Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

FEM 21024: Advanced Marketing Models

Dennis Fok
Econometric Institute
October - December, 2023

Unless indicated otherwise, the copyright of these educational materials belongs to the
Erasmus University Rotterdam.
These educational materials may not be reproduced, stored in an automated data file or
made public in any form whatsoever, either electronically, by photocopying, recording or
in any other way without the prior written permission of the EUR.

Contents

II Multivariate models for sales and market shares 2

1 Last week’s material 2

2 Multivariate sales models 3

2.1 Specification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.2 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

3 Models for market shares 6

3.1 Functional forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.2 Restricted forms and identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3.3 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
3.4 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.5 Model selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.6 Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.7 Forecasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.8 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

4 Exam questions 16

Page 1
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Part II
Multivariate models for sales and market
shares
1 Last week’s material
Slide 2 Expectation of the log-normal
Given: Z ∼ N (µ, σ 2 ), calculate E[exp(Z)].

1 (z − µ)2
Z Z  
1
E[exp(Z)] = exp(z)ϕ(z; µ, σ 2 )dz = exp(z) √ exp(− )dz
σ 2π 2 σ2
1 (z − µ)2 − 2σ 2 z
Z  
1
= √ exp(− )dz
σ 2π 2 σ2

Intermezzo:
(z − µ)2 − 2σ 2 z = z 2 − 2(µ + σ 2 )z + µ2
want to rewrite this as (z − a)2 + b = z 2 − 2az + a2 + b
ˆ a = µ + σ2

ˆ b = µ2 − a2 = µ2 − (µ + σ 2 )2 = −2σ 2 (µ + 12 σ 2 )

So, (z − µ)2 − 2σ 2 z = (z − (µ + σ 2 ))2 − 2σ 2 (µ + 12 σ 2 )

Slide 3 Expectation of the log-normal

The integral becomes

1 (z − µ)2 − 2zσ 2
Z  
1
E[exp(Z)] = √ exp(− )dz
σ 2π 2 σ2
1 (z − (µ + σ 2 ))2 − 2σ 2 (µ + 21 σ 2 )
Z  
1
= √ exp(− )dz
σ 2π 2 σ2
1 (z − (µ + σ 2 ))2
Z  
1 1
= √ exp(− ) exp(µ + σ 2 )dz
σ 2π 2 σ2 2
1 (z − (µ + σ 2 ))2
Z  
1 2 1
= exp(µ + σ ) √ exp(− )dz
2 σ 2π 2 σ2
| {z }
=1
1
= exp(µ + σ 2 )
2

Slide 4 Last week

ˆ Any questions?

ˆ Programming exercise?

ˆ Videos ok?

New questions
1: True or false? + Motivation
Consider the model yi = α + βDi + γxi + εi , where Di is a dummy which equals 1 if individual i is
female and 0 otherwise. Next, xi is a vector of individual characteristics. The parameter β represents
the difference in expected value of yi between females and males.

Page 2
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

2: Consider the model

log St = α + β log Pt + γDt + δFt + εt
with St sales in units and standard notation + assumptions.
Suppose that the marginal costs of the product are c.

a) Which price maximizes the expected profit?

b) Does the value of β matter? How?

2 Multivariate sales models

Slide 5 Last week
Last week we discussed a model for sales of a single brand

We considered the general case where we have

ˆ Sales data for one product over time (St )

ˆ Prices for this product (Pt )

ˆ Display indicator (Dt )

(special positioning of product in the store)
ˆ Feature indicator (Ft )
(advertisement in flyer)

We finally came to the model:

St = α · Ptβ · γ Dt · δ Ft · exp(εt )
with εt ∼ N (0, σ 2 ) and α, γ, δ > 0

→ What about competitors?

Slide 6 Multivariate sales models

In a multivariate sales model we consider the sales of all brands in a product category

Why consider multivariate models?

ˆ Brands within a product category are obviously related

ˆ A promotion of one brand will draw sales from competing brands

→ Competition

ˆ Study the market structure

→ Which brands compete, and which do not?
ˆ Measuring cross-elasticities

ˆ Some parameters may be common across brands

ˆ Even if there are no cross-effects, simultaneously considering all equations is more efficient

Page 3
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

2.1 Specification
Slide 7 Allowing for competitive effects
Denote the sales of brand i in week t by Sit , i = 1, . . . , J

Include all marketing mix variables of competing brands:

 
J
β D F β D F
Y Y
Sit = αi Pitβi γiDit δiFit exp(εit ) Pjtji γji jt δjijt = αi  Pjtji γji jt δjijt  exp(εit )
j̸=i j=1
| {z }
competition

and consider all brands i = 1, . . . , J simultaneously

Notation:
ˆ βji denotes the cross-effect of the price of brand j on the sales of brand i, and βii denotes the own
effect.

ˆ Disturbance terms εt = (ε1t , . . . , εJt )′ are probably correlated, we assume εt ∼ N (0, Σ).

Slide 8 Parameter restrictions

The current model
J
β D F
Y
Sit = αi Pjtji γji jt δjijt exp(εit )
j=1

contains many parameters.

Possible parameter restrictions to test:

ˆ βii = β own for all i (=common own-price effects)

ˆ βji = βjcross for all i ̸= j (=common effect of price of j on all other brands)

ˆ etc.

(more on plausible parameter restrictions later today)

2.2 Interpretation
Slide 9 Cross-effects
The cross parameters can be interpreted similarly as the own effects (last week).

ˆ Cross-price elasticity is given by βji (j ̸= i)

∂E[Sit |Pt , Dt , Ft ] Pjt

= βji
∂Pjt E[Sit |Pt , Dt , Ft ]

Usually we expect βji > 0 for j ̸= i (products are substitutes).

ˆ Cross-display effect
E[Sit |Pt , Dt , Djt = 1, Ft ]
= γji
E[Sit |Pt , Dt , Djt = 0, Ft ]
We expect γji < 1.

Page 4
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

2.3 Estimation
Slide 10 Estimation
Again consider the log-linearized form
J
X
log Sit = α̃i + (βji log Pjt + γ̃ji Djt + δ̃ji Fjt ) + εit
j=1

Stack all observations over time: log Si = Xθi +εi where X = (ι, log P1 , . . . , log PJ , D1 , . . . , DJ , F1 , . . . , FJ )

Stack over brands       

log S1 X 0 ... 0 θ1 ε1
 log S2   0 X ... 0  θ2   ε2 
 ..  =  .. ..   ..  +  .. 
      
.. ..
 .  . . . .  .   . 
log SJ 0 0 ... X θJ εJ

This can be summarized as:

log S = Zθ + ε, with ε ∼ N (0, Σ ⊗ IT ) and Z = (IJ ⊗ X)

Slide 11 Estimation
log S = Zθ + ε, with ε ∼ N (0, Σ ⊗ IT )

Estimation of θ and Σ can be done by Feasible Generalized Least Squares [FGLS] (=Seemingly Unrelated
Regressions [SUR])

1. Set an initial guess for Σ̂ (eg. IJ )

2. Assuming Σ = Σ̂: θ̂ = (Z ′ (Σ̂−1 ⊗ IT )Z)−1 (Z ′ (Σ̂−1 ⊗ IT ) log S)

3. ei = log Si − X θ̂i
4. et = (e1t , . . . , eJt )′
5. Σ̂ = T1 t et e′t
P

6. Go to 2 and iterate until convergence

Slide 12 More general specification

In the previous example the same X matrix is used for all equations
→ In this special case (F)GLS is the same as OLS per equation

In general this is not the case

Examples:
ˆ some cross-effects are not included for some brands

ˆ some variables only play a role for some brand

ˆ parameter restrictions across brands (eg. common price elasticities)

In this case construct the Z matrix properly and perform FGLS (̸= OLS)

Page 5
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 13 Model selection and forecasting

ˆ As for the univariate model, we can use all the model specification tests

ˆ Standard errors can be obtained using

d θ̂) = (Z ′ (Σ̂−1 ⊗ IT )Z)−1
Var(

ˆ Given the estimated parameters forecasting is straightforward.

3 Models for market shares

Slide 14 Modeling market shares
Sometimes it may be better to study market shares

ˆ Market shares are directly derived from sales data

Market share of brand i at time t:
I
X
Mi,t = Si,t / Sj,t , i = 1, . . . , I and t = 1, . . . , T
j=1

ˆ Market shares automatically take into account competition

Why not consider sales instead?

ˆ A market share is a relative performance measure

→ Market shares can be compared across different markets or geographical regions
ˆ Sales are sometimes not stationary:

– Category expansion or contraction

– Seasonal/cyclical movements
ˆ (Sometimes only share data is available)

Slide 15 Modeling Market Shares

Market share models allow us to study market & competitive structures across time, place and competi-
tors.

Popular market share models are designed to be:

ˆ Competitive: take into account actions by competitors

ˆ Descriptive: explain why things happen

ˆ Predictive: have forecasting power

ˆ Profit oriented: practical

ˆ Logically consistent: satisfy

– 0 ≤ Mi,t ≤ 1
PI
– i=1 Mi,t = 1

Page 6
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

3.1 Functional forms

Slide 16 Some other candidate models
Linear model:
ˆ Mi,t = µi + βi pi,t + εi,t
ˆ Predictions of market shares can become larger than 1 or smaller than 0.
Multiplicative Model:
ˆ ln Mi,t = µi + β ln pi,t + εi,t
ˆ Positive predictions.
ˆ The price elasticity equals
∂Mi,t pi,t ∂ ln Mi,t
= =β
∂pi,t Mi,t ∂ ln pi,t
ˆ Elasticity always equals β even if Mi,t approaches 1.
Both models violate logical consistency:
ˆ 0 ≤ Mi,t ≤ 1
PI
ˆ i=1 Mi,t = 1

Slide 17 Attractions
Bell et al. (JMR, 1975) introduce an attraction specification.

The only determinant of market share is the attraction (Ai,t ) consumers feel toward each available brand.
Axioms:
PI
1. Ai,t ≥ 0, i=1 Ai,t > 0
2. Ai,t = 0 ⇒ Mi,t = 0
3. Ai,t = Aj,t ⇒ Mi,t = Mj,t
4. The change in Mi,t due to change in Aj,t (j ̸= i) does not depend on j.

Axioms imply that

Ai,t
Mi,t = PI i = 1, . . . , I
j=1 Aj,t
→ Market share equals relative attraction.

Slide 18 Basic Attraction Model

Attraction of brand i at time t
K
Y
Ai,t = exp(µi + εi,t ) xβk,i,t
k
, where
k=1

ˆ µi : constant attraction for brand i

ˆ xk,i,t : marketing instrument k for brand i
ˆ βk : effect of marketing instrument k
ˆ εi,t : unexplained attraction
The unexplained attraction of all brands is assumed to have a multivariate normal distribution with
covariance matrix Σ
εt = (ε1,t , ε2,t , . . . , εI,t )′ ∼ N(0, Σ)
Important assumptions (in this specification)
ˆ Effect of marketing instrument on attraction equal for all brands
ˆ No cross effects of instruments on attractions!

Page 7
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 19 Features of MCI model

Attraction model is also known as MCI (multiplicative competitive interaction) model.
Advantages:
ˆ Model imposes that 0 ≤ Mi,t ≤ 1
PI
ˆ Model imposes that i=1 Mi,t = 1
ˆ Partial effects of changes in marketing instruments make sense (see below)
ˆ The (price) elasticities make sense (see below)

Slide 20 Fully Extended MCI Model

Allow for cross effects and brand specific effects:

Attraction of brand i at time t

I Y
K
β
Y
k,j,i
Ai,t = exp(µi + εi,t ) xk,j,t
j=1 k=1

The attraction for brand i now contains marketing instruments of all I brands and the βk parameters
are different across brands.

Model can be further extended by including

ˆ Lagged marketing instruments
ˆ Lagged market shares

3.2 Restricted forms and identification

Slide 21 Linearization
Take brand I as the benchmark brand
QI QK βk,j,i
Mi,t Ai,t exp(µi + εi,t ) j=1 k=1 xk,j,t
= = QI QK βk,j,I
MI,t AI,t exp(µI + εI,t ) j=1 k=1 xk,j,t
Take logs
I X
X K
ln Mi,t − ln MI,t = (µi − µI ) + (βk,j,i − βk,j,I ) ln xk,j,t + εi,t − εI,t
j=1 k=1
Intuition for identification problem
→ When you know I − 1 shares you know the I-th
Identified parameters:
ˆ µ∗i = µi − µI , i = 1, . . . , I − 1
ˆ βk,j,i
∗
= βk,j,i − βk,j,I , i = 1, . . . , I − 1, k = 1, . . . , K, j = 1, . . . , I
ˆ Σ∗ , the (I − 1) × (I − 1) covariance matrix of ηt = (η1t , . . . , ηI−1,t )′ where ηit = εi,t − εI,t . It holds
that ηt ∼ N(0, Σ∗ )

Slide 22 Restrictions
The fully extended [FE] model contains many parameters. In practice we will want to impose (and test)
various restrictions:
ˆ Restricted competition [RC]
The attraction of brand i only depends on the explanatory variables concerning brand i.
ˆ Restricted effects [RE]
RC + equal parameters across brands
ˆ Restricted covariance matrix [RCM]
No correlations between attraction shocks (diagonal covariance matrix Σ)

Page 8
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 23 Restrictions
Restrictions #Parameters #Restrictions
FE MCI – KI(I − 1) –
RC βk,j,i = 0 for j ̸= i KI KI(I − 2) (vs. FE)
RE βk,i = βk K K(I − 1) (vs. RC)
1 2
FCM – 2 [(I − 1) + (I − 1)] –
1
RCM Σij = 0 if i ̸= j I 2 (I − 1)(I − 2) − 1

RCM Example: I = 4, RCM implies

cov(εt ) = diag(σ12 , . . . , σ42 )

then
σ12 + σ42 σ42 σ42
 
∗
Σ =  σ42 σ2 + σ42
2
σ42 
σ42 σ42 2
σ3 + σ42

or Σ∗ = diag(σ12 , σ22 , σ32 ) + σ42 ιι′ , where ι = (1, 1, 1)′

Slide 24 Attraction specification and reduced form

RC
K
β
Y
k,i
Ai,t = exp(µi + εi,t ) xk,i,t
k=1
K
X K
X
ln Mi,t − ln MI,t = µi − µI + βk,i ln xk,i,t − βk,I ln xk,I,t + εi,t − εI,t
k=1 k=1 ↑
same coef. in all equations (i = 1, . . . , I − 1)

RE
K
Y
Ai,t = exp(µi + εi,t ) xβk,i,t
k

k=1
K
X
ln Mi,t − ln MI,t = µi − µI + βk (ln xk,i,t − ln xk,I,t ) +εi,t − εI,t
k=1
| {z }
eg. log relative price

3.3 Estimation
Slide 25 Estimation
Linearized model can each time be written as:
′ ′
y1,t = w1,t b1 + z1,t a + η1,t
′ ′
y2,t = w2,t b2 + z2,t a + η2,t
.. .. .. .
. = . + . + ..
′
yI−1,t = wI−1,t bI−1 + zI−1,t a + ηI−1,t

where

ˆ yi,t = ln Mi,t − ln MI,t ,

ˆ ηt = (η1,t , . . . , ηI−1,t )′ ∼ N(0, Σ∗ ),

ˆ wi,t : vectors of expl. var. with equation-specific parameter vector bi

ˆ zi,t : vectors of expl. var. with common parameter vector a

Page 9
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 26 Estimation
Define for i = 1, . . . , I − 1
ˆ yi = (yi,1 , . . . , yi,T )′
ˆ Wi = (wi,1 , . . . , wi,T )′ ,
ˆ Zi = (zi,1 , . . . , zi,T )′
ˆ ηi = (ηi,1 , . . . , ηi,T )′
Matrix notation
      
y1 W1 0 ... 0 Z1 b1 η1
 y2   0 W2 ... 0 Z2  ..   η2 

 .. =
 
.. .. .. .. ..

 . +
 
..


 .   . . . . .   bI−1   . 
yI−1 0 0 ... WI−1 ZI−1 a ηI−1
or
y = Xb + η
∗
with η ∼ N(0, (Σ ⊗ IT )), where ⊗ denotes the Kronecker product.

Slide 27 Estimation
In the special cases
ˆ Σ∗ is a diagonal matrix and zi,t = 0 ∀i, t or
ˆ same regressors per equation (wi,t = wj,t and zi,t = 0 ∀i, j, t)
the OLS estimator is efficient PT
→ β̂OLS = (X ′ X)−1 X ′ y and Σ̂∗ = 1
T t=1 η̂t η̂t′ where η̂ = y − X β̂OLS .

In other cases we need to use a (iterated) SUR estimator (FGLS)

b̂SU R = (X ′ (Σ̂∗−1 ⊗ IT )X)−1 X ′ (Σ̂∗−1 ⊗ IT )y,
(iterative SUR = maximum likelihood estimator)

3.4 Dynamics
Slide 28 Dynamics
The fully extended MCI model in this form does not allow for dynamic effects.
→ Marketing efforts in period t do not have an effect on the market shares in period t + 1.

In practice there will of course be dynamic effects.

Examples
ˆ The effect of an promotion can last for more than 1 period
ˆ A small market share in period t often implies a small market share in period t + 1.

Slide 29 Modeling dynamics

Most general specification:
Attraction of brand i at time t
I Y
K I Y
P
β α
Y Y
k,j,i p,j,i
Ai,t = exp(µi + εi,t ) xk,j,t Mj,t−p
j=1 k=1 j=1 p=1

αp,j,i : the effect of the p-periods lagged market share of brand j on the attraction of brand i

The fact that the attraction of brand i can depend on the lagged market shares of all other brands
leads to a very general dynamic specification. (more on the interpretation of dynamic effects in the next
lecture)

Page 10
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 30 Dynamic specifications

As with the other regressors a number of restricted specifications are possible
ˆ Restricted dynamics [RD] (αp,j,i = 0, i ̸= j)
The attraction of brand i at time t only depends on its own lagged market share
→ P I(I − 2) restrictions
ˆ Common dynamics [CD] (αp,j,i = 0, (i ̸= j) & αp,i,i = γp ∀i)
The dynamic effects are equal for all brands
→ P (I − 1) additional restrictions
ˆ No dynamics

Slide 31 Attraction specifications and reduced-form models

RD

I Y
K P
β α
Y Y
k,j,i p,i
Ai,t = exp(µi + εi,t ) xk,j,t Mi,t−p
j=1 k=1 p=1
I X
X K P
X
ln Mi,t − ln MI,t = µ∗i + ∗
βk,j,i ln xk,j,t + (αp,i ln Mi,t−p − αp,I ln MI,t−p ) + ηi,t
j=1 k=1 p=1

CD

K
I Y P
β γ
Y Y
k,j,i p
Ai,t = exp(µi + εi,t ) xk,j,t Mi,t−p
j=1 k=1 p=1

X K
I X P
X
ln Mi,t − ln MI,t = µ∗i + ∗
βk,j,i ln xk,j,t + γp (ln Mi,t−p − ln MI,t−p ) + ηi,t
j=1 k=1 p=1

3.5 Model selection

Slide 32 Model Specification strategy
ˆ start with a fully extended model
ˆ select the appropriate lag order
– tests for residual autocorrelation (LM-test)
– general information criteria: Akaike (AIC) or Schwartz (SIC)
SIC = ln(|Σ c∗ |) + κ ln T , where κ = number of parameters.
T

ˆ test for restricted covariance matrix

ˆ test for restricted dynamics and common dynamics
ˆ test for restricted competition and restricted effects
Impose restrictions if you cannot reject.

Slide 33 Testing restrictions

To test, use standard likelihood ratio tests:
LR = −2(ℓ(b̂0 , Σ̂∗0 ) − ℓ(b̂a , Σ̂∗a )) = T (ln |Σ̂∗0 | − ln |Σ̂∗a |) ∼ χ2 (ν)
asy

where
ˆ ℓ(·): log-likelihood function
ˆ (b̂0 , Σ̂∗0 ), (b̂a , Σ̂∗a ): ML estimates under H0 or Ha
ˆ ν: number of parameter restrictions

Page 11
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

3.6 Interpretation
Slide 34 Interpreting parameter estimates
Interpretation of parameters is often difficult:
ˆ Only a reduced form model can be estimated.

ˆ Often there is a very large number of parameters.

ˆ Estimated parameters in the fully specified model only give the effect on relative market shares.

ˆ The presence of dynamic effects also make interpretation difficult.

Just “looking” at estimated parameters does not lead to much insight in models with many brands.

Slide 35 Interpretation
Simplified model:
I
β
Y
Ai,t = exp(µi + εi,t ) pj,tj,i
j=1

Partial effect of price on market share (omitting the conditional expectations in the notation)
I
∂Mi,t X ∂Mi,t ∂Ak,t
=
∂pi,t ∂Ak,t ∂pi,t
k=1

where
∂Ak,t βi,k
= Ak,t
∂pi,t pi,t
∂Mi,t 1−Mi,t ∂Mi,t −M
= PI and = PI i,t for (k ̸= i) such that
∂Ai,t j=1 Aj,t
∂Ak,t j=1 Aj,t

I
∂Mi,t Mi,t X
= (βi,i − βi,k Mk,t )
∂pi,t pi,t
k=1

Slide 36 Price Elasticity

Price elasticity:
I
∂Mi,t pi,t X
= βi,i − βi,k Mk,t
∂pi,t Mi,t
k=1

This seems reasonable because when your market share approaches one the elasticity goes to zero.
Under Restricted Competition
ˆ βi,k = 0, i ̸= k
∂Mi,t pi,t
ˆ ∂pi,t Mi,t = βi,i (1 − Mi,t )

ˆ Elasticity does not depend on other market shares or attractions.

⇒ is similar to independence of irrelevant alternatives in logit models
Under Restricted Effects
ˆ βi,k = 0, i ̸= k and βi,i = β
∂Mi,t pi,t
ˆ ∂pi,t Mi,t = β(1 − Mi,t )

ˆ Elasticities are the same for each brand when they have the same market share.

Page 12
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

3.7 Forecasting
Slide 37 Forecasting relative market shares
Denote relative market shares (Mi,t /MI,t , i = 1, . . . , I − 1) by mi,t .
The reduced form of the MCI model is
ln mi,t = ln Mi,t − ln MI,t
I X
X K P X
X I
= µ∗i + ∗
βk,j,i ln xk,j,t + ∗
αp,j,i ln Mj,t−p + ηi,t
j=1 k=1 p=1 j=1

= λi,t + ηi,t

Recall that (η1,t , . . . , ηI−1,t )′ ∼ N (0, Σ∗ ) so that mi,t ∼ LogN(λi,t , Σ∗i,i ).

Therefore
b i,t = E(mi,t ) = exp(λi,t + 1/2Σ∗i,i ), i = 1, . . . , I − 1
m

Slide 38 Forecasting Market Shares

One of the purposes of the model is to be able to make forecasts for market shares.

A forecast of the market share of brand i = 1, . . . , I for period t + 1 is in fact

E[Mi,t+1 |Xt+1 , Mt ]

with
ˆ Xt+1 : information on explanatory variables up to period t + 1
ˆ Mt : information on market shares up to period t

The estimated coefficients in a MCI model refer to ln mi,t . Forecasts for the market shares are therefore
not easy to make.

Slide 39 Forecasting Market Shares

To be able to forecast market shares we need an expression relating Mi,t to mi,t
1
MI,t = PI−1
1+ j=1 mj,t
mi,t
Mi,t = MI,t mi,t = PI−1
1+ j=1 mj,t

(Unbiased) forecasts of Mi,t+1 are therefore equal to

" #
m i,t+1
M
ci,t+1 = E PI−1 Xt+1 , Mt , i = 1, . . . , I − 1
1 + j=1 mj,t+1
" #
1
MI,t+1 = E
c PI−1 Xt+1 , Mt
1 + j=1 mj,t+1

Slide 40 Forecasting Market Shares

When forecasted (expected) values for mi,t are inserted in the previous equations we do not obtain
unbiased forecasts.

The main reason is that E(X/Y ) ̸= E(X)/E(Y ).

In fact it turns out that it is not possible to get an expression for an unbiased forecast. We therefore
will have to use simulation to calculate the forecasts.

Page 13
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 41 Forecasts with Simulation

Consider the following general case:

ˆ The variable of interest is: Y

ˆ The model specifies a density function for a transformation of Y : g(Y ) ≡ Z (Y = g −1 (Z))

→ denote this density by f (z)
ˆ Note again that E[Y ] ̸= g −1 (E[Z])

To obtain an (unbiased) forecast of Y :

1. Simulate z̃i from f (z) for i = 1, . . . , N
2. Calculate ỹi = g −1 (z̃i )

3. E[Y ] ≈ N1 i ỹi
P

By the weak law of large numbers:

1 X
plim ỹi = E[Y ]
N →∞ N i

Slide 42 Simulation schema for market shares

Unbiased forecasts are obtained by using the following simulation schema.

1. Draw realizations of the disturbances ηi,t

i = 1, . . . , I − 1 from the estimated distribution.
2. Given these disturbance realizations, calculate the relative market shares using the estimated model.
3. Transform the relative market share realizations to market share realizations.

4. Repeat 1,2 and 3 for a large number of times

5. An unbiased forecast of the market share is given by the average market share over the iterations.

Slide 43 Simulation schema

Drawing of disturbance realization:

ηt = (η1,t , . . . , ηI−1,t )′ ∼ N (0, Σ

b ∗)

Decompose Σb ∗ using the Choleski decomposition. This means: find a (lower-triangular) matrix P such
′ b∗
that P P = Σ

Denote by Z a vector of (I − 1) independent drawings from a normal distribution (easy).

Note that
Cov(Z) = I
Cov(P Z) = P IP ′ = P P ′ = Σ
b∗

b ∗)
P Z is therefore a draw from N (0, Σ

Page 14
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

3.8 Illustration
Slide 44 Interpreting an empirical model (Detergent example)
Available data
Market share Mi,t
Relative price Pi,t
Relative distribution Di,t
Promotion indicator P ri,t
132 weekly observations of a 5 brand detergent market
Table 1: Data hara teristi s 5-brand detergent data
Brand
1 2 3 4 5
Average market shares (%) 24.49 22.22 22.62 8.22 22.45
Average relative pri e 1.12 1.09 1.10 1.09 1.11
Average relative distribution 0.85 0.91 0.93 0.78 0.94
Fra tion of promotions 0.27 0.21 0.39 0.34 0.10
 Fra tion of 132 weeks in whi h the brand is on promotion.

Slide 45 Empirical model

Selected model includes:
ˆ 1 period lagged market shares

ˆ own price, distribution and promotion effects

ˆ all cross effects

ˆ restricted covariance matrix for attraction disturbances

Page 15
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Slide 46 Estimation results

Table 2: Estimated parameter values in an empiri al MCI model for 5 detergent
brands (standard errors in parentheses)
M1;t M2;t M3;t M4;t

M5;t M5;t M5;t M5;t

Inter ept 1.308 (0.69) 2.136 (0.898) 1.121 (0.941) 1.387 (0.894)
P1;t -7.828 (1.139) -2.220 (1.481) -0.274 (1.554) -2.647 (1.476)
P2;t 0.054 (0.404) -3.345 (0.526) 0.003 (0.552) -0.254 (0.524)
P3;t -0.186 (0.555) -0.072 (0.722) -5.241 (0.758) 0.207 (0.72)
P4;t -0.196 (0.796) 2.487 (1.035) -0.465 (1.085) -6.682 (1.031)
P5;t 7.419 (1.329) 3.987 (1.729) 5.243 (1.813) 8.887 (1.722)
D1;t 0.081 (0.288) 0.157 (0.374) -0.080 (0.392) -0.428 (0.373)
D2;t 0.014 (0.162) 1.124 (0.211) -0.367 (0.221) 0.104 (0.21)
D3;t 0.330 (0.363) -0.068 (0.472) 1.466 (0.495) -0.584 (0.47)
D4;t 0.027 (0.113) 0.137 (0.147) -0.084 (0.154) 1.015 (0.146)
D5;t -0.224 (0.402) -1.535 (0.523) -1.042 (0.548) 0.174 (0.521)
exp(P r1 ;t ) 0.070 (0.039) 0.080 (0.051) 0.047 (0.054) -0.064 (0.051)
exp(P r2 ;t ) -0.012 (0.051) 0.149 (0.067) -0.047 (0.07) 0.020 (0.066)
exp(P r3 ;t ) 0.037 (0.038) -0.004 (0.05) 0.094 (0.052) 0.001 (0.05)
exp(P r4 ;t ) -0.076 (0.044) 0.010 (0.058) -0.091 (0.06) -0.007 (0.057)
exp(P r5 ;t ) -0.012 (0.063) -0.225 (0.083) -0.097 (0.087) -0.014 (0.082)
M1 ;t 1 0.394 (0.113) 0.118 (0.147) 0.143 (0.155) 0.212 (0.147)
M2 ;t 1 0.127 (0.107) 0.558 (0.139) 0.193 (0.146) 0.276 (0.139)
M3 ;t 1 0.180 (0.141) 0.567 (0.183) 0.384 (0.192) 0.621 (0.182)
M4 ;t 1 0.129 (0.061) 0.084 (0.079) 0.091 (0.083) 0.400 (0.079)
M5 1 -0.192 (0.138) 0.081 (0.179) -0.103 (0.188) -0.229 (0.179)
02 2 2 1 0 1
;t

^1 + ^5 ^5 ^52 ^52 0:024 0:015 0:015 0:015

B
B ^52 ^22 + ^52 ^52 ^52
C
C B
B 0 015 0 040 0 015 0 015
C
C
^  = B C =B C
: : : :
B
 ^52 ^52 ^32 + ^52 ^52 A 0:015 0:015 0:044 0:015C
C B A
^52 ^52 ^52 ^42 + ^52 0:015 0:015 0:015 0:039
Slide 47 New programming assignment
Choose one of the product categories from last week and
ˆ Specify a market share model (ignore dynamics for now)
1
ˆ Perform model selection by testing a number of restricted specifications (do not use the last half
year of data to do this)

ˆ Forecast the last half year (given prices etc) and compare the forecasting performance of a naive
forecast to the simulation-based forecast

4 Exam questions
Slide 48 Question 1
True or false? Give a short motivation:

Page 16
 Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

One does not need simulation to calculate unbiased forecasts of relative market shares in the market
share attraction model.

Slide 49 Question 2
Consider the following market share attraction model
α
Ait = exp(αi + εit ) exp(Dit )γi exp(Fit )δ Ij=1 Pjtji ,
Q
for i = 1, . . . , I , t = 1, . . . , T and where Dit and Fit are 0/1 indicators for display and feature of brand i at time
 2 
σ1
2
 σ2 
t, Pit is the price of brand i at time t, and finally (ε1t , . . . , εIt ) ∼ N (0, Σ) with Σ =  .
 
..
 . 
σI2
4a) Give one reason why one would consider to model market shares instead of sales.
4b) Explain why the exponential transformations of display and feature have to be used.
4c) For each model component (display effect, feature effect, price effect, and the error terms) give the name of
the specification or the restrictions imposed.
4d) Give the expression for the reduced form model, that is, the model that can be used to estimate the parameters.
4e) Specify which (functions) of the parameters are identified.

Slide 50 Question 3
To explain the development of the market shares of 3 brands, a researcher specifies a market share attraction
model. He specifies a reduced form model for the log relative market share of brand 1 and brand 2 relative to
brand 3. As explanatory variables he uses an intercept and the log prices of all three brands for both equations.
Denote the market share of brand i at time t by Mit and the price of brand i at time t by Pit .
a) Give the complete attraction specification that corresponds to this reduced form model. Use (Greek) letters
for the parameters.
The researcher obtains the following parameter estimates for the reduced form model

log M1t − log M3t = 1.05 − 2.45 log P1t + 0.75 log P2t + 0.25 log P3t + η1t
log M2t − log M3t = 0.75 + 0.45 log P1t − 1.95 log P2t + 0.89 log P3t + η2t

b) Suppose that in general the prices equal e2.71. Which brand will have the largest market share, and which
brand will have the smallest?
c) Give an interpretation to each of the three price effects in the equation for log M1t − log M3t .
d) Another researcher claims that it is better to take brand 2 as the base brand. Give the reduced form model
and the estimated parameter values for this case. Can it be true that this reduced form model is better? Motivate
your answer.

Slide 51 Literature
Background info:

ˆ Klapper and Herwartz (2000)

ˆ Cooper and Nakanishi (1988)

ˆ Kumar (1994)

→ Study Fok et al. (2002)

References
Bell, David E., Ralph L. Keeney, and John D. C. Little (1975), “A Market Share Theorem”,
Journal of Marketing Research, 12, 136–141.
Cooper, Lee G. and M. Nakanishi (1988), Market Share Analysis: Evaluating Competitive Marketing Effectiveness,
Kluwer Academic Publishers, Boston.
Fok, Dennis, Philip Hans Franses, and Richard Paap (2002), “Econometric Analysis of the Market Share
Attraction Model”, in P. H. Franses and A. L. Montgomery (eds.), Advances in Econometrics, vol. 16,
chap. 10, JAI Press, pp. 223–256.

Page 17
Advanced Marketing Models 2023-2024 ©Erasmus School of Economics

Klapper, D. and H. Herwartz (2000), “Forecasting Market Share Using Predicited Values of Competitor
Behavior: Further Empirical Results”, International Journal of Forecasting, 16, 399–421.
Kumar, V. (1994), “Forecasting Performance of Market Share Models: An Assessment, Additional In-
sights, and Guidelines”, International Journal of Forecasting, 10, 295–312.

Page 18