Design –
Colour for the colourblind
• http://www.colorbrewer2.org/

Contact information
Please direct enquiries to Jefferson Lynch:
jefferson.lynch@red-olive.co.uk
Office: 01256 831100
Mobile: 07860 353027
32

1505 Statistical Thinking course extract

More Related Content

What's hot (20)

Similar to 1505 Statistical Thinking course extract (20)

1505 Statistical Thinking course extract