Upload
rachel-bell
View
213
Download
0
Tags:
Embed Size (px)
Citation preview
MILAN| 23 June 2011
Success factors of models used for supporting sales‐related allocation decisions-
Sönke AlbersProfessor of Marketing and InnovationKühne Logistics University, Hamburg
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
AGENDA
2
Success factors for empirical estimation of aggregate sales response function
Success factors for optimizing allocation of sales effort
Conclusion
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Success Factors for Empirical estimation of Aggregate Sales Response Function
3
1. Functional form must be appropriate and matters!2. Functional form should exhibit decreasing elasticity3. Long-term effects should be modeled with the help of
stock models (not Koyck)4. Separation of effects of first contact and repeated
contacts5. Heterogeneity across customers should be taken
into account6. Trends and seasonality should be taken into account
I data are from a panel
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of The Aggregate Response Function
4
Observation: Marketing & sales instruments (except for price) exhibit diminishing marginal returns at some point, otherwise marketing & sales expenditures would not be optimizable!
Proposition 1: Linear relationships are therefore inappropriate for aggregate sales response models.
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
5
Example for an assumed linear relationship from a top journal:
Source: Natalie Mizik and Robert Jacobson: Are Physicians "Easy Marks"? Quantifying the Effects of Detailing and Sampling on New Prescriptions Management Science Vol. 50 (2004), No. 12, 1704-1715
We employ the following dynamic fixed-effects distributed lag regression model to assess the effect of detailing and sampling on new prescriptions:
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
6
Proposition 2a: The functional form is very important when diminishing marginal returns exist!
Name Function Elasticity
Constant Elasticity
Semi-logarithmic
Diminishing Elasticity
Modified Exponential
Log-Reciprocal
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
7
0,00
1.000,00
2.000,00
3.000,00
4.000,00
5.000,00
6.000,00
7.000,00
8.000,00
9.000,00
0 5 10 15 20 25 30
Sales
Linear
Quadratic
Constant Elasticity
Diminishing Elasticity
Modified Exponential
Log-Reciprocal
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
8
Proposition 2b: The functional form for a relationship with diminishing marginal returns is important because it leads to quite different optimal solutions!
Functional Form Optimal No. Calls
Quadratic 17.7
Constant Elasticity 77.8
Diminishing Elasticity
18.6
Modified Exponential 14.1
Log-reciprocal 24.3
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
9
Observation: Very often, researchers prefer to work with nonlinear functions that are linearizable.
Proposition 3a: Adding quadratic terms does not help unless the interval with the maximal or minimal outcome is supported by a sufficient number of observations.
Proposition 3b: Estimation of linearized models (e.g., taking logs) comes at the cost of creating a bias in the error terms (implicitly weighting lower versus higher values).
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Taking Logs in linear estimation is different from nonlinear estimation (Proposition 3b)
10
0,00
1.000,00
2.000,00
3.000,00
4.000,00
5.000,00
6.000,00
7.000,00
8.000,00
9.000,00
10.000,00
0 5 10 15 20 25 30
Sales
Constant Elasticity
Log-log
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Functional Form of the Aggregate Response Function
11
Proposition 3b: Estimating constant elasticity response functions with a linearized log-log function can lead to dramatically different results for the optimal call level compared to a nonlinear estimation
Functional Form Optimal No. Calls
Log-Log 265.8
Constant Elasticity 77.8
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
AGENDA
12
Success factors for empirical estimation of aggregate sales response function
Success factors for optimizing allocation of sales effort
Conclusion
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Success Factors for optimal allocation of sales effort
13
1. Estimation method should have maximized information value rather than goodness-of-fit
2. Results should have face validity but also some surprising aspects
3. Any model should be offered in EXCEL4. Rather than providing numerical optimization
understandable heuristics are better accepted
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Informational value versus goodness-of-fit
14
Proposition 4: Response models may be good for reproducing sales utilizing the response function, but may nevertheless be problematic if they lead to optimization errors (e.g., in the paper by Proppe and Albers (2009) where one wrong estimate affects the entire allocation task, or in case of uncertainty when non-significant variables are set to be zero)
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
15
Research questions
RESEARCH QUESTION CAN BEST BE SOLVED BY SIMULATION STUDY
Simulation
Dennis Proppe and Sönke Albers: Choosing Response Models for Budget Allocation in Heterogeneous and Dynamic Markets: Why Simple Sometimes Does Better, Marketing Science Institute Special Report 09-202, April 2009
Key question: Which econometric methods deliver the most reliable estimation results that can be used for a successful budget allocation task?
Which data properties are especially influential for good estimation when it comes to optimal budget allocation?
Under what circumstances may simple allocation heuristics perform better than the allocation based on econometric estimation?
When analyzing real data, the true model remains unknown.
Results of econometric estimation procedures can only be evaluated by goodness-of-fit-statistics and not by their closeness to the true relationship.
In a simulation study the data is generated by a true model which is specified by the researcher.
Thus, the real model parameters are known and the estimation result can be compared to the true relationship.
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
16
DATA QUALITY HAS A SUBSTANTIAL INFLUENCE ON THE OPTIMALITY OF THE BUDGET ALLOCATION
Dennis Proppe and Sönke Albers: Choosing Response Models for Budget Allocation in Heterogeneous and Dynamic Markets: Why Simple Sometimes Does Better, Marketing Science Institute Special Report 09-202, April 2009
Best case: Small error, many observations, many response units
Worst case: Large error, few observations, few response units
Scenario Measure Simple least squares
regression
Fixed effects Hierarchical Bayes
Maximum simulated likelihood
Empirical Bayes
Equal allocation
Proportion-al-to-sales
Mean 91.98% 94.21% 94.92% 96.24% 95.85% 80.86% 87.90%
Standard deviation
(7.22%) (2.55%) (4.81%) (3.39%) (3.92%) (5.83%) (3.10%)
5* 4* 3 1 2* 7 6*
Best Case, N=30
Rank (*=significant difference to next lower
rank)
Scenario Measure Simple least squares
regression
Fixed effects Hierarchical Bayes
Maximum simulated likelihood
Empirical Bayes
Equal allocation
Proportion-al-to-sales
Mean 6.03% 71.61% 55.37% 68.27% 26.78% 80.86% 87.28%
Standard deviation
(22.92%) (29.33%) (38.16%) (30.03%) (34.95%) (5.83%) (3.22%)
7 3* 5* 4* 6* 2 1*
Worst Case, N=30
Rank (*=significant difference to next lower
rank)
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
17
GOOD ECONOMETRIC TECHNIQUES BENEFIT FROM HETEROGENEITY AND a LARGE No. of OBSERVATIONS
Dennis Proppe and Sönke Albers: Choosing Response Models for Budget Allocation in Heterogeneous and Dynamic Markets: Why Simple Sometimes Does Better, Marketing Science Institute Special Report 09-202, April 2009
Heterogeneous parameters, large number of observationsScenario Measure Simple least
squares regression
Fixed effects Hierarchical Bayes
Maximum simulated likelihood
Empirical Bayes
Equal allocation
Proportion-al-to-sales
Mean 76.40% 93.33% 88.56% 95.35% 90.97% 75.18% 84.86%
Standard deviation
(28.00%) (3.10%) (14.73%) (3.79%) (10.48%) (1.00%) (0.41%)
6 2 4 1* 3* 7 5Rank (*=significant
difference to next lower rank)
Hetero-geneous,
many obser-vations
Homogeneous parameters, small number of observationsScenario Measure Simple least
squares regression
Fixed effects Hierarchical Bayes
Maximum simulated likelihood
Empirical Bayes
Equal allocation
Proportion-al-to-sales
Mean 30.69% 76.28% 62.52% 86.16% 45.07% 95.28% 95.52%
Standard deviation
(35.94%) (31.93%) (39.87%) (21.30%) (40.27%) (0.33%) (0.37%)
7 4* 5* 3* 6* 2* 1*
Homo-geneous,
few observ-ations
Rank (*=significant difference to next lower
rank)
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Implementation success depends on ease-of-use
18
Proposition 5: Managers like to make use of instruments that they can master. EXCEL is such an instrument, and thus decision support should be made available in EXCEL tools.
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Implementation success depends on ease-of-use
19
“As Albers (IJRM 2000) notes, the use of marketing models in actual practice is becoming less of an exception and more of a rule because of spreadsheet software. It is our hope that the ease with which the BG/NBD model can be implemented in a familiar modeling environment will encourage more firms to take better advantage of the information already contained in their customer transaction databases. Furthermore, as key personnel become comfortable with this type of model, we can expect to see growing demand for more complete (and complex) models—and more willingness to commit resources to them.”
Peter S. Fader, Bruce G. S. Hardie, and Ka Lok Lee: “Counting Your Customers” the Easy Way: An Alternative to the Pareto/NBD Model, Marketing Science, Vol. 24, No. 2, Spring 2005, pp. 275–284
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Implementation success depends on ease-of-use
20
Proposition 6: Managers want to understand why certain decisions are recommended. Optimization models for determining key marketing budgets will only be applied if the solution is provided in terms of understandable heuristics.
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Example for easy-to-understand Heuristic
21
Sales effort should allocated proportionally to
• Past sales• Contribution margin• Elasticity of Sales with respect to changes of sales
effort
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
AGENDA
22
Success factors for empirical estimation of aggregate sales response function
Success factors for optimizing allocation of sales effort
Conclusion
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
Conclusion
23
1. The functional form matters: Aggregate response functions should exhibit diminishing marginal returns and decreasing elasticities
2. Goodness-of-fit is not everything. Sometimes it is better to know whether an estimation provides values that lead to correct optimization, as is the case with allocation.
3. Instead of very complicated optimization approaches, we need heuristics that are understandable and easy to implement in Excel for managers.
Sönke Albers: Success factors of models used for supporting sales‐related allocation decisions
24
Thank you very much for your attention!