Background:
Faced with recent supply and cost pressures, orange juice manufacturers have been challenged to maintain a positive bottom line. To compensate for the higher cost of oranges, in 2007, every OJ brand raised its base & promoted price by more than 10%. Given the already high price sensitivity among OJ consumers, it was expected that most OJ brands would take a loss in volume.
Considering the fast moving nature of the OJ category, it is presumed that consumers will get adapted to the 10% increase in price and the volume sales will exhibit a similar pattern after a few weeks.
Business Question and Research Objectives:
Tropicana is interested in understanding the time required by consumers to get adapted to the price increase. In technical terms, the pattern of volume sales before the price increase is similar to the pattern of volume sales post price increase. We can drill down to do the analysis at the state level or market level or the client desired region level. The methodology remains the same for all levels.
Recommended Methodology:
The scope of the project can be narrowed down to understand just the average time taken for consumers to adapt to the price increase.
The analysis will be split into two parts. Firstly, it will utilize a regression model, controlling for own and competitive price/promotion/distribution. This will be the fixed effects model similar to any price and promotion or marketing mix model. Marketing inputs will be added to the data based on the client input.
We can either model on the individual pack sizes or on the total Tropicana as a whole. In either approach, we will need approximately 5 years of data to observe the volume sales pattern. With 5 years of data, we assume that there will be pattern related to both the old and the new price.
Also, a strong brand will have a stable elasticity i.e. beyond a certain time point; the brand volume will not be affected much due to the increase / decrease in price. We know that the entire OJ category has experienced an increase in price which means the brand elasticity would be unstable for certain time period and then considering the brand equity, it should come back to normal over time.
We will follow the same standard IRI process (Price-Promo / Market Mix) for building a regression model on the 5 years of data controlling for advertising and coupons. This will give us the estimates for all the variables at the TUS level. The base price elasticity in this case will take into account both the price increase and the old price. So, intuitionally, with 5 years of data, we assume the brand to be inelastic
The next step will be to understand the model fit. We will run the simulation program for the obtained fixed effects model and use the aggregate week level file (the output of simulation which is volume due to) as a base for further calculations. The structure of this file is attached for clarifications.
In the fixed effects model, our endeavor is to try and capture as much of variability in the data as possible. In practice, it is possible that the target / competition would have run some special events on a week to week basis plus some extraneous information which is captured by the error term. It is practically possible that the unexplained variation captured by the error term helps us draw reasonable insights due to the price increase. It will isolate the effect due to any other control variables viz. promotion / advertisement / competitor activities etc.
Residual Way of Modeling:
Therefore, as a second stage of analysis, we will calculate the residuals from the first model; and model for residuals with base price as the only independent variable. This will help us understand the impact of base price on the brand’s volume more accurately. From the week level aggregate file, we will calculate the residuals (See Appendix). This will be done for the entire 5 years of data and then residuals will be used as the dependant variable.
Our objective is to understand the impact of price increase on the brand elasticity and how much time do the consumers take on an average to get used to the price increase. The elasticities we have are short term elasticity’s. For capturing the impact of price increase on a long term basis, we will need to estimate the long term elasticities.
For calculating long term elasticities, we will create a moving window for the entire 5 years of data. This concept is based on one of the research papers published in the marketing science journals written by Mela, Lehmann and Gupta titled “the long term impact of promotion and advertising on consumer brand choice”.
E.g. we have 260 weeks and we will roll up the data for each 12 week period. Suppose we have data from week 1 to week 52. We create data sets from W1-W12, W5-W16, -------, W41-W52. Empirical research in marketing science says that this window is called the moving window. Instead of 12 weeks which is practiced for understanding the long term Ad effectiveness, forecasting etc we can choose some other number but for consistency sake, we can go ahead with the number 12.
For implementing the above logic, we will need to aggregate the residuals and the entire movement data at the quad week level (as explained above). We will do it only for the target variables and in our case it will be only the base price. Hence, for 5 years of data, we will have the movement data at the quad week level i.e. 66 data points at the quad week level instead of 260 at the store week level. In our case, we will sum up the residuals at the quad week level and average the base price at the quad week level.
The next step after this data preparation will be to model for residuals at the quad week level with base price as the independent. The model form will now be
Residuals = α + Base Price
The model procedure will be a standard Proc Mixed in SAS at the quad week level which will give us 66 estimates. Once we get these estimates, we will calculate the elasticities and use them appropriately.
It is likely that the plot will have high kinks (both upward and downward). In that case, it will be difficult to read a pattern. If this exists, we will use a survival analysis to fit the decay curve and smoothen out the kinks for easy interpretation.
Interpretation:
The last step in the analysis will be to plot all these 66 base price elasticities across weeks along with the base price and observe the pattern change. We expect to see a stable pattern for a long time frame due to the old price, then a sudden increase / decrease in the pattern for some time period and lastly a stable pattern which indicates that there is a rebound in the volume sales.
As a diagnostic, we can also compare the new elasticities with our TUS fixed effect elasticity value. This will help us explain the impact of change observed over 5 years of data. When we say TUS fixed effect elasticity value over the 5 years of data, our assumption is that that the consumers have already got adapted to the new price; hence it should not be quite a high number based on the OJ category standards.
If we do not observe a significant change in the pattern of elasticities, we will get a direction that the overall decline may not be just due to the base price but it can be due to the decline in promotional activity, increase in the competition promotions etc.
Appendix:
The scope of this project could be scaled to as high a level as possible. Empirical research in marketing science considers many variables to find a solution to this problem. The variables considered can be location of the store, population of the state / region, average income of people in that state / region etc. If we model for all of these variables, we will arrive at a solution which will classify the important variables that lead for consumer adoption. This will require us to get to the consumer level data i.e. panel data. If we try to scope it this way, it can grow as big as we want it.
Residuals = Actual Volume – Volume due to promotions.
Expanded Scope:
This approach is a special case in Marketing Science and involves a lot of criteria. Ideally, it would be useful to have fast moving products to be able to see a pattern, a significant price increase which will help the client fix the best price for their product depending on the average length for rebound etc. With this analysis, the client will not only be able to know the average length for rebound but depending on that length, they would be able to decide on the % increase / % decrease they would want to fix.
If the clients are interested in knowing the effect of Increase / Decrease at a granular level viz. Region, CORP RMA etc, we can very well use the Randomization approach using PROC MIXED in SAS.
The scope of this approach could be used to help the client understand the average length taken for the advertisement effect / campaign to decay. This will be very helpful to the clients since they could utilize their advertisement expenses better.
Faced with recent supply and cost pressures, orange juice manufacturers have been challenged to maintain a positive bottom line. To compensate for the higher cost of oranges, in 2007, every OJ brand raised its base & promoted price by more than 10%. Given the already high price sensitivity among OJ consumers, it was expected that most OJ brands would take a loss in volume.
Considering the fast moving nature of the OJ category, it is presumed that consumers will get adapted to the 10% increase in price and the volume sales will exhibit a similar pattern after a few weeks.
Business Question and Research Objectives:
Tropicana is interested in understanding the time required by consumers to get adapted to the price increase. In technical terms, the pattern of volume sales before the price increase is similar to the pattern of volume sales post price increase. We can drill down to do the analysis at the state level or market level or the client desired region level. The methodology remains the same for all levels.
Recommended Methodology:
The scope of the project can be narrowed down to understand just the average time taken for consumers to adapt to the price increase.
The analysis will be split into two parts. Firstly, it will utilize a regression model, controlling for own and competitive price/promotion/distribution. This will be the fixed effects model similar to any price and promotion or marketing mix model. Marketing inputs will be added to the data based on the client input.
We can either model on the individual pack sizes or on the total Tropicana as a whole. In either approach, we will need approximately 5 years of data to observe the volume sales pattern. With 5 years of data, we assume that there will be pattern related to both the old and the new price.
Also, a strong brand will have a stable elasticity i.e. beyond a certain time point; the brand volume will not be affected much due to the increase / decrease in price. We know that the entire OJ category has experienced an increase in price which means the brand elasticity would be unstable for certain time period and then considering the brand equity, it should come back to normal over time.
We will follow the same standard IRI process (Price-Promo / Market Mix) for building a regression model on the 5 years of data controlling for advertising and coupons. This will give us the estimates for all the variables at the TUS level. The base price elasticity in this case will take into account both the price increase and the old price. So, intuitionally, with 5 years of data, we assume the brand to be inelastic
The next step will be to understand the model fit. We will run the simulation program for the obtained fixed effects model and use the aggregate week level file (the output of simulation which is volume due to) as a base for further calculations. The structure of this file is attached for clarifications.
In the fixed effects model, our endeavor is to try and capture as much of variability in the data as possible. In practice, it is possible that the target / competition would have run some special events on a week to week basis plus some extraneous information which is captured by the error term. It is practically possible that the unexplained variation captured by the error term helps us draw reasonable insights due to the price increase. It will isolate the effect due to any other control variables viz. promotion / advertisement / competitor activities etc.
Residual Way of Modeling:
Therefore, as a second stage of analysis, we will calculate the residuals from the first model; and model for residuals with base price as the only independent variable. This will help us understand the impact of base price on the brand’s volume more accurately. From the week level aggregate file, we will calculate the residuals (See Appendix). This will be done for the entire 5 years of data and then residuals will be used as the dependant variable.
Our objective is to understand the impact of price increase on the brand elasticity and how much time do the consumers take on an average to get used to the price increase. The elasticities we have are short term elasticity’s. For capturing the impact of price increase on a long term basis, we will need to estimate the long term elasticities.
For calculating long term elasticities, we will create a moving window for the entire 5 years of data. This concept is based on one of the research papers published in the marketing science journals written by Mela, Lehmann and Gupta titled “the long term impact of promotion and advertising on consumer brand choice”.
E.g. we have 260 weeks and we will roll up the data for each 12 week period. Suppose we have data from week 1 to week 52. We create data sets from W1-W12, W5-W16, -------, W41-W52. Empirical research in marketing science says that this window is called the moving window. Instead of 12 weeks which is practiced for understanding the long term Ad effectiveness, forecasting etc we can choose some other number but for consistency sake, we can go ahead with the number 12.
For implementing the above logic, we will need to aggregate the residuals and the entire movement data at the quad week level (as explained above). We will do it only for the target variables and in our case it will be only the base price. Hence, for 5 years of data, we will have the movement data at the quad week level i.e. 66 data points at the quad week level instead of 260 at the store week level. In our case, we will sum up the residuals at the quad week level and average the base price at the quad week level.
The next step after this data preparation will be to model for residuals at the quad week level with base price as the independent. The model form will now be
Residuals = α + Base Price
The model procedure will be a standard Proc Mixed in SAS at the quad week level which will give us 66 estimates. Once we get these estimates, we will calculate the elasticities and use them appropriately.
It is likely that the plot will have high kinks (both upward and downward). In that case, it will be difficult to read a pattern. If this exists, we will use a survival analysis to fit the decay curve and smoothen out the kinks for easy interpretation.
Interpretation:
The last step in the analysis will be to plot all these 66 base price elasticities across weeks along with the base price and observe the pattern change. We expect to see a stable pattern for a long time frame due to the old price, then a sudden increase / decrease in the pattern for some time period and lastly a stable pattern which indicates that there is a rebound in the volume sales.
As a diagnostic, we can also compare the new elasticities with our TUS fixed effect elasticity value. This will help us explain the impact of change observed over 5 years of data. When we say TUS fixed effect elasticity value over the 5 years of data, our assumption is that that the consumers have already got adapted to the new price; hence it should not be quite a high number based on the OJ category standards.
If we do not observe a significant change in the pattern of elasticities, we will get a direction that the overall decline may not be just due to the base price but it can be due to the decline in promotional activity, increase in the competition promotions etc.
Appendix:
The scope of this project could be scaled to as high a level as possible. Empirical research in marketing science considers many variables to find a solution to this problem. The variables considered can be location of the store, population of the state / region, average income of people in that state / region etc. If we model for all of these variables, we will arrive at a solution which will classify the important variables that lead for consumer adoption. This will require us to get to the consumer level data i.e. panel data. If we try to scope it this way, it can grow as big as we want it.
Residuals = Actual Volume – Volume due to promotions.
Expanded Scope:
This approach is a special case in Marketing Science and involves a lot of criteria. Ideally, it would be useful to have fast moving products to be able to see a pattern, a significant price increase which will help the client fix the best price for their product depending on the average length for rebound etc. With this analysis, the client will not only be able to know the average length for rebound but depending on that length, they would be able to decide on the % increase / % decrease they would want to fix.
If the clients are interested in knowing the effect of Increase / Decrease at a granular level viz. Region, CORP RMA etc, we can very well use the Randomization approach using PROC MIXED in SAS.
The scope of this approach could be used to help the client understand the average length taken for the advertisement effect / campaign to decay. This will be very helpful to the clients since they could utilize their advertisement expenses better.
No comments:
Post a Comment