Scholarly article on topic 'Evaluation of Promotional and Cross-Promotional Effects Using Support Vector Machine Semiparametric Regression'

Evaluation of Promotional and Cross-Promotional Effects Using Support Vector Machine Semiparametric Regression Academic research paper on "Economics and business"

CC BY-NC-ND
0
0
Share paper
Academic journal
Systems Engineering Procedia
OECD Field of science
Keywords
{"Support Vector Machines" / "price promotion" / "machinery learning" / "statistical learning"}

Abstract of research paper on Economics and business, author of scientific article — María Pilar Martínez-Ruiz, José Luis Rojo-Álvarez, Francisco Javier Gimeno-Blanes

Abstract Present article illustrates how system engineering procedures and techniques could be applied to marketing and pricing models. In this research it is evaluated, over retail grocery products sales, the promotional and cross-promotional effects based on Support Vector Machines Semiparametric Regression (SVM-SR) technique. More specifically, in this work it is evaluated the interaction effects of combined promotional, for differentiated types of brands. Database was developed using one year scanned sales records from a Spanish hypermarket. Mayor findings were: (i) higher direct sales increment in larger package sizes promoted articles of national premium brands products not incorporating additional or functional ingredients; (ii) relevant cross price effects (both asymmetric and neighbourhood); (iii) higher sales grow on Friday and Saturday; and (iv) enhanced results in combined price discount and advertising feature promotions than any individual promotion.

Academic research paper on topic "Evaluation of Promotional and Cross-Promotional Effects Using Support Vector Machine Semiparametric Regression"

Available online at www.sciencedirect.com

SciVerse ScienceDirect PfOCSCl Í(

Systems Engineering

Systems Engineering Procedia 1 (2011) 465-472

Evaluation of Promotional and Cross-Promotional Effects Using Support Vector Machine Semiparametric Regression

María Pilar Martínez-Ruiza*, José Luis Rojo-Álvarezb, Francisco Javier Gimeno-Blanesc

aUniversity of Castilla la Mancha, Commercialization and Market Research Area, Av. de los Alfares, 44, 16071 Cuenca, Spain bUniversity Rey Juan Carlos, Signal Theory and Communication Department, Camino del Molino s/n, 28943 Fuelabrada, Madrid, Spain cUniversity Miguel Hernández, Signal Theory and Communication Area, Av. de la Univesrsidad s/n, 03202 Elche, Alicante, Spain

Abstract

Present article illustrates how system engineering procedures and techniques could be applied to marketing and pricing models. In this research it is evaluated, over retail grocery products sales, the promotional and cross-promotional effects based on Support Vector Machines Semiparametric Regression (SVM-SR) technique. More specifically, in this work it is evaluated the interaction effects of combined promotional, for differentiated types of brands. Database was developed using one year scanned sales records from a Spanish hypermarket. Mayor findings were: (i) higher direct sales increment in larger package sizes promoted articles of national premium brands products not incorporating additional or functional ingredients; (ii) relevant cross price effects (both asymmetric and neighbourhood); (iii) higher sales grow on Friday and Saturday; and (iv) enhanced results in combined price discount and advertising feature promotions than any individual promotion.

© 2011 Published by Elsevier B.V. Selection and/or peer-review under responsibility of the Organising Committee of The International Conference of Risk and Engineering Management.

Keywords: Support Vector Machines, price promotion, machinery learning, statistical learning

1. Introduction

Sales promotions is becoming a key activity for grocery retailers, especially under current economic conditions, where consumers are especially reactive to price incentives and manufacturers/retailers face turn-over reductions due to economic crisis, and both of them are reacting actively by enforcing theirs competitive profiles. Under this circumstances, retailers, and also manufacturer, find in the promotional activities one strategy, defensive or not, to move forward. Promotion provides mainly a commercial tool to achieve major benefits and sales in the short term, and secondly challenges other competitors' actions. Not surprisingly, consumer response to a sale promotion has motivated considerable research and literature [1, 2, 3, 4, 5]. In particular, several research lines in marketing and management areas have focused their attention on price and promotion impact over sales. Attending to relevant studies, it is generally accepted that price promotion has a clear effect on sales, although its magnitude and how it is translated into real benefits, it would be also related to more detailed and concrete terms.

Price promotions have been long recognized as an effective tool for managing brands [6] and to boost brand performance, not only in the short term [7, 8, 9, 10]. It is remarkable the relevance the promotion activities reached

2211-3819 © 2011 Published by Elsevier B.V. Selection and/or peer-review under responsibility of the Organising Committee of The International

Conference of Risk and Engineering Management.

doi:10.1016/j.sepro.2011.08.068

in recent years, when a rapid growth headed retailers to spend a significant share of their marketing budgets (up to 50% in same cases) in this activities [11]. This fact encourages industry managers to concentrate efforts to know whether and under which conditions, the promotions are most effective to achieve targeted goals. Bemmaor and Mouchoux in 1991 [12] noticed that additional research was needed to understand why certain brands presented better response to specific promotions combinations than other. Mentioned study did not elucidate differential effects and promotions interactions related to differentiated tier brands. This perspective was analysed in other studies [6] that elaborated on different promotions interaction and brand characteristics as a way to maximize synergies among elements inside a certain promotional mix. This research showed the interactions among displays, feature advertising, price promotions and brands inside different price-quality tiers. Using information from scanner and specific test performed, these authors concluded that: (1) separated promotional actions presented a higher response over high-tier brands than over low-tier (displays, feature advertising, or pricing discount); (2) meanwhile, this differentiated effect was not that visible on multi-promotional actions; (3) the combined effects of price deals well with displays or with feature advertising, was higher on low-tier brands.

From the analysis technique stand point, Support Vector Machines Semiparametric Regression (SVM-SR) have been recently reported as suitable quantitative model to evaluate sort-time demand effect due to price-deals, (regardless this increment is related to a higher consumption or to a future needs anticipation) [3, 4, 5].

Published literature and the important latest increase in promotional initiatives from retailers and manufacturer, inspired present research to explore the joint effect of specific promotional tools together with different brands strategies. More specifically, we seek to analyse the interactions among different promotional tools (i.e., deal discounts, buy three get one free promotions and feature advertising) in different price-quality tiers. We studied own-item discount effects as well as two kinds of cross-price effects: asymmetric cross-price effects and neighbourhood cross price effects. It was also observed sales evolution along week-days as well. Promotion effectiveness was also invetigated. SVM-SR methodology was proposed as statistical learning method for this analysis.

The document is structured as follows. The next section describes the SVM-SR model, and the variables used. The third sections presents the database and estimation results. The last numbered section discus the results obtained and summarise research main findings.

2. Model Description and Statistical Specification

2.1. Description of the SVM-SR methodology

Different types of regression methods have been mostly presented in relevant published papers to assess the sales response to temporary price discounts. Aside from parametric and nonparametric regression, a third type of regression, known as semiparametric regression (SR) which arises from the combination of the previous types of regression methodologies, is achieving greater importance. Given the advantages of SR to estimate promotion effects -for example, when adequately combined, SR can provide the benefits of both parametric regression such as efficiency and low variance, and the benefits of nonparametric regression, such as flexibility and small bias [5] -, this is not surprising that this type of regression has achieved growing importance during the few past years [13, 5, 4, 14, 15].

Bearing in mind the benefits of the SR methodology, [5, 4, 14] use Support Vector Machines (SVM) to estimate, among other effects, the shape of the deal effect curve by carrying out a SVM-SR model. In their model, the parametric (linear) component accounted for dichotomic variables, whilst the nonlinear component -the Nadayara-Watson constant kernel estimator- was used to capture and describe the complex cross-effects between metric variables. As these authors suggested, a comparatively high number of data is needed for this method to properly perform. The main reason is due to the nonparametric component of the model, aside from the fact that data overfitting could be easily shown. Although these aspects might limit the applicability of these methods in grocery retailing due to the reduced number of observations, the possibility to work with daily data due to the development

of information and communication technologies such as electronic point-of-sale (POS) and more recently, radio-frequency identification (RFID), is making it increasingly possible to work with such methodologies.

In the SVM-SR method two important elements are used: (i) a robust cost function, the e -Huber cost, for dealing with heteroscedasticity, non-Gaussian noise, and sparse models; and (2) a Mercer's kernels, which makes it possible to yield non-linearity in the model on a flexible way [16].

2.2. Variables definition

In our application of the SVM-SR model, we express brand sales as the sum of a nonparametric function of metric variables -price indices of the own-and-competing brands -and a parametric function of other relevant promotional dichotomic predictors -day-of-week of promotional days and day-of-week of nonpromotional days -.

Hence, our predicted variable is a metric variable, Sales(i, which is the units sold of brand i in day t; and our predicted variables are: PI(l), which is a metric variable that reflects the price index of brand i in day t; a set of dichotomic variables reflecting both the day-of-week of promotional days, Daypro(l), as well as the day-of-week of non-promotional days, Daynonpro\l). In addition, between the predictor variables two more dichotomic variables were incorporated in order to reflect whether there is or not a promotion supported on the use of feature advertising consisting of: (i) a deal discount, Df(), or a bonus pack discount, Bpft(,).

Hence, for each given brand (i) our SVM-SR model can be defined as fo1llows:

Sales((i) = m(piI)+ aT XD + et (1)

where:

• Sales(i : unit sales of brand (i), i=1,...,12 in day t, t=1,...,310;

• m(X,M),: nonparametric function;

• PI1: vector of price indices of brand i, i=1,...,12. Each price index is calculated as the ratio of current to regular price of brand i in day t.

• Xf : vector of D dichotomic variables (day t) consisting of:

O Daypro(): indicators of the day of week -Monday (1) to Sunday (7)- during promotional periods in brand (i);

O Daynonpro\l): indicators of the day of week -Monday (8) to Sunday (14)- during non-promotional periods in brand (i);

O Df(): indicator of price discounts supported by feature advertising in brand (i);

O Bpf(): indicator of bonus pack supported by feature advertising in brand (i);

O a : (transposed) vector of effects of dichotomic descriptor variables;

O et: disturbance term.

As previously highlighted, this model fulfils the metric variables requirement enabling the modelling of complex interactions by means of the nonparametric part of the model. But also, the dichotomic variables are retained by the linear part of the model, making it easy to analyse their effects. As no interaction effects are expected to emerge among these latter variables, there is no need to model them nonparametrically. Finally, it should be noted that confidence intervals (CI) for linear coefficients (95% level) in the parametric part of the models were estimated using bootstrap resampling.

3. Data description and estimation results

3.1. Data

Our database was drawn from a hypermarket located in the northeast of Spain. The information was registered on a daily basis and contains store-level scanner data over a period of one year (from 1st September 2005 to 31st August

2006). For this study 310 daily observations were considered. We focused our study on one product category with major sales during the considered period: milk. This category has been selected on the basis the following criteria. First, this is a frequent purchased product category, pervading sufficient volume to be analysed statistically. The frequent promotions in this category will allow validating the model for any cross-relation effect among brands and products (different type of promotions and brand strategies are found in the category).

Type of Brand Package Format Package Size (Litres) Additional Ingredient

Asturiana clásica 1.5 litros National Premium Tetrabrik 1.5 -

Asturiana combi clásica 1.5 litros National Premium Bottle 1.5 -

Asturiana fibra National Premium Tetrabrik 1 Fibre

Asturiana Naturlínea National Premium Tetrabrik 1 Tonalin

Ato clásica National brand with limited (regional) distribution Tetrabrik 1 -

Ato clásica 1.5 litros National brand with limited (regional) distribution Tetrabrik 1.5 -

Castillo clásica Type of brand: National brand Tetrabrik 1 -

RAM energía y crecimiento National Premium tetrabrik 1 calcium, 12 vitamins and other minerals

Llet clásica National brand with limited (regional) distribution Tetrabrik 1 calcium, 12 vitamins and other minerals

Store Brand Milk Store Brand Tetrabrik 1 -

Pascual clásica National Premium Tetrabrik 1 -

Puleva calcio National Premium Tetrabrik 1 Calcium

Table 1. Product category details

Brand R2 RMSE Brand R2 RMSE

Training Test Training Test Training Test Training Test

Asturiana clásica 1.5 0.6440 0.2113 31.1875 47.9366 Castillo clásica 0.4608 0.1383 10.666 17.6877

Asturiana combi clásica 1.5 0.7549 0.3323 79.7838 116.9781 RAM energía y crecimiento 0.1649 0.0086 7.1301 4.549

Asturiana fibra 0.4103 0.3141 4.1964 5.7212 Llet clásica 0.74609 0.1883 48.4889 75.7389

Asturiana Naturlínea 0.3093 0.1908 6.5803 5.8697 Marca blanca 0.6771 0.2507 50.4195 64.3604

Ato clásica 0.5185 0.1980 26.2471 32.8986 Pascual clásica 0.7436 0.2118 57.6201 111.8934

Ato clásica 1.5 0.6751 0.3710 48.604 93.1814 Puleva calico 0.9905 0.1906 2.277 113.5363

Table 2. Estimation results

The category analysed contain seven national premium brands, three national brands with regional distribution, one national brand and one store brand. Category breakdown is provided in the Table 1.

3.2. Results

The estimation results obtained with SVM-SR model are shown in Table 2. 75% of daily information was applied for training and estimation, and the remaining 25% of the data was used for validation proposes. The R2 and the conventional root mean squared error (RMSE) were calculated in each model for both training and test subsets, in order to prevent from overfitting.

Assessment of own-price effects

The SVM-SR model enables to evaluate how brand sales respond to its own promotions by means of the examination of the own-item deal effect curves. The analysis of the curves for this category has evidenced how the greatest sales increments have been observed, in this order, in Asturiana combi clásica 1.5 litros, Ato clásica and Asturiana clásica 1.5 litros, whilst the lowest have been detected in Asturiana Naturlínea, Puleva calcio and RAM energía y crecimiento. Thus, the greatest effects have been observed mainly in those national premium brands not containing any additional/functional ingredient and they are also commercialized in the largest package sizes. Figure 1 shows an example of the results obtained for the own-item deal effect curve obtained for the brand with the greatest predicted increment in sales units, Asturiana combi clásica.

Figure 1. Own-item deal effect curve for Asturiana combi clásica.

Assessment of cross-price effects

The SVM-SR model makes it possible to capture the complex nature of the relationship between sales and price promotions since the nonparametric part of Equation (1) allows accommodating flexible interaction effects betweeen different brands promotions. In particular, the nonparametric part of Equation (1) allows us to build up three-dimensional deal effect surfaces to assess cross-price effects.

In Figure 2 we it is presented an example of a three dimensional deal effect surface. The vertical axis of this surface represents the increments in the predicted sales volume for one brand, whilst the other two axes represent the price index of that particular brand as well as a competing brand. By analysing different curves within the three dimensional deal effect surface, it is possible to detect substitution patterns within the category. Hence, the curve AB indicates the sales obtained by the own-brand in response to its own promotions, and in the absence of promotions in the competing brand considered (i.e., this is the own-item deal effect curve). The curve C-D evidences how own-brand sales vary when this brand offers different discount levels and the competing brand is offered at its maximum discount. The curve B-C is called cross-item deal effect curve, and shows how own-brand sales change in response to the different levels of discounts offered in the competing brand (considering also that there is not any promotion

in the own brand). The curve D-A illustrates how own-brand sales vary considering that the own brand is offered at its maximum discount and the competing brand is sold at different discount levels.

¿- u.34 us;

PI Asturiana combi clasica p, Aslutiana 15 da3lca

Figure 2. Example of a three-dimensional deal effect surface: interactions between the discount levels of Asturiana clásica 1.5l

A deep evaluation of the three-dimensional deal effect surfaces obtained in the category, led us to conclude that asymmetric cross-price effects was present as promotions offered by higher-priced brands affected lower-priced brands sales. Deals offered by lower-priced brands also affect higher-priced brands sales, although the magnitude was smaller. This finding was consistent with the price promotions published in literature [17, 18, 19]. Neighbourhood cross price effects was also detected as it was found that similarly priced brands had a larger cross-price effects than brands not similarly priced [19].

Bootstrap resampling estimates

Confidence intervals and statistical validation of the model was obtained based on bootstrap resampling. Results showed sales increase magnitude for each day of the week. As models are additive, estimations with confidence intervals not including 0 indicate: (i) positive values showed a sales increase for that considered day; and (ii) negative value indicates sales decrease. Figure 3 shows an example of bootstrap resampling estimates.

Sale units Intercept

0 5 10 15 16 17 18

# variable

Figure 3. Example of bootstrap resampling estimate: results obtained for Asturiana combi clásica.

Results showed for all brands, with the only exception of Puleva calcio, a growing pattern in sales during Friday and Saturday. It was also observed a negative trend on Sunday. These findings allow us to confirm the positive influence of Fridays and Saturdays, as detected by previously studies [3, 4, 5].

It was noticed - again with the only exception of Puleva Calcio - that the sales increment led from featured promotions were larger for deals discounts than the ones led by three get one free promotions. This finding is also

consistent with the results obtained by previous studies evaluating the efficiency of bonus pack promotions in general. Some of these works even evidenced how bonus pack promotions lack credence [20].

4. Final Discussion

This work has evaluated the interaction and combined effects of different promotional tools (i.e., deal discounts, buy three get one free promotions and feature advertising) with different price-quality tiers brands. In order to achieve this goal, a SVM-SR methodology has been used. In particular, as evidenced by previous studies [5] this methodology can provide the benefits of both nonparametric and parametric regression.

In order to illustrate the usefulness of SVM-SR to estimate future promotion effects in retail, the software application was developed using one year daily real scanner data from Spanish hypermarket database. The present study took different brands (i.e., national premium brands; national brands; national brands with regional distribution and one store brand) and products packages for a specific product category with frequent promotions during the considered period.

Several relevant effects were observed. First, the analysis of the own-item deal effect curves -obtained from the nonparametric part of the model - showed greatest sales increments mainly in the national premium brands not containing any additional ingredient and commercialized in the largest package sizes. Second, detailed evaluation of three-dimensional deal effect surfaces - also obtained from the nonparametric part of the model - confirmed the existence of both, asymmetric and neighbourhood cross price effects in the category. So, it could be stated that promotions offered by higher-priced brands affect sales of lower-priced brands more than the reverse. And also brands similarly priced had larger cross-price effects than brands priced further apart. Findings were consistent with particular previous studies also analysing cross-price effects [19].

The parametric part of the model enabled us to obtain the parameter estimates for the brands in the category considered. At this point it should be highlighted that for all brands, with the only exception of Puleva Calcio, estimates indicated a growing pattern in sales during two days of the week, Friday and Saturday. This was consistent with previous works [3, 4, 5]. In addition, a descendant pattern was present on Sundays.

It was also confirmed that the positive effect of featured promotions were larger for price discounts than for buy three get one free deals. This effect is also consistent with specific previous works in the promotions literature that also showed lack credence for bonus packs promotions [20].

Readers should be aware of the following limitations in this study. First, the analysed promotion effects might be particular to the considered category as well as to the specific conditions of the study: store format (hypermarket); number and types of brands sold by the retailer in the category (relatively high number of national premium brands versus low priced brands); competitive environment. Thus, further research is required to widen results and generalize these findings to other categories, and study other promotional tools (e.g., displays) and effects (e.g., store switching).

Acknowledgements

This research has been funded by Fundación Ramón Areces (Spain).

References

1. R. Blattberg and S. Neslin, Sales promotion: Concepts, methods, and strategies. Prentice Hall Englewood Cliffs, NJ, 1990.

2. H. V. Heerde, P. Leeflang, and D. Wittink, "Semiparametric analysis to estimate the deal effect curve," Journal of Marketing Research, vol. 38, pp. 197-215, May 2001.

3. M. Martinez, A. Molla, and J. Rojo, "Using daily store-level data to understand price promotion effects in a semiparametric regression model," Retailing and Consumer Services, vol. 3, no. 13, pp. 193-204, 2006.

4. M. Martnez-Ruiz, A. Molla-Descals, M. Gomez-Borja, and J. Rojo-Alvarez, "Evaluating temporary retail price discounts using semiparametric regression," Journal of Product & Brand Management, vol. 15, no. 1, pp. 73-80, 2006.

5. M. Martinez-Ruiz, M. Gomez-Borja, A. Molla-Descals, and J. Rojo-Alvarez, "Using support vector semiparametric regression to estimate the effects of pricing on brand substitution," International Journal of Market Reserach, 2008.

6. K. Lemon and S. Nowlis, "Developing synergies between promotions and brands in different price-quality tiers," Journal of Marketing Research, vol. 39, no. 2, pp. 171-185, 2002.

7. V. Nij s, M. Dekimpe, J. Steenkamp, and D. Hanssens, "The category-demand effects of price promotions," Marketing Science, pp. 1-22, 2001.

8. R. Slotegraaf and K. Pauwels, "The impact of brand equity and innovation on the long-term effectiveness of promotions," Journal of Marketing Research, vol. 45, no. 3, pp. 293-306, 2008.

9. S. Srinivasan, K.Pauwels, D. Hanssens, and M. Dekimpe, "Do promotions benefit manufacturers, retailers, or both?," Management Science, vol. 50, no. 5, pp. 617-629, 2004.

10. J. Steenkamp, V. Nijs, D. Hanssens, and M. Dekimpe, "Competitive reactions to advertising and promotion attacks," Marketing Science, pp. 35-54, 2005.

11. J. Quelch, "How to market in a recession," Marketing KnowHow, September 2008.

12. A. Bemmaor and D. Mouchoux, "Measuring the short-term effect of in-store promotion and retail advertising on brand sales: A factorial experiment," Journal of Marketing Research, vol. 28, no. 2, pp. 202-214, 1991.

13. P. Robinson, "Root-n-consistent semiparametric regression," Econometrica: Journal of the Econometric Society, pp. 931-954, 1988.

14. E. Martínez and T. Montaner, "The effect of consumer's psychographic variables upon deal-proneness," Journal of Retailing and Consumer Services, vol. 13, no. 3, pp. 157-168, 2006.

15. W. Steiner, A. Brezger, and C. Belitz, "Flexible estimation of price response functions using retail scanner data," Journal of retailing and consumer services, vol. 14, no. 6, pp. 383-393, 2007.

16. J. Rojo-Álvarez, M. Martnez-Ramón, M. de Prado-Cumplido, A. Artés-Rodrguez, and A. Figueiras-Vidal, "Support vector method for robust arma system identification," Signal Processing, IEEE Transactions on, vol. 52, no. 1, pp. 155-164, 2004.

17. R. Blattberg and K. Wisniewski, "Price-induced patterns of competition," Marketing Science, vol. 8, no. 4, pp. 291-309, 1989.

18. R. Sethuraman, "A model of how discounting high-priced brands affects the sales of low-priced brands," Journal of Marketing Research, vol. 33, no. 4, pp. 399-409, 1996.

19. R. Sethuraman, V. Srinivasan, and D. Kim, "Asymmetric and neighborhood cross-price effects: Some empirical generalizations," Marketing Science, pp. 23-41, 1999.

20. B. Ong, F. Ho, and C. Tripp, "Consumer perceptions of bonus packs: an exploratory analysis," Journal of Consumer Marketing, vol. 14, no. 2, pp. 102-112, 1997.