MKT ANALYTICS - Final Exam

what is the CLV (customer lifetime value)

NPV of all future profits a customer generates over life of their relationship with the company

ingredients of the CLV formula

m = margin (profit)

r = retention rate

i = discount rate

AC = acquisition cost

g = growth rate

CLV formula (no growth rate)

m(r/(1 + i - r) - AC

margin multiple

(1 + i - r)

CLV formula (with growth rate)

m(r/(1 + i - r(1+g))) - AC

effect of m on CLV

increases (more profit increases CLV)

effect of i on CLV

decreases (more discount decreases PV for future sales)

effect of r on CLV

increases (customer staying longer leads to higher profits)

effect of AC on CLV

decreases (subtracted from CLV)

effect of g on CLV

Increases (if margin grows over time, value increases)

most important CLV component for firm value

retention rate

unsupervised learning

no outcome value (Y is unknown) - used for finding patterns + clustering

supervised learning

input values have outputs (Y is known) - used for classification + regression

overfitting

when the model works perfectly with training data but poorly with testing because it learned the exact pattern of the training data

what prevents overfitting

data splitting (split data into training, validation, testing)

what does CART stand for

Classification And Regression Trees

Confusion Matrix

	X	Y
X	True Positive (TP)	False Negative (FN)
Y	False Positive (FP)	True Negative (TN)

compute hit-rate (accuracy) using confusion matrix

accuracy = (TP + TN) / total observations

how to identify what model delivered higher AUC using ROC curves

1. The higher the area under the curve the better (has a higher AUC)

2. ROC curve compares TP vs TN

when is AUC more appropriate compared to hit-rate (accuracy)

AUC = more appropriate when classes (Y) are unbalanced

• ex: 95% zeros, 5% ones

type 1 of ensemble-learning

bagging - separate data into model subsets & aggregate predictions of individual models using majority vote

type 2 of ensemble-learning

boosting - model = trained sequentially and misclassifications are weighed more heavily in next model

what type of ensemble-learning is XGBoost?

tree-based

does contextual targeting perform better than behavioral

no

contextual < behavioral < full

critical requirement for A/B testing approach to be valid

randomly assigning participants

is the demand curve usually upward/downward sloping

downward sloping (higher price = less demand)

where on the demand curve is the niche segment

top left (high price, low demand)

where on the demand curve is the mainstream segment

bottom right (low price, high demand)

why do we need a multiplicative demand model over a linear one

• multiplicative = suitable for regression analysis

• multiplicative models real world better (niche vs mainstream markets)

what is the price elasticity of demand

% change in quantity demanded relative to the given % change in price

price elasticity formula

Price elasticity = (% Change Q) / (% Change P)

expected sign of own-price elasticity

negative (1% increase in price is associated with 2.8% decrease in sales)

How to tell whether demand is elastic

-1 < price elasticity < 0

How to tell whether demand is inelastic

price elasticity < -1

relationship between profits and prices if the demand is elastic

revenue increase = price decrease * quantity increase (opposite effect)

relationship between profits and prices if the demand is inelastic

revenue increase = price increase * quantity decrease (same effect)

expected sign of the cross-price elasticity

positive (because products are substitutes, when price of product B increases, Demand for A increases)

What are the two dimensions of panel data

1. Time dimension (weeks, months)

2. Cross Section (stores, markets, regions)

What are fixed effects?

dummy variables for cross sectional markets that have different demand levels => remove bias from unobserved differences