While time series (TS) methods capable of explaining and forecasting behavior have been available for several decades, they have received relatively little attention from marketers. There are a number of reasons for this including the fact that most marketing researchers lacked training in TS methods and there were no user-friendly software packages for them to turn to for help. On top of this, there were very few adequate data sources. Now, however, all that is beginning to change. Given the ever-increasing size of marketing data sets, along with the emergence of Internet data sources and a growing interest in the finance/marketing interface, TS analysis is finding a new role among those who study marketing.

Some Background
While most marketing scientists are well trained in standard econometric and experimental-design techniques, only a few receive formal training in TS analysis. Of the 43 applicants for a tenure-track marketing position at UCLA in fall 1996, only six had taken a graduate course in TS analysis. In addition, TS techniques are data driven, specifically historical data—historically observed patterns are extrapolated to derive forecasts—and as stated above, until quite recently, good data has been hard to come by.

A major reason for the scarcity of historical marketing data has to do with how most firms have viewed data collection. Managers have typically had little incentive to build databases of historical performance and marketing effort for their products and services. Only current and future performance is rewarded, in addition to which, many managers argue that as the market place is constantly changing, historical data are less relevant. At the same time, again until quite recently, assembling a dataset of historical spending, pricing, and performance used to require the retrieval of old accounting records, which, being designed for other purposes, did not always make for the best data.

The fact is, there was no single substantive marketing area where TS modeling was adopted as the logical research tool.

The Present
In recent years, many of these inhibiting factors have begun to disappear. Indeed, the 2nd editions of two important marketing-modeling textbooks, Hanssens et al. (2001), and Leeflang et al. (2000), devote considerably more attention to TS techniques than their first edition. Similarly, the year 2000 marketing faculty applicant pool at UCLA that was trained in TS analysis rose to 21%. Also, unlike years ago, there are now several user-friendly software packages on the market (packages such as EViews, Forecast Pro, and SCA) that significantly facilitate the implementation of TS techniques.

As for the theoretical character of the approach, two recent developments, cointegration analysis and VARX models, offer more confirmatory potential. Cointegration analysis allows you to test for the existence of theoretically based equilibrium relationships between variables, while structural VARX models are specifically designed to supplement sample-based information with managerial judgment and/or marketing theory.

Though current applications of cointegration analysis are still exploratory, rather than confirmatory, these two developments give TS techniques more credibility. There is also a growing openness in the marketing community toward empirical generalizations derived from the repeated application of data-driven techniques to multiple datasets. Even more important, the advent of new data sources based on the automatic, real-time recording of purchase and consumption transactions has led to many effectiveness and segmentation studies regarding the level of brand choice, purchase quantity, and purchase timing. As these databases have continued to grow, covering longer and longer time spans, it has also become possible to study their longitudinal dimension using TS techniques.

Another important occurrence has been the development of techniques specifically designed to disentangle short-run from long-run movements: i.e. unit-root tests, cointegration and error-correction modeling, and persistence estimation, as explored in Dekimpe and Hanssens (1995, 1999). These have provided a natural match between TS analysis and one of marketing’s major areas of interest: quantifying the long-run impact of tactical and strategic marketing decisions.

The Future
Various recent developments are expected to result in an ever growing use of TS methods by marketers. These developments include the expansion of marketing data sets along several dimensions (more variables, longer time spans, a finer temporal aggregation level and a more detailed level of entity aggregation), the increasing rate of change in many market environments, a growing interest in exploring the finance/marketing interface and, last but not least, the emergence of Internet data sources.

More Variables
Traditionally, TS modelers were able to analyze the over-time impact of one or two marketing-mix variables—often advertising and/or price—and competitive information was, at best, available for a limited set of players or for the combined competition. In scanner datasets, price, feature, display and other information is available for all players in the market, even at the SKU level.

In terms of variables, information sets may not only be extended along the usual dimensions, such as more detailed marketing-mix control variables for a larger and more detailed set of competitors. Another potentially interesting data development is the emergence of TS of attitudinal variables that can be matched with transactional observations. In the past, attitudinal data such as awareness and preference were rarely collected, and it was impossible to use the power of modern TS techniques on such variables. Some industries and specific companies have now begun to build tracking databases of attitudes and preferences. For example, many banks now use teller, telephone, or computer transaction occasions as opportunities to gauge key customer attitudes such as satisfaction. Hanssens (1998) used a TS of inferred consumer preference from conjoint measurement to estimate the elasticity of sales with respect to customer-defined product quality. Even though the collection of prolonged, equally spaced attitudinal data may pose some additional challenges in terms of, among other things, costs, panel attrition, and error structures inherent in those kinds of data, these expanded behavioral/attitudinal databases are sure to raise new and important research questions such as: Are there long-term sales consequences of short-lived customer dissatisfaction? Are there threshold values of awareness that are associated with long-term sales growth? Is there an equilibrium between customer attitudes and sales performance?

An Increased Length of the Time Span
As marketing data capture becomes automated and virtually costless, longer TS will become available for research, offering substantial opportunities. First, as samples improve, so does the reliability and precision of our insights. Second, as we gain a better understanding of long-term—and perhaps irreversible—impacts, we can make more powerful strategic inferences. Third, we can better explore changes in regimes, be it in the business environment itself or in the role of marketing. The latter issue is especially important to reconcile statistical and managerial considerations when trying to make inferences about marketing’s long-run effectiveness.

Moving-window regression and VAR models, as well as time-varying parameter models, are ideally suited to exploit the advantages of longer time spans. With the estimation technology firmly in place, the potential for new marketing knowledge generated from long time-span studies is considerable. As an example, researchers may want to revisit product life cycle and diffusion-of-innovation theory with these powerful methods. Do the traditional definitions of introduction, growth, maturity, and decline pass some rigorous TS tests on evolution vs. stability? Or do we need new stage definitions that are more in line with the changing role of word-of-mouth, marketing, and competition over time?

A Finer Level of Temporal Aggregation
Not only will databases cover longer periods of time, they will also incorporate finer time grids. While previous studies often used quarterly or monthly data, recent trends point toward weekly, daily, and even hourly observations. These finer time grids allow for accurate estimation of post-promotional dips and competitive reaction patterns, phenomena that are harder to detect with temporally aggregated data. One example in this respect is the recent transfer function analysis of the hourly effects of direct-response television advertising, the first study to identify peaks and declines in advertising effects within the same day (Tellis et al., 2000). Such databases offer unique opportunities for the study of fast, tactical adjustments in the marketing-mix, and can be expected to have a major impact on the practice of ad copy development and replacement, and price promotion management, especially in the direct marketing arena. At the same time, it will allow marketers to learn to what extent superiority in tactical execution creates long-term strategic advantages for marketers.

From an estimation perspective, one can of course apply the different techniques to these micro-grid databases, but care must be exercised to ensure a “logical correspondence” between the frequency of sampling and the rate of change in the phenomenon at hand. For instance, to assess whether the diaper market is stable or evolving, sampling on an hourly basis does not increase the reliability of unit-root and cointegration tests, nor does it offer additional managerial insights. To analyze eye-movement data, on the other hand, finer sampling is essential for studying the intricacies of the dynamic inter-relationships.

A More Detailed Level of Entity Aggregation
In addition to finer time grids, data are now available at more disaggregate levels, e.g. the individual consumer level and the SKU and/or store level. TS studies have traditionally worked with more aggregated variables, the appropriateness of which can be questioned given recent evidence that it may lead to biased estimates in case of heterogeneity in consumer preferences, and/or marketing-mix activity across stores. This evidence, however, has mostly focused on short-run response parameters, and more research is needed to assess the sensitivity of long-run inferences to aggregation biases.

Entity disaggregation may reflect a movement toward more analyses at the individual store or SKU level, in which case traditional aggregate TS techniques can still be applied, or at the individual consumer household level, where individual choice models seem more appropriate. Indeed, weekly data on the purchases of an individual household may predominantly contain zeros, interrupted by an occasional positive spike, making them less appealing to TS analysts.

While this evolution may not seem to bode well to TS modelers, it is interesting to note that recent developments in the individual choice literature try to integrate the dynamic flexibility of TS techniques into multinomial logit and probit models. Similarly, the application of structural equation models to longitudinal data may also involve the integration of TS concepts. In combination, these different data developments will create great opportunities for applied TS modelers and users.

Rate of Change in the Market Environment
In the past, product markets evolved relatively slowly—30 to 40 years elapsed between introduction and maturity for major household appliances launched in the 1930s. Marketing-mix models were therefore built on time samples that constituted only a fraction of the evolutionary cycle, and often appeared stationary. Indeed, most of the scholarly knowledge base of market response has been developed on databases that were assumed to be stationary. As a result, we know a great deal about short-term marketing-mix elasticities and their determinants, but relatively little about how the marketing-mix functions in a nonstationary trending environment.

The new reality of marketing decisions operates much more in rapidly changing and turbulent environments. Even relatively short time windows—say 1–2 years of weekly data—are now sufficient to capture most of the relevant evolution of a product, and the assumption of stationarity in the data will increasingly become untenable. An empirically tested knowledge base can and should develop that accounts for marketing’s role over the entire life of a product.

One area of study that will be significantly affected is competitive reaction. A recent meta-analysis on 442 product categories established that the predominant form of competitive reaction to price promotions is no reaction (Steenkamp, Nijs, Hanssens & Dekimpe 2001). Can such a finding, based on data from the 20th Century, be expected to hold in the high-information context of the 21st Century? Our casual observation of the competition between book e-tailers Amazon and Barnes and Noble suggests otherwise. For example, when Amazon recently announced a substantial discount on its best-seller books, Barnes and Noble matched the discount within hours. Similar fast reactions are now routinely observed in the airline industry. While game theory has developed the principles of optimal strategic behavior with competition, its empirical knowledge base is small, and much is likely to be learned by empirically examining game-theoretic principles in the context of fast-changing markets.

The Marketing–Finance Relationship
The goals of marketing have traditionally been formulated from a customer perspective. There is a growing interest, however, in expanding this view to include the investor or shareholder perspective in such a way that the separated disciplines of marketing and finance may serve a common purpose. The finance discipline, with its inherent focus on growth and the valuation of assets over time, has long embraced TS analytic techniques. However, the finance literature has only rarely examined how marketing actions affect the valuation of the firm, the exception being some work on the effects of new-product announcements and brand equity on stock prices. Recent work has established that marketing investments in advertising, promotions and product improvement can have a long-term impact on sales, and therefore cash flows and profits.

On the other hand, such marketing investments are costly and may have a negative impact on short-term profit streams. Thus, the important question arises to what extent the investor community either rewards or punishes firms for engaging in these marketing activities. Given that firm valuation data are becoming widely available over long TS, that question may be addressed by expanding the scope of TS models with a firm valuation measure.

The Emergence of Internet Data Sources
The emergence of Internet data sources is likely to make the developments described above even more prominent. Indeed, they allow one to easily integrate transactional and attitudinal data, and to quickly create high-frequency data at the individual consumer level. In much the same vain as we can currently track weekly sales performance, we will be able to track weekly, or even daily changes in consumer interest (e.g. number of Web site visits, shopping intensity derived from visitation patterns, and first-time and repeat purchases), not only at the aggregate level, but also at the individual consumer level. These data can be supplemented with frequent Web enabled market research studies on customer satisfaction, usage rates, and future purchase intentions, with the resulting databases providing an excellent laboratory for experimentation. TS models can be used to infer the long-term effect of an awareness or preference increase on the revenue and profitability of a brand. Internet data will also enable a more detailed understanding of the dynamics of consumer choice, as they will quickly generate sufficient degrees of freedom to integrate TS techniques with existing logit/probit specifications.

The internet will not only create vaster and more detailed datasets, it is also expected to cause a faster diffusion of new ideas and more turbulent competitive environments. Moreover, it will likely create a radical break in the way business is conducted, both in terms of the channel relationships with suppliers and intermediaries, and in terms of interactions with customers. From a TS perspective, these abrupt and radical changes pose significant challenges. Indeed, the constant-parameter assumption in many persistence-based models may no longer be appropriate to test the long-run implications of Internet-related decisions. Instead, structural-break analyses will likely be needed to determine whether, for instance, the introduction of a dual, Internet-based distribution channel elevated stationary category sales to a new and higher level, cannibalized existing revenue streams, and/or affected prevailing competitive reaction patterns.

Still, given recent developments, it is almost certain that the role of TS in marketing will only continue to grow.

References
Dekimpe, M.G., Hanssens, D.M., 1995. The Persistence of Marketing Effects on Sales. Marketing Science 14 (1), 1-21.

Dekimpe, M.G., Hanssens, D.M., 1999. Sustained spending and persistent response: a new look at long-term marketing profitability. Journal of Marketing Research 36, 397–412.

Hanssens, D.M., 1998. Order forecasts, retail sales, and the marketing mix for consumer durables. Journal of Forecasting 17, 327–346.

Hanssens, D.M., Parsons, L.J., Schultz, R.L. 2001. Market Response Models. Boston, MA: Kluwer Academic Publishers, 2nd edition.

Leeflang, P.S.H., Wittink, D.R., Wedel, M., Naert, P., 2000. Building Models for Marketing Decisions. Kluwer Academic Publishing, The Netherlands.

Steenkamp, J.-B., Nijs, V.R., Hanssens, D.M., Dekimpe, M.G. 2001. Competitive Reactions and the Cross-Sales Effects of Advertising and Promotion. UCLA Center for Marketing Studies, Working Paper No. 363, September.

Tellis, G.J., Chandy, R.K., Thaivanich, P., 2000. Which ads work, when, where, and how often? Modeling the effects of direct television advertising. Journal of Marketing Research 37, 32–46.

----------

This article was based on the original paper, "Time-Series Models in Marketing: Past, Present and Future" by M.G. Dekimpe and D.M. Hanssens, which appeared in the September 2000 issue of the International Journal of Research in Marketing.

More Features