THE “VERY” SHORT-TERM
EFFECT OF TV AND RADIO
ADVERTISEMENT ON ONLINE
SHOPPING BEHAVIOR
Hylke Vietmeijer
THE “VERY” SHORT-TERM EFFECT
OF TV AND RADIO ADVERTISEMENT
ON ONLINE SHOPPING BEHAVIOR
Hylke Vietmeijer S2408414
Master thesis MSc Marketing Rijksuniversiteit Groningen First supervisor: dr. P.S. (Peter) van Eck
Second supervisor: dr. E. (Evert) de Haan 2020-01-11
Management summary
Preface
Table of contents MANAGEMENT SUMMARY ... 2 PREFACE ... 3 INTRODUCTION ... 5 LITERATURE REVIEW ... 6 DATA ... 8 METHODOLOGY ... 11 RESULTS ... 13
CONCLUSIONS & RECOMMENDATIONS ... 20
Introduction
For a marketing department it is crucial to quantify the effects of their operations, not only to evaluate efforts in retrospect, but also to optimize their marketing mix. The task of
quantifying the effects is not an easy one and has been the subject of many researches (Joshi and Hanssens 2010, Sethuraman et al. 2011, Edeling and Fischer 2016). The internet has greatly increased the tangibility of marketing campaigns, thanks to all the tracking and logging that is possible in a digital environment. This has allowed companies to apply
attribution modelling, where a key performance indicator is attributed to a specific or multiple marketing campaigns, in order to estimate the actual impact of their marketing. With sales attributed to marketing campaigns, the return on investment (ROI) can be calculated and used to optimize the marketing mix. A common tool to evaluate ROI and optimize the allocation of marketing resources is marketing mix modelling (MMM) (Kitchen 2010). The success of MMM greatly depends on the quality of the data used and the accuracy of the estimated effects.
While the internet has greatly altered the world of marketing, traditional media, i.e., TV and radio, still represents a majority of the marketing budget for many companies. The impact of traditional media cannot be measured as easily as online advertisement, as it is unfeasible to log who was exposed to the advertisement and there is no direct tracking of behavior after the advertisement. Due to the aforementioned constraints, companies often rely on earlier research in order to estimate the impact of their campaign. A substantial amount of research has been done on the long-term and day to day effect of traditional media on sales. Online shops who want to attribute their marketing on the most granular level possible, however, prefer results on a lower aggregation level than day and sales, because they are able to track their performance on a session level and measure metrics such as visits, basket value and conversion rates.
a large Dutch online retailer for which the session, transaction and advertisement data are all combined into one dataset. Two models will be applied on this dataset, both a statistical model and a machine learning model. In marketing research both types are used and there are arguments for both to make. Statistical models are easier to use and have easier interpretable results, whereas machine learning models usually outperform the statistical models in fitting and predicting the data. The statistical model will be a linear regression as it is the simplest and of the most widely used models applied. The machine learning model will also be a regression but using gradient boosting algorithms. Gradient boosting has shown great results and is often used to win online machine learning competitions.
This research will help the online retailer to gain a better understanding on the effect of their offline marketing and increase the accuracy of their attribution and MMM. These improved accuracies will in turn lead to improved budget allocation and an easier time for the marketing department to justify their budget. In addition, the short-term effects can be used to predict and better handle the increased site load that occurs after TV commercials.
For academic contributions, this paper adds to the existing body of literature by estimating the effects of offline marketing in an online setting on a very granular level. In addition, the synergy between TV and radio is analyzed, which has not been done on a very granular level. The robustness of the effects is also extra challenged in this paper compared to equivalent researches, by utilizing both a statistical model and a machine learning model, which are both used extensively in practice.
The rest of the paper is organized as follows. First the existing literature is reviewed to develop a theoretical basis for the chosen models and the hypotheses are stated. Next the data is described and the methodology is formulated. Subsequently the models are estimated and the hypotheses are validated. Finally, the results are interpreted and afterwards the research will be discussed
Literature review
This review will start with the broad problem of measuring the effectiveness of advertising spend and will gradually move towards the exact problem of measuring the short-term effect of offline advertisement on online shop behavior.
allocations optimized using the results from the model and the last click attribution used by the company. The budget proposed using the model yielded a 21% revenue increase over the status quo, whereas the last click attribution led to 10%-12% less revenue opposed to the status quo. This highlights the importance of correctly estimating effectiveness of
advertisements for companies. Good estimations can lead to improved budget allocations while bad attributions can lead to worse allocations. In addition, the long-term effect of traditional media can’t be inferred in a straightforward manner and takes a long time to measure. Long term effects are therefore most commonly calculated using approximations from established researches and often reported as a multiplier of the short-term effect (Wildner and Modenbach 2015, Wood and Poltrack 2015). The short-term effect is usually derived using statistical methods, i.e., regression, that compare the results of a day or week when an advertisement was shown to days or weeks before and or after the advertisement. While this method has been successfully used to estimate the short-term effect (Lambin 1976, Leone and Schultz 1980, Assmus, Farley and Lehmann 1984, Tellis 1988, Kanetkar,
Weinberg, and Weiss 1992, Sethuraman and Tellis 1991, Pedrick and Zufryden 1993,
Srinivisan et al. 2016, Liaukonyte et al. 2015, Nibbering et al. 2017) the results vary between studies and there is no consensus on the existence of a significant short-term effect. A
possible explanation for the conflicting results comes from Tellis and Weiss (1995) they hypothesized that the positive effects of advertising found where “spurious results reached by aggregation of the data over time and households”. Kitts et al. (2014) had some success creating a “web-based TV conversion tracking system” by linking customers to television advertisements using the website traffic spikes that followed after the advertisement showing that there is an immediate effect of TV advertisements on websites. Therefore, it is plausible that other aggregations on a too high level can lead to incorrect results and a very short-term approach might lead to more consistent results. This leads to the research question:
Is there a significant short-term effect of TV and radio advertisements on online shopping behavior?
significant effects of offline advertisement on website traffic, online transactions and search ad referrals. Both these studies were done using data which only looked at a period of 2 hours or less around each advertisement. This shows that there is a direct response of customers that view offline advertisement on website behavior. However, they did not take into account website behavior such as conversion rate and visit lengths. While the effect of advertisement on website visits will most likely be positive as found by earlier research. The effects of conversion rate and visit lengths are less straightforward. While the advertisements could lead to increased engagement and willingness to buy it also possibly creates an influx of customers for which these levels are lower than those of the customers that would visit the site
irrespective of the advertisements, which in turn leads to overall lower conversion rates and average site lengths. The earlier research question is narrowed down further in order to fit to an online setting creating the following three sub questions:
Do TV and radio advertisements result in a short-term uplift in website visits? What is the short-term effect of TV and radio advertisement on conversion rate? What is the short-term effect of TV and radio advertisement on visit lengths?
Advertisements are often part of bigger marketing campaigns and often multiple marketing channels are used during the same period. Lim 2015 found that participants to their research “have greater perceived message credibility, ad credibility, and brand credibility than counter- parts exposed to repetitive ads from a single medium.” Multiple quantitative
researches have also studied synergy effects between channels and found a significant effect on a day-to-day level (Lin et al. 2013, Naik and Raman 2003, Naik and Peters 2009 and Stolyarova and Rialp 2014). For the very short-term this synergy may also exist. However, for the very short-term the advertisement effects are based more on impulsive responses the synergy effects might be a lot less pronounced. This leads to the sub question:
Is there a significant short-term synergy effect between TV and radio advertisements on online customer behavior?
Data
As mentioned in the introduction the data used for this research comes from a large online retailer in the Netherlands. The data spans a period of almost three and a half year (2017-07-01 to 2020-12-06) during which over 34 thousand radio and TV commercials were
information such as time and place of showing, but also measures of impact in the form of gross rating points (GRP).
GRP is a measure of exposures relative to the size of a target population. It is
calculated by taking the percentage of the target population that has seen your advertisement and multiply it by the number of exposures. So, when twenty percent of your target audience has seen the ad and they were exposed on average three times to the ad, the GRP would be sixty. GRPs are frequently used in research when calculating the effects of offline media and Nibbering et al. (2017) who also studied the effect of offline media on online behavior used GRP to model TV and radio. Neijens and Voorveld (2015) state that the value of GRPs has increased since the advent of the internet, however Findley et al. (2020) found that the value for radio and TV has decreased the value of GRP and proposed an alternative measure called persuasion rating point (PRP). For this research PRP was not yet available as it is a new method and has not been widely adopted.
The session data has been collected using Google Analytics. Website and app behavior is collected by snippets of code present on each page that sends all the events that happen on the page to a server. A session is a bundle of events which can be seen as a single shopping session of the customer. The conversion rate is calculated by dividing the number of
seasonality’s and events should be taken into consideration as marketers are more likely to time their advertisements in accordance to popular moments to shop.
FIGURE 1.
Distribution of total sessions per hour and weekday.
FIGURE 2.
Distribution of total sessions per day.
TABLE 1. Sessions smoothed Conversion rate Visit length TV GRP Radio GRP TV adstock Radio Adstock Mean 33573 0.0404 276 0.0425 0.0396 0.7325 1.2611 Std dev 8557 0.0109 51 0.3911 0.3493 2.1887 3.6064 Min 3708 0.0000 16 0.0000 0.0000 0.0000 0.0000 Max 266393 0.1889 1203 35.0873 11.84053 36.1986 33.8390
Descriptive statistics of primary variables.
Methodology
A common solution for the aforementioned seasonality is to include variables to the model that control for the variation. However, this introduces many extra variables which can lead to overfitting and also does not allow simple models to change the effect of other independent variables based on the specific time. For advertisement it makes sense that the effect is proportional to level of traffic on the site as these correspond with moments in time when customers are more likely to be available to shop. To circumvent these issues the day and week seasonality are controlled by smoothing the sessions based on the average share per 5-minute interval per weekday. The monthly seasonality will be controlled by using dummy variables as the seasonality is less profound and stable over time. For other exogenous effects, such as weather, promotions, holidays, pandemic, etc. the lag of the dependent variable is added.
It is expected that the advertisements not only have an effect on the sessions in the 5-minute interval in which it is shown but also in the following 5-minutes and hours. The commercials that were shown prior to the current interval are added to the model as
dependent variables in two forms. First as lags which measure direct effects after x amount of minutes have passed. Secondly in the form of advertising adstock, which sums all the
intervals for which the adstock lingers is set to 48 which corresponds to four hours. The interaction between TV and radio is added as the product of the two adstocks. The decay function follows the following formula, where x is the input GRP, decay is the decay value and t is the amount of intervals that have passed:
𝑎𝑑𝑠𝑡𝑜𝑐𝑘 = 𝑥 ∗ 𝑑𝑒𝑐𝑎𝑦−𝑡 where t < 49
For estimating the effects of offline advertisement, a linear regression will be estimated. A linear regression has the big advantage of having results which are easily interpretable. The coefficients can be seen as the direct impact if the underlying variable was increased with one. Due to the smoothing of the session the relation is not completely linear anymore for the sessions model, this means that the effects are relative to the specific hour and weekday. This makes sense as it is expected that an advertisement will likely create more sessions when it is affecting a time and day when people are more inclined to shop. The formula for the sessions, conversion rate and visit length models are:
𝑦 = 𝛽0+ 𝛽1𝑇𝑉 + 𝛽2𝑇𝑉−1 + 𝛽3𝑇𝑉−2+ 𝛽4𝑇𝑉−3+ 𝛽5𝑅𝑎𝑑𝑖𝑜 + 𝛽6𝑅𝑎𝑑𝑖𝑜−1 + 𝛽7𝑅𝑎𝑑𝑖𝑜−2
+ 𝛽8𝑅𝑎𝑑𝑖𝑜−3+ 𝛽9𝑇𝑉𝑎𝑑+ 𝛽10𝑅𝑎𝑑𝑖𝑜𝑎𝑑 + 𝛽11𝑇𝑉𝑎𝑑∗ 𝑅𝑎𝑑𝑖𝑜𝑎𝑑+ 𝛽12𝑦−1 + 𝛽13−24𝑀𝑜𝑛𝑡ℎ𝑠
𝑦𝑡: Dependent variable at time t
𝛽0: Intercept
𝑇𝑉𝑡: TV GRP’s at time t 𝑅𝑎𝑑𝑖𝑜𝑡: Radio GRP’s at time t
𝑇𝑉𝑎𝑑: TV adstock 𝑅𝑎𝑑𝑖𝑜𝑎𝑑: Radio adstock
𝑀𝑜𝑛𝑡ℎ𝑠: Months from January to December, leaving out February
also effect of the advertisement variables can be inferred using shapley values and by comparing the model fit for models with and without the advertisement variables. Shapley values have enjoyed much interest from the data science community for uncovering the relations within “black box” models. It is a solution concept within game theory introduced by Shapley 1951. It is a measure of the added value of each player in a game. It does so by calculating the end result for every combination of players in the game. For the XGBoost model it means that it is a measure of the impact of each variable in determining the end result. In the simplest case it calculates the impact of a variable on one customer by comparing the output for both a model that does and does not contain this variable.
Results
The first model estimated is the linear regression with the sessions as independent variable for which the results can be found in Table 1. Model fit was excellent with an r-squared of 0.92 however, this is mainly driven by the inclusion of the lagged sessions. All the TV GRP coefficients are significant and show that there is direct positive effect between TV advertisements and sessions. The TV adstock and third lag of the TV GRP are negative, however this does not mean that the advertisement effect is negative. This is due to the lag being present in the model, for example the direct effect of TV GRP is 415.88 during the interval where the advertisement was shown for the next interval this effect persists through the lag of the sessions and 0.95*415.88 of the effect remains. To get a better understanding the effect of a single TV GRP on sessions for each interval after airing is presented in figure 3. This shows that there is a spike in the uplift in sessions for the intervals that start after 5 and 10 minute and afterwards there is a persisting but fast declining uplift which dies down just after three hours. For Radio only the first lag of radio GRP and the adstock are
significant. The effect of 1 radio GRP on sessions for multiple intervals can be found in figure 4. This shows that radio has a smaller effect than TV but the effect over multiple intervals is completely different. The effect of radio only starts after 5 minutes with a small uplift which keeps increasing for subsequent intervals until around the two-hour mark after which it gradually declines. The numbers in these figures are dependent on the time of airing because the sessions data was smoothed over the hours and weekdays. The interaction between TV and radio adstock was not significant and thus not synergy effect was present in the model.
TABLE 2. Coefficients Std Error Intercept 1513.96*** 27.45 sessions t-1 0.95*** 0 TV GRP 415.88*** 14.49 TV GRP t-1 452.55*** 14.11 TV GRP t-2 52.87*** 14.07 TV GRP t-3 -45.47*** 13.98 Radio GRP -4.34 15.95 Radio GRP t-1 65.88*** 15.78 Radio GRP t-2 -0.26 15.65 Radio GRP t-3 23.07 15.77 January 143.83*** 27.68 March 116.03*** 27.49 April 410.39*** 28.33 May 232.82*** 27.89 June 229.84*** 27.77 July 81.77*** 25.84 August -50.1*** 25.68 September 173.46*** 25.92 October 188.6*** 26.26 November 566.95*** 26.97 December 336.72*** 27.47 TV adstock -19.49*** 3.39 Radio adstock 17.88*** 1.99 TV * Radio adstock 7.84 1.06 Notes: ***p<0.01, **p<0.05, *p<0.1
FIGURE 3.
The effect of 1 TV GRP on sessions for each interval after airing.
FIGURE 4.
The effect of 1 Radio GRP on sessions for each interval after airing.
The results of the conversion rate model can be found in Table 2. While there are significant effects of both radio and TV in the model, none of the coefficients are of sufficient magnitude to have impact in the real world. Model fit is subpar with an r-squared of 0.54 considering that the model included a lag of the dependent variable.
TV GRP t-1 452.55*** 14.11 TV GRP t-2 52.87*** 14.07 TV GRP t-3 -45.47*** 13.98 Radio GRP -4.34 15.95 Radio GRP t-1 65.88*** 15.78 Radio GRP t-2 -0.26 15.65 Radio GRP t-3 23.07 15.77 January 143.83*** 27.68 March 116.03*** 27.49 April 410.39*** 28.33 May 232.82*** 27.89 June 229.84*** 27.77 July 81.77*** 25.84 August -50.1*** 25.68 September 173.46*** 25.92 October 188.6*** 26.26 November 566.95*** 26.97 December 336.72*** 27.47 TV adstock -19.49*** 3.39 Radio adstock 17.88*** 1.99 TV * Radio adstock 7.84 1.06 Notes: ***p<0.01, **p<0.05, *p<0.1
Linear regression results for conversion rate model.
TABLE 4. Coefficients Std Error Intercept 70.54*** 0.48 Visit length t-1 0.74*** 0 TV GRP 2.22*** 0.2 TV GRP t-1 2.01*** 0.2 TV GRP t-2 0.84*** 0.19 TV GRP t-3 1.21*** 0.19 Radio GRP -0.92*** 0.22 Radio GRP t-1 -0.25 0.22 Radio GRP t-2 -0.46** 0.22 Radio GRP t-3 -0.51** 0.22 January 1.15*** 0.38 March -1.22*** 0.38 April 1.35*** 0.38 May 2.08*** 0.38 June 1.97*** 0.38 July 2.91*** 0.36 August 0.42 0.36 September 1.05*** 0.36 October -3.73*** 0.36 November -3.98*** 0.36 December -0.98*** 0.37 TV adstock 0.5*** 0.05 Radio adstock 0.24*** 0.03 TV * Radio adstock 0.09*** 0.01 Notes: ***p<0.01, **p<0.05, *p<0.1
FIGURE 5.
The effect of 1 TV GRP on visit length in seconds for each interval after airing.
FIGURE 6.
The effect of 1 Radio GRP on visit length in seconds for each interval after airing.
To find the added effect of the advertisement variables in a machine learning model a gradient boosted regressor was created for both a dataset with and without the advertisement variables. The r-squared and root mean square error (RMSE) are in table 4. This shows that the
advertisement variables only slightly improved the model fit with the RMSE decreasing with almost 3%. To investigate the fitted model the shapley values are calculated and are shown in figure 7. The ordering of the variables is based on their importance, a measure of how much of the predictions used this variable, for the model and the x axis for each variable are all the observations whose prediction depended on that variable. The x axis shows the shapley value
so observations on the left were given a negative output by the variable and observations on the right were given a positive output. The color corresponds to the value of the variable, blue low and red high. It is clear from this figure that the lag of sessions does almost all the work, it has the highest importance and has relatively a lot of observations tied to it. For the
advertisement variables that have been used by the model the impact is ambiguous as there is not clear correlation between the feature value and shapley value. Except for TV GRP, its lag and the interaction between TV and radio adstock, it is clear that most high valued
observations (red) got a positive effect on the end result and vice versa for the interaction. But overall, the model does not fit much observations on variables other than the lag of sessions.
TABLE 5.
Without With
r-squared 0.9300 0.9337
RMSE 2279 2218
Model performance measures XGBoost model with and without advertisement variables.
FIGURE 7.
Conclusions & Recommendations
In line with the findings from Liaukonyte et al. (2015) and Nibbering et al. (2017) the linear regression models found a significant and positive short-term effect of TV and radio
advertisement on the number of sessions. In addition, the model for visit length also found a significant and positive effect for TV and radio advertisement. This indicates that
advertisement in the short-term not only works as a pull mechanism, but it can also influence how customers go through the shop. For conversion rates there were also significant effects found by the model but the magnitude of the effects was negligible. As expected, the models also showed that the effect of advertisement lingers over time with distinct differences over time between TV and radio. While TV has a high peak in the first ten minutes and a steady decline afterwards, radio shows a gradual increase up to around two hours and a slow decline afterwards. This shows that the optimal timing for the television advertisements is around ten minutes before the peak shopping time, whereas for radio it should be planned two hours before the peak. This can be explained by the fact that people listen radio while focusing on some other task, such as driving, studying or gaming. TV on the other hand requires more focus and watching it is often the main focus of people while consuming it. In addition, people often have an electronic device with them while watching TV which allows for quick website visits.
No significant synergy effects were found for both the sessions and conversion rate models. For the visit length model there was a significant and positive effect however, the magnitude of the effect was very small and would add less than a millisecond when looking at average GRP’s. These results stand in contrast to the existing literature, but the existing literature on synergy effects only has looked on day and week levels of aggregation. For the short-term the synergy may not work impulsively enough to make a big enough difference in a short time period. Also, for short-term it is very unlikely for a customer to have both been exposed to a radio and a TV advertisement within a time period of 4 hours. Thus, a lack of actual synergy might also have dampened the results.
The XGBoost model showed that the advertisement variables had a modest effect on the model fit and that they played no major roles in the prediction for the majority of
could not explain. To further investigate this hypothesis future research can try statistical models that allow for nonlinear relations. In the dataset it is also not known whether a customer has seen the advertising or not and for the largest percentage the answer would be no. This means that changes in the behavior of customers that were actually exposed to the advertisement is difficult to extract from the giant pool of customer behavior not exposed. A good follow up research would be to investigate these research questions but with a dataset where it is known per customer who was exposed to an advertisement.
This research has used GRP’s as a measure for the advertisement exposures, but Findley et al. (2020) has shown that PRP is a better alternative when looking at TV and radio. Future research might consider using PRP over GRP and research in what ways the results differ between the two measures. Another limitation of this research was that the control for exogenous variables was taken care in a non-elegant way by introducing a lag of the
dependent variable. However, this method is not guaranteed to give good results. A model that has more control variables might be able to get a cleaner picture of what the effect of advertisement is.
References
• Assmus, G., Farley, J. U., & Lehmann, D. R. (1984). How advertising affects sales: Meta-analysis of econometric results. Journal of Marketing Research, 21(1), 65-74. • Chen, T., & Guestrin, C. (2016, August). Xgboost: A scalable tree boosting system.
In Proceedings of the 22nd acm sigkdd international conference on knowledge
discovery and data mining (pp. 785-794).
• De Haan, E., Wiesel, T., & Pauwels, K. (2016). The effectiveness of different forms of online advertising for purchase conversion in a multiple-channel attribution
framework. International Journal of Research in Marketing, 33(3), 491-507.
• Edeling, A., & Fischer, M. (2016). Marketing's impact on firm value: Generalizations from a meta-analysis. Journal of Marketing Research, 53(4), 515-534.
• Findley, F., Johnson, K., Crang, D., & Stewart, D. W. (2020). Effectiveness and Efficiency of TV's Brand-Building Power: A Historical Review: Why the Persuasion Rating Point (PRP) Is a More Accurate Metric than the GRP. Journal of Advertising
• Ilfeld, J. S., & Winer, R. S. (2002). Generating website traffic. Journal of Advertising
Research, 42(5), 49-61.
• Joshi, A., & Hanssens, D. M. (2010). The direct and indirect effects of advertising spending on firm value. Journal of marketing, 74(1), 20-33.
• Kanetkar, V., Weinberg, C. B., & Weiss, D. L. (1992). Price sensitivity and television advertising exposures: Some empirical findings. Marketing Science, 11(4), 359-371. • Kitchen, P. J. (2010). Integrated brand marketing and measuring returns. In Integrated
brand marketing and measuring returns (pp. 1-8). Palgrave Macmillan, London.
• Kitts, B., Bardaro, M., Au, D., Lee, A., Lee, S., Borchardt, J., ... & Wadsworth-Drake, J. (2014, August). Can television advertising impact be measured on the web? web spike response as a possible conversion tracking system for television. In Proceedings
of the Eighth International Workshop on Data Mining for Online Advertising (pp.
1-9).
• Lambin, J. J. (1976). Advertising, competition and market conduct in oligopoly over time; an econometric investigation in western European countries.
• Leone, R. P., & Schultz, R. L. (1980). A study of marketing generalizations. Journal
of Marketing, 44(1), 10-18.
• Liaukonyte, J., Teixeira, T., & Wilbur, K. C. (2015). Television advertising and online shopping. Marketing Science, 34(3), 311-330.
• Lin, C., Venkataraman, S., & Jap, S. D. (2013). Media multiplexing behavior: Implications for targeting and media planning. Marketing Science, 32(2), 310-324. • Naik, P. A., & Peters, K. (2009). A hierarchical marketing communications model of
online and offline media synergies. Journal of Interactive Marketing, 23(4), 288-299. • Naik, P. A., & Raman, K. (2003). Understanding the impact of synergy in multimedia
communications. Journal of marketing research, 40(4), 375-388.
• Neijens, P., & Voorveld, H. (2015). Cross-Platform Advertising: Current Practices and Issues for the Future: Why the Gross Rating Point Metric Should Thrive in Today's Fragmented Media World. Journal of Advertising Research, 55(4), 362-367. • Nibbering, D., Frasincar, F., & Vandic, D. (2017). Analysing the effect of offline
media on online conversion actions. International Journal of Web Engineering and
Technology, 12(2), 165-185.
• Nielsen, D. (2016). Tree boosting with xgboost-why does xgboost win" every"
• Sethuraman, R., & Tellis, G. J. (1991). An analysis of the tradeoff between advertising and price discounting. Journal of marketing Research, 28(2), 160-174.
• Sethuraman, R., Tellis, G. J., & Briesch, R. A. (2011). How well does advertising work? Generalizations from meta-analysis of brand advertising elasticities. Journal of
Marketing Research, 48(3), 457-471.
• Shapley, L. S. (1951). Notes on the n-Person Game—II: The Value of an n-Person Game.
• Sheih, G. W., Srinivisan, S. K., Velayutham, S. K., Mahajan, R., Iyer, S., & Khan, H. (2016). U.S. Patent No. 9,276,972. Washington, DC: U.S. Patent and Trademark Office.
• Stolyarova, E., & Rialp, J. (2014). Synergies among advertising channels: An efficiency analysis. Journal of Promotion Management, 20(2), 200-218.
• Tellis, G. J. (1988). The price elasticity of selective demand: A meta-analysis of econometric models of sales. Journal of marketing research, 25(4), 331-341.
• Tellis, G. J., & Weiss, D. L. (1995). Does TV advertising really affect sales? The role of measures, models, and data aggregation. Journal of Advertising, 24(3), 1-12. • Wildner, R., & Modenbach, G. (2015). The long-term ROI of TV advertising in a
digital world. Marketing Intelligence Review, 7(1), 54-60.
• Wood, L. A., & Poltrack, D. F. (2015). Measuring the long-term effects of television advertising: Nielsen-CBS study uses single-source data to reassess the “two-times” multiplier. Journal of Advertising Research, 55(2), 123-131.
