lecture 6.1 Flashcards
multiplication model: advertising
formal concept
Q = k *outofhomeadv^a * TVadv^b
a = sales effect of OoH advertising
b = Sales effect of TV advertising
k= scaling perameter
Implications of nonlinearity
Absolute effect:
dQ/DtvAdv = k * OoHAdvâ * b TVAdv^b-1 = b/TVAdvQ
Effect depends on the level of the advertising that determines the sales level
Relative effect:
(dQ/Q)/dTVAdv/TVAdv = dQ/dTVAdv * TVAdv/Q = b
b = elasticity corresponds to the exponent in the multiplication model
in the multiplicative model we can see the elasticity right away because the elasticity is the exponent in the multiplicative model
Modified exponential model
There is a saturation limit. upper bound for sales no matter what i do (Qmax)
Sometimes sales might even go down because i put too much effort into advertising.
Q = Qmax(1-e^-a1x)e^u
Ex = (Qmax - Q/Q) a1x
Log-reciprocal model
S shaped model
Marketing activities are not effective until they reach a level, then they increase heavily and at some point decrease again
Q = e^(a0-a1*(1/x)+u)
Ex=a1/x
Logistic model
S shaped model
Marketing activities are not effective until they reach a level, then they increase heavily and at some point decrease again
Q = (Qmax/1+e^(ao+a1*x)) *eu
Ex = a1x((Qmax - Q)/Qmax)
logistic model if i have a 1 0 coded variable as a dependent variable and the outcome variable can only be predicted in the for of a prbobability
Logistic model what is it all about
starding point:
outcome variable can only be predicted in the form of probability (e.g. purchase probability, adoption probability)
Event to be explained/predicted is binary (0 or 1)
Response (yes/no to an offer in direct marketing
Inclusion of a product in the assortment of a retailer
First use of a new product (e.g. insulin pump) and buyer (adoption)
Implications:
Robust specification required i.e. predictions lie in the interval 0 to 1
Robustnessi mplies nonlinear regression function
Special assumptions regarding the distribution of the error term required (logistic)
Adequate estimation methods require (e.g. maximum likelihood)
Adequate fit dimensions required (classic R^2 cannot be used)
Logistic regression, why is it useful for customer success
1) able to extract interpretable factors
2) can calculate risk of churn scores (retain or churn)
Logistic model example mailings
you are sending mailings with an offer. can you predict the response rate based on individual variables? can you optimize the price and advertising?
A customer resonds (y= 1) or not (y=0)
Your decision variables are:
The price to charge per unit of product (20, 25,30)
whether to send a reminder email 8yes or no)
whether to include a coupon (yes or no)
How much to spend on salesperson interaction($)
A pilot mailing was performed varying these variables according to a randomized trial. We obtained data of whether the individual responded, and have information on prices, email, coupon and salesforce expenditure
THe pilot was conducted with 200 respondents
Dynamic effects: sales response models (dynamic response model)
Qt = b0 + b1X1 + lamda* Qt-1
Qt = sales depend on time t
beta1X1 = current effect
lamda*Qt-1 = carry over effect
Customers, retailers and competitors might need a certain time to react to marketing activities (delayed response effects: execution delays, noting delays, purchase delays, recording delays)
especially the impact of advertising is considered to be a dynamic process
Customers, retailers and competitors might even react to marketing activities in advance (if they anticipate an expected action)
Dynamic effects: customer holdout model
Q slowly increases, reaches level then decreases over time when sales campaign is over
Delayed response model (customer holdout model)
Alternative it only slowly decreases, this is “hysteresis” effect