Lecture 1 - Uncertainty
Customers are assets that generate profits over time. Marketing used to be product-centric and
transaction-focused, and is now customer-centric and relationship-focused.
Customer lifecycle:
1. Customer acquisition: how customers are born / first contact with the firm
2. Customer development: change in behavior over time; buying more (upselling) or different
things (cross-selling)
3. Customer retention: preventing customer death or churn
Testing
Testing: obtaining more information before committing a large amount of resources, and (hence)
reducing the risk of possible failure
1. Randomly select some customers (test sample) (size = n)
2. Send these customers a mailing, and collect and analyze responses
3. Use results to decide whether to send to the rest of the population ( rollout sample) (size =
N - n)
Sample should be representative for people outside of the sample (randomized sampling).
Test results
→ Assume a test sample of size 5000; thus, randomly selecting 5000 customers and sending them the
mailing
→ Results of test mailing
→ 175 out of 5000 respond; estimated response rate: p̂ =
→ Margin/profit per response is €50 (assume): m = 50
→ Should the rollout be done? How much profit is expected if it is sent to the rest (rollout sample)?
E(rollout profit) = (N - n)(m * p̂ - c)
→ (N - n): number of customers
→ (m * p̂ - c): profit per customer
→ m: margin (profit) per response
→ p̂ : estimate of response rate
→ c: cost of marketing
→ Only roll out when E(rollout profit) > 0 (thus, p̂ > c/m)
Option Value
When the expected rollout profit is positive, it is rolled out to the rest of the sample.
Roll out when: E(rollout profit) > 0 (p̂ > c / m)
→ Bad campaigns are only tested, good campaigns are tested and rolled out (when profit per customer is
positive = p̂*m*c)
, 1. Assume perfect information
2. Test predicts
→ Success (p = 0.05); m * p - c = 1.00
→ Failure (p = 0.01); m * p - c = -1.00
3. Success occurs 30% of the time
Limiting losses to the test means only losing €5000
No-test option = €0
Test = €11,500 (if cost of running the test would be
lower, you wouldn’t do it)
Uncertainty
p: true unobserved population response rate
Sample mean estimate: (what we observe)
Standard error:
Central limit theorem: for a large enough sample, distribution of the sample mean is approximately
normal
Probability of a mistake:
Bootstraps
Bootstrap: sample with replacement from the original sample, using the same sample size (imaging
what other samples would give you)
→ b = 1 gives you B bootstrap samples (some are sampled often, some aren’t at all)
1. Resample with replacement
2. Calculate estimate using this
resample set
, → You now have a distribution
Test whether the estimated
Where
Alfa = response rate < breakeven (type I error; rollout while
you shouldn’t)
Instead of using this “all or nothing” approach, we can also use the test to identify profitable groups
and target mailing to them (test sample, targeted rollout sample (sent), untargeted rollout sample
(not sent)).
Data to use:
Most common
1. Demographics (gender, ethnicity, age, income, family size, occupation, marital status,
education, homeowner/renter, length of residence)
2. Transaction data (past purchases, amounts, dates, discounts)
Best, but unavailable for prospects
3. Marketing (past mailings, content mailings, date, costs)
4. Survey data (psychographics, attitudes, interests, activities)
Even if the untargeted mailing campaign would be profitable, selecting customers usually is more
profitable.
Thus, testing resolves (some, usually not all) uncertainty about the benefit of marketing. Testing
gives the option to rollout if test results are positive, and is even more valuable when you use it to
target better.
5.
Customers are assets that generate profits over time. Marketing used to be product-centric and
transaction-focused, and is now customer-centric and relationship-focused.
Customer lifecycle:
1. Customer acquisition: how customers are born / first contact with the firm
2. Customer development: change in behavior over time; buying more (upselling) or different
things (cross-selling)
3. Customer retention: preventing customer death or churn
Testing
Testing: obtaining more information before committing a large amount of resources, and (hence)
reducing the risk of possible failure
1. Randomly select some customers (test sample) (size = n)
2. Send these customers a mailing, and collect and analyze responses
3. Use results to decide whether to send to the rest of the population ( rollout sample) (size =
N - n)
Sample should be representative for people outside of the sample (randomized sampling).
Test results
→ Assume a test sample of size 5000; thus, randomly selecting 5000 customers and sending them the
mailing
→ Results of test mailing
→ 175 out of 5000 respond; estimated response rate: p̂ =
→ Margin/profit per response is €50 (assume): m = 50
→ Should the rollout be done? How much profit is expected if it is sent to the rest (rollout sample)?
E(rollout profit) = (N - n)(m * p̂ - c)
→ (N - n): number of customers
→ (m * p̂ - c): profit per customer
→ m: margin (profit) per response
→ p̂ : estimate of response rate
→ c: cost of marketing
→ Only roll out when E(rollout profit) > 0 (thus, p̂ > c/m)
Option Value
When the expected rollout profit is positive, it is rolled out to the rest of the sample.
Roll out when: E(rollout profit) > 0 (p̂ > c / m)
→ Bad campaigns are only tested, good campaigns are tested and rolled out (when profit per customer is
positive = p̂*m*c)
, 1. Assume perfect information
2. Test predicts
→ Success (p = 0.05); m * p - c = 1.00
→ Failure (p = 0.01); m * p - c = -1.00
3. Success occurs 30% of the time
Limiting losses to the test means only losing €5000
No-test option = €0
Test = €11,500 (if cost of running the test would be
lower, you wouldn’t do it)
Uncertainty
p: true unobserved population response rate
Sample mean estimate: (what we observe)
Standard error:
Central limit theorem: for a large enough sample, distribution of the sample mean is approximately
normal
Probability of a mistake:
Bootstraps
Bootstrap: sample with replacement from the original sample, using the same sample size (imaging
what other samples would give you)
→ b = 1 gives you B bootstrap samples (some are sampled often, some aren’t at all)
1. Resample with replacement
2. Calculate estimate using this
resample set
, → You now have a distribution
Test whether the estimated
Where
Alfa = response rate < breakeven (type I error; rollout while
you shouldn’t)
Instead of using this “all or nothing” approach, we can also use the test to identify profitable groups
and target mailing to them (test sample, targeted rollout sample (sent), untargeted rollout sample
(not sent)).
Data to use:
Most common
1. Demographics (gender, ethnicity, age, income, family size, occupation, marital status,
education, homeowner/renter, length of residence)
2. Transaction data (past purchases, amounts, dates, discounts)
Best, but unavailable for prospects
3. Marketing (past mailings, content mailings, date, costs)
4. Survey data (psychographics, attitudes, interests, activities)
Even if the untargeted mailing campaign would be profitable, selecting customers usually is more
profitable.
Thus, testing resolves (some, usually not all) uncertainty about the benefit of marketing. Testing
gives the option to rollout if test results are positive, and is even more valuable when you use it to
target better.
5.
