Loading field map
Marketing science emerged from a persistent tension: how to build formal, quantitative models of markets and consumer behavior without sacrificing the explanatory depth that makes those models useful for managers and theorists. The subfield is defined by its commitment to mathematical and statistical methods—drawn from economics, operations research, and statistics—to describe, predict, and guide marketing decisions. Its history is a series of debates about what kind of model is most trustworthy: one that fits historical data well, one that recovers causal relationships, one that captures deep behavioral primitives, or one that maximizes predictive accuracy on new data.
The earliest systematic quantitative work in marketing, from the 1950s through the 1970s, took the form of market response models. These were regression-based frameworks that linked marketing inputs—advertising spending, price changes, sales force effort—to output measures such as sales or market share. The core assumption was that the relationship between a marketing action and a market outcome could be estimated from historical data and then used to optimize future spending. Market response models gave managers a practical tool for budget allocation, but they treated the consumer as a black box. They could say that a 10 percent price cut raised sales by 5 percent, but they could not explain why. This limitation set the stage for a wave of frameworks that tried to open that black box.
In the late 1960s, two major consumer behavior models appeared almost simultaneously, each attempting to model the mental processes that intervene between a marketing stimulus and a purchase decision. The Howard-Sheth Model (1969) was built around the idea of dynamic learning: consumers move through stages of attention, comprehension, attitude, and intention, and their decision rules change as they gain experience with a product category. The model emphasized feedback loops—a purchase updated the consumer's confidence and brand knowledge, which then altered future processing. The Engel-Kollat-Blackwell (EKB) Model (1968) took a different approach. It laid out a detailed stage-based information-processing sequence: problem recognition, search, alternative evaluation, choice, and post-purchase outcomes. Where Howard-Sheth stressed the evolution of decision rules over time, EKB provided a static but comprehensive map of the cognitive steps in a single decision episode.
These two frameworks were not direct competitors in the sense that one replaced the other. They coexisted as complementary research programs, each inspiring a distinct empirical tradition. Howard-Sheth led to studies of learning curves and brand loyalty dynamics; EKB became the template for survey-based studies of consumer decision stages. Both, however, shared a common difficulty: they were too rich to be fully estimated as formal models. Their many constructs—perceptual bias, search propensity, evaluative criteria—were hard to measure reliably, and the models offered little guidance on how to test the whole system against data. This gap between theoretical ambition and empirical tractability created pressure for a different kind of modeling.
While consumer behavior modelers were wrestling with individual-level complexity, Frank Bass in 1969 proposed a model that deliberately ignored individual psychology and focused on aggregate adoption patterns. The Bass Diffusion Model described how a new product spreads through a population using just two parameters: the coefficient of innovation (external influence, such as advertising) and the coefficient of imitation (internal influence, such as word of mouth). The model was parsimonious, fit historical adoption curves remarkably well, and could be estimated with ordinary least squares on sales data. It did not replace the Howard-Sheth or EKB frameworks—those continued as research traditions—but it narrowed the ambition of marketing science in a productive way. Instead of modeling every mental step, the Bass model showed that a simple aggregate law could forecast adoption timing and peak sales with enough accuracy to guide launch decisions. The framework remains active today, extended with covariates for price, marketing mix, and competitive effects, and it is still used by firms to forecast new product trajectories.
At roughly the same time, a parallel methodological school was developing around stochastic choice models. These frameworks treated consumer choice as inherently probabilistic: even under identical conditions, a consumer might choose brand A today and brand B tomorrow. Rather than trying to explain each choice deterministically, stochastic models estimated the probability distribution over alternatives as a function of brand loyalty, purchase event feedback, and heterogeneity across consumers. The most influential early example was the Dirichlet model of repeat purchasing, which showed that brand choice shares in stationary markets could be predicted from penetration and purchase frequency alone. Stochastic choice models coexisted with the Bass model—both were aggregate-level, data-driven approaches—but they addressed a different question: not how a new product diffuses, but how established brands compete in mature categories. This school remains active, now often integrated with Bayesian estimation methods.
By the 1980s, marketing scientists had grown uneasy with the limitations of simple regression-based market response models. Those models assumed that marketing variables were exogenous—that advertising budgets and prices were set independently of demand shocks. In reality, managers raise advertising when they expect strong sales, and they cut prices when inventory piles up. This reverse causality meant that ordinary least squares estimates were biased. Econometric modeling entered marketing science as a methodological school that prioritized causal identification. Researchers imported techniques from labor econometrics: instrumental variables to correct for endogeneity, panel data methods to control for unobserved heterogeneity, and time-series models to handle dynamic feedback. The shift was not a rejection of market response models but an absorption of them into a more rigorous framework. The same regression equations were now estimated with instruments—for example, using competitor prices as instruments for a firm's own price—so that the coefficients could be interpreted causally. Econometric modeling remains a core toolkit in marketing science, especially for measuring the sales impact of advertising, promotions, and distribution changes.
If econometric modeling corrected the bias in reduced-form estimates, structural modeling (emerging around 1995) went further by asking what deep behavioral primitives generated the observed data. Structural models specify a utility function for consumers and a profit function for firms, then solve for the equilibrium that arises when both sides optimize. The goal is to recover parameters—such as price sensitivity or brand preference—that are invariant to policy changes, so that the model can predict outcomes under counterfactual scenarios that have never been observed. For example, a structural model of the automobile market can estimate consumer willingness to pay for fuel economy and then simulate what would happen if a carbon tax were imposed. This is something reduced-form econometric models cannot do, because their parameters are conditional on the existing pricing regime. Structural modeling did not replace econometric modeling; the two coexist, with structural methods used when the research question demands counterfactual prediction and reduced-form methods used when the goal is to estimate a specific causal effect with minimal assumptions. The tension between them—primitives versus flexibility—remains one of the subfield's liveliest debates.
The most recent methodological school, machine learning in marketing (accelerating around 2010), introduced a fundamentally different philosophy. Where structural and econometric models are theory-driven—they impose a functional form derived from economic assumptions—machine learning models are data-driven: they let flexible algorithms (random forests, gradient boosting, neural networks) discover patterns in high-dimensional data without pre-specified equations. Machine learning excels at prediction tasks: targeting individual customers for promotions, personalizing website content, forecasting churn. It has absorbed some of the stochastic choice tradition (probabilistic predictions) and some of the econometric tradition (causal forest methods for heterogeneous treatment effects), but it also challenges both. The challenge to structural modeling is that machine learning often predicts better without any economic theory; the challenge to econometric modeling is that machine learning's black-box predictions are hard to interpret causally. The school is not a replacement for earlier frameworks but a new infrastructure that has expanded the subfield's empirical reach while reopening old questions about the role of theory.
Today, marketing science is a pluralistic field. The leading active frameworks—structural modeling, machine learning, stochastic choice models, and econometric modeling—each occupy a distinct niche. Structural modeling is the method of choice when the research question requires counterfactual policy simulation, such as merger analysis or the evaluation of a new pricing scheme. Machine learning dominates when the task is prediction at the individual level with large datasets, such as real-time bidding in digital advertising. Stochastic choice models remain the standard for understanding brand competition in frequently purchased categories. Econometric modeling is the default for estimating causal effects in field experiments and quasi-experimental settings.
What the leading frameworks agree on is that validation matters: a model should be tested on hold-out data, its assumptions should be stated transparently, and its predictions should be benchmarked against simpler alternatives. What they disagree on is how much theory a model needs. Structural modelers argue that without economic primitives, predictions are unreliable when the environment changes. Machine learning advocates counter that in high-dimensional settings, theory imposes false constraints and reduces accuracy. Econometricians occupy a middle ground, using theory to guide instrument selection but letting the data speak about effect sizes. This disagreement is not a sign of weakness; it is the engine that drives the subfield forward, forcing each school to sharpen its methods and clarify what it can and cannot do.
Marketing science has never settled on a single definition of what a good model is. That unresolved question—formal precision versus explanatory depth, causal identification versus predictive accuracy—is exactly what gives the subfield its coherence and its momentum.