The MECOVMA Framework: Implementing Machine Learning Under Macroeconomic Volatility for Marketing Predictions

Abstract

The methodological framework introduced in this paper, MECOVMA, is a novel framework that guides the application of Machine Learning specifically for marketing predictions within volatile macroeconomic environments. MECOVMA has been developed in response to the identified gaps displayed by existing frameworks—when it comes to consolidation, relevance, interdisciplinarity, and individuality—and in light of the polycrises occurring in the current decade. The methodology to develop MECOVMA comprises three phases: firstly, synthesizing existing frameworks based on their thematic relevance to select MECOVMA’s process steps; secondly, integrating the evidence provided by a systematic literature review to design the content of these process steps; and thirdly, using an expert evaluation, structured through a qualitative content analysis, to validate MECOVMA’s applicability. This leads to the final framework with four overarching P_MECOVMA process steps, guiding the Machine Learning application process in this context with specific tasks. These include, for example, the processing of multidimensional data inputs, complexity reduction in a dynamic environment, and training methods adapted to particular macro-conditions. In addition, features are provided on how Machine Learning can be put into marketing practice, incorporating both narrower statistical- and broader business-oriented evaluations, and iterative feedback loops to mitigate limitations.

1. Introduction

Over the last decade, Machine Learning (ML) has gained significant traction in marketing practice [1,2,3]. Since then, ML has opened up new opportunities to gain deeper insights into demand behavior and enhance predictive marketing analytics [4,5]. The resulting interest in ML, which can extend or even surpass traditional econometric models in different cases [6], implies an obligation for the marketing discipline to further engage with its effective implementation [7]. Consequently, Herhausen et al. [8] (p. 1) conclude, “Companies increasingly use algorithms to generate predictions for marketing decisions… not surprisingly, Machine Learning [ML] is a trending topic for marketing researchers and practitioners”.

Especially since the beginning of this decade, the marketing environment has been subject to strong fluctuations and a series of disruptive events, leading to an overall volatile period [5,9]. In the early 2020s, the pandemic and subsequent measures affected economic development [10,11], and “in response, customers worldwide have radically changed their purchasing behavior” [12] (p. 1). The geopolitical and economic landscape is characterized by significant challenges, including the beginning of the conflict in the Middle East toward the end of 2023, the escalation of the Russian–Ukrainian war in 2022 [13,14], and tensions between the United States and China [15]. Macroeconomically, these phenomena are accompanied by dynamic inflation rates, recession expectations, and the resulting monetary policy reactions of the Federal Reserve, which in turn, can provoke volatile demand responses [16,17,18]. Basdekis et al. [19] (p. 2) argue, “What is certain is that, currently, we live in a vulnerable period, which may be maintained for the coming years, forcing all the competent bodies to take the necessary measures”.

To cope with this volatile marketing environment and sustain the accuracy of predictions on future developments, ML can serve as an important tool: “The existing complexity and uncertainty of business processes open the space for ML”, as stated by Reis et al. [20] (p. 240). However, the implementation of ML prediction practices, especially under these demanding conditions, is associated with considerable challenges and necessitates a suitable methodological approach [21,22]. MacKay et al. [23], for example, point out that—despite a sharp increase in the number of advanced ML algorithms—there is a noticeable gap between their theoretical potential and their successful application. This gap is echoed by Van Giffen et al. [7], who criticize an often unclear and non-transparent approach to the practical application of ML algorithms. In addition, Herhausen et al. [8] (p. 1) emphasize that “marketing decision-makers today often struggle to adequately capture and transform (big) customer data into meaningful insights”—a difficulty that also reflects the complexity of consumer decision-making behavior [24,25]. Available studies indicate that the use of ML requires a specific approach depending on the area of application and conditions [26,27]. Currently, a specific methodological framework adapted for this highly pertinent application context is lacking, as, for example, multiple recent approaches tend to rely on guidelines for ML deployment in broader contexts [8,28].

This paper, therefore, presents a new methodological framework developed to implement ML in this particular context, termed MECOVMA (Machine learning in MacroECOnomic Volatility for MArketing Predictions). The MECOVMA framework contributes to methodological progress in this area by providing a scientific answer to the following four identified gaps in the research landscape: (i) consolidation, as various individual approaches have already been developed to address the challenges in this thematic area, which, however, usually serve as isolated solutions to specific problems [29,30,31,32,33,34]; (ii) relevance, as the existing research body is not sufficiently aligned with environmental conditions, which “potentially facilitates theoretical closure and taken-for-granted ignorance to novel developments in practice” [35] (p. 221) and highlights the urgency for further investigation; (iii) interdisciplinarity, as ML terminologies and criteria are usually not fully in line with the needs of the marketing discipline and thus fail to promote a holistic perspective [7,36]; and (iv) individuality, as generic and non-customized frameworks do not meet the specific requirements of ML marketing prediction—especially considering that “the selection of the correct methodology from the vast number available, based on the context and many other aspects, is critical” [37] (p. 1086).

This paper is designed as a research study, resulting in a robust methodological framework synthesized from recent academic insights [38]. A methodological framework is understood here in line with the widely accepted definition [39,40], which describes it as “structured practical guidance or a tool to guide the user through a process, using stages or a step-by-step approach” [41] (p. 2). The design of this study focuses primarily on a conceptual synthesis and is formally structured as a three-phase methodology. This approach broadly follows the phases of McMeekin et al. [41] for developing a methodological framework, which is used in various scientific analyses [42,43,44]. As a result, the present paper is organized in the following manner. An introduction (Section 1) is provided before a theory overview (Section 2) of methodological frameworks in general and the foundations of ML in marketing. Beginning with the three-phase methodology, in Methodology Phase I (Section 3.1), the overall process steps for the MECOVMA framework are selected by means of synthesizing data and evaluating the relevance of existing frameworks. Then, Methodology Phase II (Section 3.2) involves the content design needed to substantiate the previously defined process steps based on a systematic literature review. The last phase of the three-phase methodology, Methodology Phase III (Section 3.3), contains an expert discussion at an international symposium to validate the framework, while its applicability is assessed through a qualitative analysis of the content. Subsequently, a further discussion (Section 4) and limitations (Section 5) are presented. Finally, a conclusion (Section 6) completes this study, which aims to create an amalgamation that provides methodological support in using ML for marketing prediction in the recent macroeconomically volatile times.

2. Theoretical Overview

2.1. Fundamentals of Methodological Frameworks

A methodological framework typically offers researchers the foundational building blocks that guide the approach and direction of their study or project [35]. According to Corallo et al. [39], there are terms that are often used interchangeably—despite having different scientific meanings—such as “theoretical framework”, “conceptual framework”, “process improvement framework”, “methodology”, “model”, “method”, or “system”. Recognizing this semantic variability and the fact that, from a terminological standpoint, a standardized and universally accepted definition that delineates the scope of a methodological framework does not exist, McMeekin et al. [41] (p. 2) provide a working definition. There, they describe a methodological framework as “a tool to guide the developer through a sequence of steps to complete a procedure”. The authors refer to a broad spectrum of existing attempts to define it, such as Cruz Rivera et al. [45], who understand a methodological framework as a set of specific procedures, rules, postulates, and methods. They also reference older sources, such as Andrade et al. [46], who consider the purpose of a methodological framework to be the explicit structuring of the execution of specific tasks.

The purpose of a methodological framework, as described by Cruz Rivera et al. [45], is to reduce the likelihood of ineffective research efforts while increasing the potential for impactful research results. Rweyendela et al. [43] affirm that it supports a testable approach and thus minimizes subjectivity and user interpretation. Also articulated is a need for a methodological framework to provide structure and guidance for researchers and practitioners while still maintaining certain conceptual openness [35]. The resulting structure is intended to help maintain important principles for research quality, such as limiting the subjective leeway of the researcher (researcher bias), increasing the replicability of an experiment or study, and strengthening the reasonableness of the results [47]. Besides the benefits, methodological frameworks can also entail certain challenges. The potential complexity of frameworks can make practical implementation difficult, and in practice, there can be a “tension between frameworks that aim to capture complexity and those that aim to simplify core principles” [48] (p. 513). Other possible difficulties include the lack of clarity or intuitiveness in the specific application of frameworks, the fact that they may reflect the attitudes of their creators, a potential absence of articulation of their benefits compared with other frameworks, and the difficulty of objectively evaluating frameworks, which can make their selection challenging [48].

Although no generally established development structure emerges from the literature, McMeekin et al. [41] (p. 1) manage to find a certain consensus in the research with their approach consisting of “(a) identifying data to inform the methodological framework; (b) developing the methodological framework; and (c) validating, testing and refining the methodological framework”. Concerning the data identification to inform the methodological framework, particularly the literature search and the derivation of existing methods, these receive broader scientific recognition and have been effectively employed in several recent studies [37,42,43,49]. In the development phase, the specific framework is then elaborated. A starting point can be the development of overarching building blocks, meaning the general procedural steps in the process [43]. Subsequently, these building blocks can be successively substantiated by specifying them in detail with important theoretical and practical content, for example, from a literature review [37,43,50]. Validation can be performed in different ways. Schlumberger et al. [49] recommend that the step of practical validation, especially owing to its complexity and scope, can be either fully or partially outsourced from the original conceptual research. In cases where this step is implemented, practical applications such as piloting or incorporating a case study can be considered. Furthermore, the methodological framework and its components can be evaluated and thus validated in a focus group. Conducting written and oral surveys with specific framework-related questions in academic and industrial circles, as well as among affected stakeholders, is also possible. In this way, the validation aims to ensure that the methodological framework’s content meets the requirements of the relevant interest groups [47,51].

2.2. ML for Marketing Predictions

In the last years, ML has established itself as a key element in marketing science and practice, manifested in a diverse range of application scenarios for advanced data processing leveraging ML [8,52]. According to the scientific literature, the increasing uncertainty and complexity of business operations are likely to accelerate the need for more sophisticated applications of ML in marketing [20,21]. Nevertheless, with respect to the practical implementation of ML predictions, MacKay et al. [23] (p. 6) observe that “despite the increasing number of well-performing models, the gap between model success and… implementation has not been closing at the same rate”.

The definition of ML in the scientific literature can be observed as largely consistent in its general understanding. Li et al. [26] (p. 18) provide a terminological overview and explain, “Artificial intelligence [AI], the capability of a machine to imitate intelligent human behaviors, is facilitated by machine learning, a subfield of artificial intelligence [AI], which enables a system to automatically learn and enhance its performance through experience”. Multiple authors express a comparable view of ML as a method of learning from data or previous experiences to optimize the performance of a model [23,26,53,54,55]. This paper focuses on the ML task of prediction. In the context of marketing management, predictions are commonly understood as making statements about future states or developments grounded in existing data information [55,56,57]—or, as Kmiecik [21] (p. 100) asserts, “predicting is the inference of unknown events based on known events”. From the perspective of scientific theory, many ML models use the principle of inductive reasoning, as they learn patterns from data in order to make generalizable predictions. This involves generating generalizable ML models—essentially, empirical regularities—from systematic observations, measurements, or experiences to predict future events. Hence, ML models naturally generate estimates through learned parameters and structures that encapsulate data patterns, with the goal of inferring new, unseen data from them [38,54,58,59]. Predictive phenomena can be characterized through quantitative variables, expressed in quantities, and qualitative variables, described descriptively. The term forecasting is often used in connection with predictions that place an explicit emphasis on temporal relationships with the output [21,60]. In academic discourse, there appears to be a consensus that predictions, for example, about demand and their accuracy, can have substantial impacts and consequences—such as on customer service, inventory and production levels, or operational cost structures [21,61,62]. Consistent with this reasoning, Karl [24] (p. 8) concludes that “precise demand forecasts (“predictions”) play a pivotal role… and ultimately impact the commercial success”.

In this paper, marketing predictions are considered to be predominantly a supervised ML problem. Discussions in the scientific literature tend to differ on which specific learning paradigms ML incorporates. Predominantly mentioned learning paradigms are supervised learning, together with unsupervised learning, as well as often reinforcement learning [23,26]. Beyond them, learning paradigms like active and semi-supervised learning are also mentioned [63]. Under the premise of a supervised learning problem, a marketing prediction function (${\mathrm{f}}_{\mathrm{M}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}}$) focused on in the present study commonly incorporates input variable(s) $\mathrm{x}\in \mathrm{X}$—also called predictors or features [24]—that are related to data output values $\mathrm{y}\in \mathrm{Y}$ (target variable(s)) (${\mathrm{f}}_{\mathrm{M}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}}:\mathrm{X}\to \mathrm{Y})$. In such a process, a model $\mathrm{f}\left(\mathrm{x}\right)$ is generated, resulting in a prediction function influenced by several input variables, for example, in the multivariate case: ${\mathrm{Y}}_{\mathrm{M}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}}={\mathrm{f}}_{\mathrm{M}\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{d}}({\mathrm{X}}_1,{\mathrm{X}}_2,\dots ,{\mathrm{X}}_{\mathrm{n}};\Theta )$. During the training phase, the aim is to learn optimal parameters—represented by $\Theta$ and typically called model weights or coefficients—from the data to achieve a valid fit that enables accurate predictions for new data. Predictions can yield a real-valued result (y ∈ $\mathbb{R}$), such as demand in units, or a categorical outcome (y ∈ {0,1}), for example, purchase or non-purchase [64,65].

In addition to the highly relevant topic of predictive ML, there are further—and partially integrated or related—current ML developments. These include the application of generative methods, such as Large Language Models, which can closely resemble human language texts, or Generative Adversarial Networks to generate realistic images or videos [66,67,68]. Other trending topics include specialized Artificial Neural Network architectures—such as Graph-, Convolutional-, and Recurrent-Neural-Networks for specific types of data processing—as well as ensemble models, which combine multiple models to have more robust results [69,70,71]. Furthermore, Bundi [67] and Zhang et al. [72] highlight federated learning, by which decentralized data sources are used for ML model training to maintain data security.

3. Methodology

The methodology of this paper in developing the MECOVMA framework is broadly guided by McMeekin et al. [41]. It follows an approach with three methodological phases, as graphically depicted in Figure 1. Methodology Phase I is designed for the initial step selection of the framework and is discussed in Section 3.1 of this paper. It includes a summarized tabular overview of existing frameworks, along with an individual evaluation of their relevance based on specific criteria derived from the research question. Through a systematic synthesis of these analytical results, the overall process steps of the MECOVMA framework are defined in terms of their overall number and titles. Building on that, Methodology Phase II of this paper covers the content design of the framework, as discussed in Section 3.2. Drawing on the findings of a systematic literature review [73], the process steps are detailed in terms of their specific tasks and scopes. This leads to the final visualization of the MECOVMA framework. In Methodology Phase III (Section 3.3), the developed framework is then validated through an expert discussion held at an international symposium, and MECVOMA’s applicability is investigated via qualitative content analysis.

3.1. Methodology Phase I: Step Selection

The Methodology Phase I for step selection follows the latest recommendation by Kallingal and Firoz [37] (p. 1106), which proposes that “the evidence to inform the methodological framework… is identified by exploring existing methodologies and identifying new data through literature search”. In this context, the research process in Methodology Phase I is divided into two consecutive elements: first, existing frameworks are collected and subsequently tabulated to analyze the constituent steps (Section 3.1.1). This is essential, as different authors vary the quantity and titles of steps in their frameworks, resulting in no uniformly accepted framework for the present research purpose [8,28,74]. In order to analyze these different approaches, a relevance evaluation of the existing frameworks is implemented by taking into account the thematic priorities of this study’s research question. The evaluated frameworks are subsequently synthesized in the second part of Methodology Phase I (Section 3.1.2) within an integrative research process [43,75]. The outcome is the number and titles of the overarching process steps for the MECOVMA framework.

3.1.1. Framework Step Analysis and Relevance Evaluation

The starting point for the development of a topic-specific framework in the MECOVMA context is based on the recognized scientific practice of building on existing frameworks [37,45,76,77]. As suggested by Rweyendela et al. [43], this study systematically analyzes established scientific databases—specifically Web of Science and Scopus platforms—to identify potential existing frameworks. The literature is subjected to a review utilizing both backward and forward searching techniques, known as “backward and forward literature snowballing” [43] (p. 2228). The initial inclusion criteria for selecting the frameworks are (a) publication in a peer-reviewed journal to ensure general scientific standards, (b) being written in German or English for comprehension purposes, (c) publication in a year ≥2020 due to significant shifts in macroeconomic conditions since then, and (d) demonstrating procedural eligibility, operationalized by a methodological approach or framework that includes transparently specified process steps. Studies are excluded if they lack (e) specificity to the broader MECOVMA context and (f) substantiation, as operationalized by the absence of a formal theoretical, conceptual, or empirical foundation underpinning the framework.

Following this broad filtering with inclusion criteria and exclusion criteria, fine filtering with a thematic relevance evaluation R_i of the frameworks is performed. Here, the frameworks initially selected are evaluated based on their alignment with the three thematic priorities defined for this study. These include (T₁) AI/ML/Statistical Learning based Prediction, (T₂) Marketing/Consumer Behavior, and (T₃) Volatility/Uncertainty/Macroeconomic Environment. This thematic triad is visualized in Figure 2.

The relevance evaluation R_i, in accordance with the defined triad of thematic priorities (T₁–T₃) in Figure 2, is conducted for each selected framework and is presented in

Table 1. Framework Overview: Number of Steps and Thematic Relevance.

Study	Framework (Name)	P	${\mathbf{R}}_{{\mathbf{T}}_1}$ (AI/ML/Statistical Learning Based Prediction)		${\mathbf{R}}_{{\mathbf{T}}_2}$ (Marketing/Consumer Behavior)		${\mathbf{R}}_{{\mathbf{T}}_3}$ (Volatility/Uncertainty/Macroeconomic Environment)		${\sum }_{\mathbf{i}=1}^6{\mathbf{R}}_{\mathbf{i}}$
			Study	Frw.	Study	Frw.	Study	Frw.
Tsao et al. [34]	Forecasting Framework for Value of ML and External Information Index	3	✓	✓	✓	✗	✓	✓	5/6
Ma and Sun [63]	AI-Driven Marketing Landscape	3	✓	✓	✓	✓	✓	✗	5/6
Efat et al. [78]	Forecasting Pipeline	4	✓	✓	✓	✗	✓	✗	4/6
Gharibshah and Zhu [79]	User Response Prediction Workflow	3	✓	✓	✓	✓	✗	✗	4/6
Herhausen et al. [8]	Simplified Framework for ML in Marketing	4	✓	✓	✓	✓	✗	✗	4/6
Kharfan et al. [32]	Methodology	3	✓	✓	✓	✗	✓	✗	4/6
Raizada and Saini [74]	Methodology for Sales Forecasting	8	✓	✓	✓	✗	✓	✗	4/6
Ngai and Wu [80]	Conceptual Framework for ML Application in Marketing	7	✓	✓	✓	✓	✗	✗	4/6
Jiang et al. [81]	Machine Learning Integrated Portfolio Rebalance Framework	2	✓	✓	✗	✗	✓	✓	4/6
Van Giffen et al. [7]	Conceptual Approach: Using Machine Learning for Marketing Problems	4	✓	✓	✓	✓	✗	✗	4/6
Potrawa and Tetereva [33]	Workflow Diagram	5	✓	✓	✓	✗	✓	✗	4/6
Brackmann et al. [28]	ML Framework	6	✓	✓	✓	✗	✗	✗	3/6
Geiler et al. [31]	ML Pipeline for Churn Prediction	4	✓	✓	✓	✗	✗	✗	3/6
Jafarzadeh et al. [4]	CRISP-DM Framework	6	✓	✓	✓	✗	✗	✗	3/6
Esmeli et al. [30]	Framework for Purchase Intention Prediction	3	✓	✗	✓	✓	✗	✗	3/6

. R_i is assigned to each evaluation category, rated using a standardized score that adopts a Boolean value, which intends to reduce subjectivity in the evaluation process. This Boolean value indicates whether the area of the thematic priority is addressed (R_i = 1, Yes = ✓) or not (R_i = 0, No = ✗). This evaluation includes both incorporating these priorities into this study’s framework and addressing them within the paper itself. The consideration of the reflection of the thematic priorities in the paper is based on the assumption that, if the paper provides the appropriate thematic contextualization, these aspects will also be implicitly recognized in the framework. Three thematic priorities (T₁–T₃), examined in both the study and framework, result in six evaluation categories. A framework is included in the further course of the paper if it is considered “thematically relevant” to the research question, which is operationalized by using the threshold value ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}\ge 3$, thereby requiring confirmation in three or more of the six relevance evaluation categories. As a result of this entire filtering process, the results presented in Table 1 are the final remaining frameworks. There are listed the corresponding studies with the author and year, along with the framework name, identified from the original text using terms that capture its key elements. Furthermore, the number of process steps (P), the individual evaluation for the thematic priorities (T₁–T₃), and the resulting overall relevance evaluation ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}$ are detailed.

There are wide variations in the specific structures of the frameworks indicated in Table 1. While Tsao et al. [34], for example, set out a three-stage framework in which external information sources are combined with ML, the concept put forward by Ma and Sun [63] proposes a marketing strategy that is driven by AI. The highest overall relevance evaluation of ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}=5$ is accorded to these two sources in terms of thematic relevance. This is closely followed by Efat et al. [78], who achieved a still high overall relevance evaluation of ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}=4$ and propose a forecasting pipeline using a hybrid modeling approach for sales predictions, incorporating factors such as marketing effects. Gharibshah and Zhu [79] matched this evaluation with their proposal of a workflow that uses ML to predict how customers will respond. Furthermore, Herhausen et al. [8] focus on applying ML specifically to the field of marketing within a straightforward four-step framework, whereas Brackmann et al. [28] set out a more conventional ML framework that is designed for wider and more general retail applications. The framework proposed by Kharfan et al. [32] is targeted toward their case study concerning the prediction of demand for seasonal products newly launched on the market. Further approaches include Geiler et al. [31], who aim to predict customer churn rates, Potrawa and Tetereva [33], who develop a general framework for an ML pricing model, and Van Giffen et al. [7], who demonstrate a typical application logic when using ML within marketing. The conceptual framework presented by Ngai and Wu [80] allows data and models to be considered individually. The methodology set out by Raizada and Saini [74] is an approach for sales forecasting, while Jiang et al. [81] focus on developing portfolio management aided by an ML framework. In addition, the authors Jafarzadeh et al. [4] directly apply the Cross-Industry Standard Process in Data Mining (CRISP-DM)—an established data science framework [82]—to their research projects, which also forms the basis for ML application in the context of small- and medium-sized enterprises. Esmeli et al. [30] propose a framework that focuses on early purchase prediction, with a specific emphasis on the e-commerce domain rather than marketing contexts. This results in a subordinate overall relevance evaluation of ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}=3$, at the lower end of Table 1, which is further supported by the fact that contextual and loyalty features are considered but not macro-environmental conditions.

A total of N = 15 frameworks meet both the defined threshold for the overall relevance evaluation (${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}\ge 3$) and the initial inclusion and exclusion criteria. The analysis of the number of overarching process steps demonstrates a range of 2 ≤ P ≤ 8, with P_max = 8 achieving the maximum value as the only framework. Figure 3 shows the distribution of P within the ultimately included frameworks as a histogram, which indicates that the majority of them tend to position their process steps predominantly in the lower single-digit range. The analysis identifies an average number of process steps of ${\mathrm{M}}_{\mathrm{P}}\approx 4.33$ and a median of ${\mathrm{M}\mathrm{d}\mathrm{n}}_{\mathrm{P}}\approx 4.00$. With this analysis as the basis, a recommended benchmark for the MECOVMA framework can be derived, suggesting a value of P_MECOVMA = 4 overarching process steps.

3.1.2. Synthesis of the Process Steps

After establishing the number of process steps in the MECOVMA framework, the titles of the P_MECOVMA = 4 steps are to be determined. For this purpose, further synthesis is conducted by listing in

Table 2. A Framework Overview: The Titling of the Steps.

Study	P (Names)	${\sum }_{\mathbf{i}=1}^6{\mathbf{R}}_{\mathbf{i}}$
Tsao et al. [34]	(1) Operational Data (Internal Data, Extremal Information Index), (2) Forecasting Framework (Company’s Model/Time Series Forecasting, Machine Learning Model/Baseline Forecasting, Clustering/Classification/Multiple Regression/Intelligent Forecasting), (3) Value of Machine Learning Model, Value of External Information Index	5/6
Ma and Sun [63]	(1) AI and Machine Learning, (2) Marketing Actions (Search, Marketing Mix, Recommendation, Engagement, Attribution), (3) Industry Trends (Interactive and Media-Rich, Personalization, Real-Time Automation, Customer-Journey Focus)	5/6
Efat et al. [78]	(1) Data Model, (2) Feature Extraction (Characteristic based Clustering, Convolutional Neural Network/Gated Recurrent Unit, 1-D Feature Array), (3) Feature Composition, (4) Prediction (Long Short-Term Memory)	4/6
Gharibshah and Zhu [79]	(1) Data (Raw Data, Labeling, Preprocessing), (2) Learning (Method, Train), (3) Output (Prediction, Ordered List)	4/6
Herhausen et al. [8]	(1) Relevant Marketing Phenomena Data, (2) Preprocessing, Annotation, and Feature Extraction, (3) Train and Validate ML Model, (4) Classification and Prediction Inform Marketing Decision	4/6
Kharfan et al. [32]	(1) Data Pre-processing (Self-through (Point of Sale) Data, Data Filtering/Data Aggregation, Feature Engineering/Feature Selection), (2) Three-Step Model (Cluster, Classification, Prediction), (3) Model Results (Comparing the Forecasting Accuracy and Bias)	4/6
Raizada and Saini [74]	(1) Dataset, (2) Data Cleaning and Preparation, (3) Splitting the Data (Training Data, Testing Data), (4) Pre-Processing, (5) Feature Selection/Feature Measurement, (6) Learning/Regression Technique, (7) Applied (Linear Regression/Random Forecast Regression/K-Nearest Neighbors Regression/Support Vector Regression Technique/Extra Tree Regression), (8) Prediction of Sales	4/6
Ngai and Wu [80]	(1) Supervised/Unsupervised/Reinforcement Learning, (2) ML Algorithm, (3) Text Analysis/Voice Analytics/Image and Video Analytics, (4) ML Tools, (5) ML Technologies, (6) Product, Promotion, People, Price, Place, Process, Physical Evidence, (7) ML Application in Marketing	4/6
Jiang et al. [81]	(1) Predictive Modelling (Historical Market Index Data, Technical Indicators, Machine Learning Models for Market Movement Prediction, Risk-Aversion Coefficient Adjustment), (2) Portfolio Optimization Modeling (Historical Asset/Stock Data, Mean-Risk Portfolio Optimization Model, Portfolio Weights Allocation)	4/6
Van Giffen et al. [7]	(1) Relevant Population, (2) (Generates) Data, (3) (Trains) Machine Learning Model, (4) Predictions Trigger Marketing Decisions and Actions	4/6
Potrawa and Tetereva [33]	(1) Gathering Data (Google Maps, Rental Website, Characteristics, Description, Photos), (2) Extracting Features (Keyword-Based Variables, Type of the View, Location-Based Variables), (3) Data Cleaning (Combine Data Sources into Final Data Set), (4) Regression (Ordinary Least Squares Model, Random Forest), (5) Explanatory Analysis (Comparison of the Models, Variable Importance, Partial Dependence Plot, Local Interpretable Model-agnostic Explanations)	4/6
Brackmann et al. [28]	(1) Business Insight, (2) Data Understanding, (3) Data Preparation, (4) Modeling Phase, (5) Evaluation Phase, (6) Go Live	3/6
Geiler et al. [31]	(1) Data, (2) Sampling (Over-/Undersampling, Hybrid), (3) Model Fitting (Supervised Learning, Semi-Supervised Learning), (4) Evaluation (Cross (Stratified) (K-fold) Validation, Metrics)	3/6
Jafarzadeh et al. [4]	(1) Business Understanding, (2) Data Understanding, (3) Data Preparation, (4) Modeling, (5) Evaluation, (6) Deployment	3/6
Esmeli et al. [30]	(1) Data Collection, Pre-Processing, and Data Generation, (2) Feature Extraction and Model Creation, (3) Data Mining Model Evaluation	3/6

all the frameworks ultimately included, along with their study data (left), the specific names of the process steps P (center)—determined through an analysis of the text elements—and the overall thematic relevance evaluation. The purpose of Table 2 is to derive the titles for P_MECOVMA by systematically aggregating the naming of other frameworks, based on the value of ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}$.

The synthesis of the names of all process steps P in Table 2 results in the following titles for the MECOVMA framework: P1_MECOVMA “Predictive Data Collection”, P2_MECOVMA “Data Preparation in a Volatile Macro-Environment”, P3_MECOVMA “Prediction Modeling”, and P4_MECOVMA “ML Evaluation and Marketing Application” (see Figure 4). The subsequent sections give a description of the procedure for the development of these titles for the individual P_MECOVMA steps.

The first step of P1_MECOVMA “Predictive Data Collection” initially draws on studies with a high overall relevance evaluation, such as that by Efat et al. [78], which emphasize the data model as the initial starting point for their forecasting model. Similarly, Tsao et al. [34] utilize operational data as an initial stage in their framework, intentionally incorporating both internal information and external indices on environmental conditions. Studies like that by Esmeli et al. [30], Gharibshah and Zhu [79], and Potrawa and Tetereva [33] further underscore the importance of collecting raw data with predictive insights. In this context, the significance of extracting data from marketing phenomena and actions is particularly highlighted by Herhausen et al. [8] and Ma and Sun [63]. Additionally, Brackmann et al. [28] interpret the initial step to include an understanding of both the data and the underlying business processes, a perspective similarly adopted by Jafarzadeh et al. [4].

In the subsequent step, P2_MECOVMA “Data Preparation in a Volatile Macro-Environment”, is particularly reflected in frameworks such as those by Kharfan et al. [32], which extensively focus on data preprocessing prior to modeling. This includes data filtering and aggregation as well as feature engineering and selection. Data preparation activities are also evident in the frameworks of Brackmann et al. [28], Esmeli et al. [30], and Herhausen et al. [8]. The framework by Raizada and Saini [74] dedicates a series of steps to this purpose (Data Cleaning and Preparation, Splitting the Data (Training Data, Testing Data), Pre-Processing, Feature Selection/Feature Measurement)—however, most other authors consolidate such elements into one step, as reflected in P2_MECOVMA.

Next, P3_MECOVMA “Prediction Modeling” is synthesized from Table 2. Among the thematically particularly relevant frameworks, Efat et al. [78] describe this process simply as prediction. Ma and Sun [63] specifically focus on modeling in marketing and extend beyond ML to include AI in general. Tsao et al. [34] refer to this step as a forecasting framework and include not only ML modeling but also alternative models. Thus, the title of P3_MECOVMA is not limited to ML alone. This step is uniformly seen as essential, as demonstrated by Herhausen et al. [8], who focus on the training of models in the marketing context, or by Gharibshah and Zhu [79], who describe it similarly to Raizada and Saini [74] as learning. Geiler et al. [31] specify the process step terminologically as model fitting, whereas Brackmann et al. [28] and Jafarzadeh et al. [4] explicitly refer to it as modeling.

The final step of “ML Evaluation and Marketing Application”, referred to as P4_MECOVMA, is based on frameworks such as that set out by Tsao et al. [34], the final phase which emphasizes the value generated by the use of ML. Ma and Sun [63] address applications within a marketing context more explicitly, as well as further aspects of the implementation of ML-driven marketing actions. Taking a comparable approach to that of Ngai and Wu [80]—who include references to how they impact the “P’s” of marketing—Herhausen et al. [8] also concentrate on marketing decision-making. In addition, Jafarzadeh et al. [4] acknowledge the evaluation of models as a key aspect of ML frameworks before their final application. Along similar lines, Brackmann et al. [28] also include an evaluation phase before completing and implementing their process. Kharfan et al. [32] underscore this by highlighting the importance of evaluation, particularly when it comes to identifying the model’s accuracy and bias. Potrawa and Tetereva [33] explicitly add an explanatory analysis at the end of their framework. As the intended result of Section 3.1, Figure 4 illustrates the developed outcome, displaying the number and titles of the overarching process steps for the MECOVMA framework—with dashed feedback loops representing the iterative execution.

3.2. Methodology Phase II: Content Design

To achieve this study’s research objective of constructing an actually ready-to-use methodological framework, Methodology Phase II builds upon the foundational structure established in Methodology Phase I. This development is grounded in theoretical insights from a systematic literature review. The foundation for this is provided by the systematic literature review conducted by Muth et al. [73], which follows the process guidelines of Xiao and Watson [83]. It scans the Web of Science database for relevant papers via a search string including terms such as “machine learning”, “demand”, “macroeconomic volatility”, and “forecasting”, which are connected by using so-called Boolean operators. The resulting 2877 hits are filtered, excluding non-relevant subject areas and thematic categories, non-English-language and non-journal articles, as well as duplicates. After reviewing the thematic relevance of titles and abstracts and assessing the internal and external validity by independent researchers, 64 relevant peer-reviewed studies remained in the final literature sample. The results from this systematic literature review specify the P_MECOVMA process steps with precise tasks and action guidelines to enhance, through the use of such, a methodological framework consistency and trustworthiness in projects [22,41,84,85,86].

3.2.1. Content Detailing with a Systematic Literature Review

Section 3.2.1 expands on the results developed in Methodology Phase I and further details the design of the process steps of the MECOVMA framework. Each of the four steps is discussed in detail and in the intended chronological order of the framework, although iterative and reverse movements between the P_MECOVMA steps are possible. This discussion provides specific guidance for practitioners and academics in applying the framework, which is ultimately visualized in Section 3.2.2. From a scientific-theoretical perspective, the MECOVMA framework combines both inductive approaches for generating new hypotheses based on empirical data and deductive approaches for systematically verifying and validating these hypotheses, which are discussed in detail in the following section.

• P1_MECOVMA “Predictive Data Collection”

As the first step of the MECOVMA framework, raw data is collected within the marketing context to support the subsequent generation of predictions [32,87,88]. First of all, a context is provided by a clearly defined marketing problem or phenomenon. This step not only includes defining and conceptualizing the research but also integrates existing evidence, along with a possible inductive formulation of the hypotheses.

The systematic literature review used to detail the content [73] of the steps provides a general guideline in this regard, pointing out the necessity of collecting multivariate data inputs to take the complex nature of demand in volatile market environments into account [87,89,90]. The data collection process can require certain prerequisites, such as a minimum sample size, which is relevant for the effective performance of ML without overfitting, as well as standards for data timeliness, representativeness, or source quality (see limitations in P4_MECOVMA). Relying only on univariate approaches—and, therefore, not initially collecting multiple potential input variables—is not recommended in the MECOVMA context, as predicting future developments based solely on historical behavior has strong limitations when anticipating structural changes in the macroeconomic environment [89,91].

Three aspects make up the first step of multivariate data input within the framework: It is recommended to consider (1) forward-looking signals and sentiments for the systematic collection of variables that can potentially serve as key indicators for predicting how demand will evolve in the future [88]. These could be traditional indicators derived from surveys, for example, regarding consumer confidence and expected spending in the forthcoming year, or indicators of economic sentiment across different industries [90]. Predicting consumer behavior using search engine trends and platforms such as Google Trends is an example of a more modern method [87]. To ensure better model predictions, including suitable macroeconomic indicators can be essential. The (2) marketing mix variables may include price, which can indicate an inflationary market environment and is, therefore, considered to be a relevant factor in the context of the research problem. Price constitutes one element of the “4 P’s” of marketing, which can encompass numerous important factors that determine demand behavior. (3) Further MECOVMA-relevant predictors should be considered given that a multitude of other variables potentially influence marketing-relevant measures such as demand. These include, for example, consumer behavior predictors, demographics, weather, product data, seasonality, and calendar variables [36,92]. For collecting data inputs, various data sources can be utilized, ranging from open-source databases to government records and other public institutions. Internal company datasets are also relevant, with an emphasis on collecting data across different company departments to overcome silos.

• P2_MECOVMA “Data Preparation in a Volatile Macro-Environment”

The second essential step within the MECOVMA framework is data preparation for final preprocessing before modeling [73]. This must be targeted toward a volatile macroeconomic environment to maximize the accuracy of the predictions while minimizing unnecessary complexity, noise, and the time it takes for the model to compute [93,94].

The respective variable selection process comprises three aspects. (1) Temporal causality is an inductive analytical approach for identifying key indicators from the previous input of multivariate data. Methods such as the (Partial) Autocorrelation Function ((P)ACF) are used to identify or analyze temporal dependencies. One purpose of these key indicators is to act like warning signals for prospective market developments, trends, and turning points, as opposed to variables that are coincident and do not provide early indications of future changes in demand [88,95]. (2) Variable extraction involves transforming existing variables into a new set of factors that capture the underlying data structure [93,94]. In order to draw attention to important data inputs, it is possible to apply unsupervised ML methods, like Factor Analysis and Principal Component Analysis [94,96,97,98]. These methods are used inductively in the context of marketing to explore underlying data structures. The purpose of (3) complexity reduction is to decrease the number of input variables to those that are potentially necessary or value-adding. One can use descriptive analyses such as correlation analyses, as well as supervised ML techniques including regression methods—such as the Least Absolute Shrinkage and Selection Operator (LASSO) for variable selection and regularization to reduce model complexity—for this purpose [99]. This process can be supported by expert assessments, alongside evaluations that are merely data-driven, through the selection of variables that are relevant to the practice [34,91,100]. Additional tasks, such as the crucial one of data cleansing, complete the step P2_MECOVMA “Data Preparation in a Volatile Macro-Environment” and can play an important role in ensuring the reliability of the ML model.

• P3_MECOVMA “Prediction Modeling”

P3_MECOVMA is the key stage of the MECOVMA framework for selecting and generating models that facilitate accurate prediction within a complex dynamic. Two main aspects make up this stage: firstly, modeling strategies that detail the generation methods for the models, and secondly, prediction algorithms specifying the chosen model. The first level of modeling strategies comprises three primary aspects. The dynamic adaptation of the model to the most recent data is enabled by (1) temporal focus. Techniques like rolling windows focus on recent data, enhancing the model’s ability to respond to environmental changes [71,92]. The process of (2) model integration such as pooling can be employed to consolidate multiple analogous sets of demands into a collective group and can enable prediction models to accurately identify structural shifts and common demand patterns [92]. The purpose of (3) model specification is to allow models to be selected variously, for example, in accordance with a particular set of consumer demands. In order to assess the model’s performance based on volatility, the Coefficient of Variation can be applied. Decomposition techniques enable models to be applied for different consumer demands, for example, a standard and a volatile component [89].

The prediction algorithms on the second level of P3_MECOVMA showcase a wide variety of algorithms—enabling the inductive formulation of hypotheses for comparison and subsequent evaluation. Despite the need for model testing on a project-specific basis, certain prediction models based on the systematic literature review are specifically recommended for testing [73]. Regarding ML architectures, Artificial Neural Networks are fundamental for both metric and discrete target values. With a specific architecture, such as recurrent networks, they are also appropriate for time-dependent contexts [71,101,102]. Furthermore, Random Forests/Decision Trees and Support Vector Machines—especially for classification tasks—are repeatedly used successfully in current topic-specific research studies [29,99]. Regression and auto-regressive models serve as robust prediction bases that are relatively simple but can provide a sophisticated benchmark for more complex ML algorithms. Additionally, other ML and non-ML models are relevant in this context [89,95,103]. The systematic inclusion of non-ML approaches is important given the rapidly changing nature of prediction algorithms (e.g., TimeGPT) and to improve the robustness of model selection through comparisons with econometric and traditional time series techniques (e.g., exponential smoothing). A combination of multiple models (hybrids, ensembles) is explicitly recommended in the MECOVMA context to reduce the effects of suboptimal individual predictions in the event of pronounced market volatility and to increase the stability of the prediction [31]. For the final modeling phase, established practices, such as partitioning data into training, testing, and validation sets, as well as cross-validation, can be employed to minimize typical ML risks, such as overfitting [73].

• P4_MECOVMA “ML Evaluation and Marketing Application”

The primary task of the final step of the MECOVMA framework, referred to as P4_MECOVMA, is to evaluate predictions, critically examine models, and determine whether they have been successfully applied to practical marketing. The following two approaches are recommended for the evaluation of predictions in the context of MECOVMA. First, applying deductive methods in a narrower statistically-oriented evaluation and making use of a variety of complementary metrics to quantitatively assess the accuracy of predictions and associated hypotheses. The use of a minimum of three prediction performance indicators is advised [73]. The prediction horizon and the uncertainties it harbors should also be a basis for evaluating performance [92,104]. Second, a broader business-oriented evaluation is recommended to be conducted to examine the wider implications of predictions when it comes to operational marketing activities and business resources [60,100].

In the second stage of P4_MECOVMA, four main limitations that can hinder the effective use of ML in marketing should be examined during the process [73]. The reliability of modeling results can be compromised by (1) data-specific limitations, including data gaps and irregularities, as well as a limited range of sample sizes and variables, such as specific macroeconomic indicators that can be difficult to forecast reliably or are subject to delayed reporting [29,56,93,101]. A feedback loop is therefore incorporated for further optimization of the data input in P1_MECOVMA. (2) Generalizability refers to removing restrictions on the transfer of ML predictions to a wider range of prediction scenarios, thereby ensuring that the results can be applied to a variety of industries, countries, or objects of prediction [56]. Consequently, a feedback loop is also integrated that refers back to the marketing issue specified in P1_MECOVMA so that the scope can be adjusted as required. Limitations that are specific to the (3) methodology and application refer to the challenge of finding a robust theoretical basis for models and their relationships and of taking into account all the relevant factors that exert an influence on predictions [56,105,106], with feedback according to the prediction modeling in P3_MECOVMA. Furthermore, as the “adoption is limited by the level of trust afforded by given models” [107] (p. 4571), it is necessary for stakeholders to understand the predictive capabilities of ML and its established relationships, underlined by the limitations of (4) interpretability in volatile environments [56,100]. ML relationships can be proactively controlled during model selection in P3_MECOVMA by choosing more transparent ML models, which may involve a trade-off with predictive performance. Retrospectively, these relationships can also be examined at the P4_MECOVMA stage using analytical tools [92,108]—such as feature importance or Local Interpretable Model-agnostic Explanations. This model transparency can also be important for the ML modelers themselves to understand the mechanisms and architecture of the model and thus to validate and proactively intervene when necessary. Finally, the application of ML predictions to marketing means implementing them in real-world scenarios, including operational marketing measures or volatile decision-making situations [56]. An inductive process can lead to the development of new hypotheses [38], which can then be tested to develop theories based on insights obtained through the MECOVMA process. Moreover, ongoing monitoring is recommended here for ML predictions after the implementation phase, with retraining often necessary [109], particularly when market conditions and demand behavior are influenced by volatility.

3.2.2. Framework Visualization

This section presents the final visualization of the MECOVMA framework, as detailed in Section 3.2.1, and emerges from the analyses in the previous Methodology Phases I and II. The framework is graphically illustrated in Figure 5 to provide a concise overview of its structure and function. In practical usage, the starting point is intuitively at the top, at P1_MECOVMA, and progresses iteratively downward, following the guidance of the overarching process steps and tasks on the right side step by step.

The content that is presented in Section 3.2.1 is supplemented with significant further operational aspects within the visualization of the MECOVMA framework in Figure 5. The iterative transitions between various P_MECOVMA steps and their sub-elements are presented with the help of dashed lines that make it easier to work toward optimal results in stages. The need to intervene at an early stage to ensure the systematic mitigation of limitations pertaining to the data, owing to inadequate input variables or biased information when predictive data is collected, is highlighted, for example, by the link between data input (P1_MECOVMA) and limitations (P4_MECOVMA). Furthermore, as indicated by the line from P2_MECOVMA to P1_MECOVMA, adjusting or extending the data collection is recommended if the results of the variable selection are inadequate. Additionally, the inclusion of business transparency requirements in the selection of prediction algorithms is emphasized by the link between the prediction modeling presented in P3_MECOVMA and the limitations to interpretation in P4_MECOVMA.

Moreover, the large arrows shown on the left of Figure 5, representing the superordinate steps, have wider implications, illustrating a situation in which the final ML evaluation and application to marketing lead to an inadequate project result. In this case, it is recommended to restart the MECOVMA process, which is shown by P4_MECOVMA being followed by a shaded P1_MECOVMA arrow, in order to build on the existing findings and make iterative improvements.

As shown on the left of Figure 5, marketing science and business management theories form the basis for the theoretical expertise that underpins the MECOVMA process. This impacts the research conceptualization in P1_MECOVMA as well as, for example, variable extraction and complexity reduction (P2_MECOVMA) and ML model generation (P3_MECOVMA). Furthermore, as represented by the broader business-oriented evaluation (P4_MECOVMA), the practical knowledge gained from the particular field to which the ML predictions are to be applied strengthens the MECOVMA process. Gennatas et al. [107] (p. 4571) state with regard to this, “In practice, we rely on human experts to perform certain tasks and on machine learning [ML] for others. However, the optimal learning strategy may involve combining the complementary strengths of humans and machines”. It is also possible, wherever permissible, to integrate current advancements in (generative) AI, especially by means of Large Language Model Applications, as a co-pilot into the MECOVMA process. This can, for instance, aid in the initial ascertainment of which factors could influence the “Multivariate Data Input” (P1_MECOVMA) or the programming of prediction algorithms (P3_MECOVMA).

3.3. Methodology Phase III: Framework Validation

Building on the foundational steps of Methodology Phases I and II, Methodology Phase III—also demonstrated in the lower part of Figure 1—focuses on validating the MECOVMA framework. A discussion and assessment of the methodological framework by a group of experts—like in a focus group—is considered a recognized research practice [47,51]. This is intended to strengthen the credibility and broad applicability of the methodological framework and to refine it to maximize its ability to meet the requirements of stakeholders [35].

3.3.1. Symposium Presentation with Expert Review

As Rweyendela et al. [43] (p. 2240) explain, the “validation of the framework… inputs from experts aided in evaluating its adoption, utilization and acceptability” is a relevant aspect when it comes to developing a framework. McMeekin et al. [41] and Mirza et al. [110] further support this step toward the final optimization of the framework. For this reason, MECOVMA is presented to international experts in the field of forecasting and AI at an international symposium to gather feedback and validate the framework. The presentation to the attending experts includes the components “Context and Objective”, “Research Gap and Thematic Focus”, “Methodology: MECOVMA Framework Development”, “MECOVMA Framework: Steps 1–4”, “Limitations and Future Research Directions”, and “Discussion” [111]. During the final “Discussion” component, two questions are raised by the author, and an expert discussion takes place, focusing in particular on the expected practical applicability of MECOVMA. From a methodological perspective, the discussion is analyzed using qualitative content analysis [112]. This analysis is primarily based on the principles outlined by Mayring [113], which serve as the guidelines for the systematic summary of the textual material. This represents a scientifically established and frequently adopted procedure [28,114]. The process of framework validation is schematically displayed in Figure 6.

3.3.2. Qualitative Content Analysis for Framework Refinement

The documentation of the expert verbal discussion is organized in a table. To serve the specific aims of this study, the first column of this table entails the statement in the interest of understanding the dynamics. The dialog is in response to the following two questions that the author poses after introducing MECOVMA: “Q1: What challenges do you foresee in the practical application of the MECOVMA framework?” and “Q2: Can you share experiences where such a framework has been (un)successfully applied?” [111]. Anonymity is maintained by not collecting or disclosing the names of experts and by omitting specific personal data [47]. Only responses provided voluntarily by experts, who self-identify by raising their hands to address particular framework questions and contribute to the further development of the research, are considered. Building on the methodological foundations of Mayring [113], an inductive qualitative content analysis technique is employed to summarize and categorize the qualitative data. As part of this approach, it is important to define that the coding units are the individual expert insights related to the research questions; the context units focus on the oral discussion directly after the presentation until the end of the question section; and the analysis units comprise the full documented material from this period. The analysis of this material is then conducted in three steps. Firstly, the key content of each original quote is paraphrased. This is followed by establishing broader themes and generalizing. The third step is to condense the key content to its essential core. Here, an additional column outlines the implications for the MECOVMA framework (framework refinement) based on the outcomes of the qualitative content analysis. This allows the framework to be optimized through an inductive process that is applied directly to the text material, excluding content that is not essential or is merely an embellishment. This systematic procedure is intended to be rule-guided and as intersubjectively verifiable as possible [113,115,116].

In this way, the presentation of the research results is intended to facilitate the collection of insights from a specialized group of experts. These experts provide a critical assessment of the existing framework and its elements in relation to the requirements of stakeholders [47]. This process supports the optimization of MECOVMA and the inclusion of additional insights in this paper. In order to optimize the framework—using the qualitative content analysis—it is revealed as important, for example, to ensure better model predictions by identifying and including suitable macroeconomic indicators, which are hence outlined in more detail in P1_MECOVMA based on this feedback. Furthermore, according to the opinion of an expert in the audience, technical challenges associated with data collection may be even less significant than organizational barriers. An example mentioned includes fragmented databases in different departments of a company, which can critically impede data collection in P1_MECOVMA. As such, this consideration is explicitly integrated into the framework at this stage. Moreover, data cleansing is highlighted by an expert as highly relevant for ensuring the robustness of predictive modeling processes and is therefore included in the P2_MECOVMA stage for data preparation. The experts also emphasize the importance of model transparency—not only for stakeholders but also for ML modelers to gain a deeper understanding of model architectures. This insight is thus incorporated into P4_MECOVMA. Furthermore, another notable expert inquiry raises the question of whether MECOVMA is capable of handling multi-level data hierarchies—such as multiple related products—or if it is limited to single-level predictions. This consideration is included as an avenue for further specification in Section 5. Overall, the validation procedure within Methodology Phase III supports the iterative refinement of the MECOVMA framework, helping to promote its wider acceptance in the scientific community [45].

4. Discussion

MECOVMA distinguishes itself from existing frameworks through its specific features, addressing the four research gaps outlined in the introduction (Section 1)—namely (i) consolidation, (ii) relevance, (iii) interdisciplinarity, and (iv) individuality.

For (i) consolidation, MECOVMA integrates data from N = 15 frameworks for the overall structure (Section 3.1), as well as from N = 64 reviewed studies for the content design (Section 3.2), leading to a consolidation that includes a harmonization of different approaches. The MECOVMA framework stands out from the existing frameworks through the integration of multidimensional data sources at the “Predictive Data Collection” (P1_MECOVMA) stage—often not addressed with this specificity [31]. Traditional statistical and ML techniques are brought together with theoretical and practical concepts in “Data Preparation in a Volatile Macro-Environment” (P2_MECOVMA) to ensure that the potential of traditional, well-established methods is exploited—compared to frameworks exclusively focused on ML [63,80]. While “Prediction Modeling” (P3_MECOVMA) focuses strongly on the algorithmic perspective with a wide range of models, the final step of “ML Evaluation and Marketing Application” (P4_MECOVMA) extends beyond this by considering an evaluation of the prediction in the business context and its impact there. This final step demonstrates a significant differentiation compared to certain analyzed frameworks in Section 3.1 [74,78].

To address (ii) the relevance and context of current macroeconomic volatility [9,55], MECOVMA responds to the criticism that established research approaches are often not aligned with the current market reality [35]. To ensure the actuality of the ML model, for example, aspects such as rolling time windows (P3_MECOVMA) are integrated, as well as state-of-the-art data sources, for example, Google Trends, and algorithmic ML trends, such combined models (see P1_MECOVMA and P3_MECOVMA). Furthermore, the prediction and what it is based on (P4_MECOVMA) are considered to be updated on an ongoing basis. Comparing the modeling strategies of MECOVMA with those of Tsao et al. [34], which holds the highest overall relevance evaluation ${\sum }_{\mathrm{i}=1}^6{\mathrm{R}}_{\mathrm{i}}$, MECOVMA differs, for example, in that it incorporates volatility into its modeling strategy, whereas Tsao et al. [34] use external information sources but do not account explicitly for these aspects. Also contrary to other frameworks [31,32,74], MECOVMA specifically addresses practical implementation steps, especially with regard to model application and operational maintenance (see P4_MECOVMA).

In order to maintain an (iii) interdisciplinary perspective and, hence, to take a more comprehensive approach, the framework integrates methods from economics, computer science, and marketing in order to ensure that it addresses objectives across different disciplines. These include quantitative academic targets represented by narrower statistics-oriented evaluations, as well as practical targets involving wider business-oriented evaluations (P4_MECOVMA). By taking a dual approach, the MECOVMA framework distinguishes itself from established approaches from the perspective of scientific theory. This extended methodology allows the confirmation of existing hypotheses, as well as the generation of new insights through the integration of both explorative and confirmatory aspects. The purpose of integrating theory and practice is to create reliable predictive models that have a solid theoretical foundation, as well as being applicable in practice [38,51]. Consequently, the focus of MECOVMA is on the requirement pointed out by Pfister et al. [35] (p. 203) of “not diminishing the wide gap between the state of theory and leading-edge practice, thus creating demand for more theoretically interesting and practically relevant theory-building studies”. Despite the interdisciplinary characteristics of frameworks, including those presented by Herhausen et al. [8], Ngai and Wu [80], and Van Giffen et al. [7], they are often based strongly on the data modeling process that either scientifically validates existing assumptions or generates new hypotheses [31,78,79].

The tailored adaptation to meet specific challenges—and thus, the (iv) individuality of the structural design of MECOVMA—is considered an important aspect in the framework development [51]. The applicability of MECOVMA across a wide range of uses is extended by an approach that is sensitive to different contexts through the selection of a prediction model in accordance with the individual characteristics of the marketing problem or phenomenon. This sets it apart from multiple other frameworks that have been analyzed, such as Gharibshah and Zhu [79], with a focus on predicting user responses in online advertising; Kharfan et al. [32], involving forecasting demand for seasonal products in fashion retailing; and Geiler et al. [31], with predictions of customer churn. MECOVMA also differentiates itself from other frameworks [74,81] because of the evident inclusion of feedback loops at different stages of the process (see dashed arrows in Figure 5) and its structure that integrates feedback to enable general self-optimization (see thick step arrows in Figure 5).

5. Limitations

Although a systematic procedure applies to the three-phase development methodology of MECOVMA [41], it is nevertheless susceptible to different limitations in terms of (1) methodology, (2) content, and (3) implementation.

In relation to (1) methodology, building MECOVMA upon existing frameworks (Methodology Phase I) carries the risk of implicitly transferring their limitations [117], even if a critical reflection and synthesis are conducted. This applies particularly when various disciplines with their own goals and practices are brought together, which can make them difficult to compare [118,119]. In addition, a risk of the unintentional exclusion of the relevant literature is posed by the selection and search criteria of both the framework search and the systematic literature review to be used, as well as by the journal selection of databases (Methodology Phases I and II) [73,120,121]. Furthermore, while for relevance evaluation R_i, the use of binary Boolean values as a more neutral metric limits subjective influences such as primacy effects, ideologies, and confirmation bias [55,122,123,124], it has the risk of diminishing nuances in the differentiation between the evaluated frameworks. The subjective influence of the researcher is limited but not excluded, for example, during data collection (Methodology Phases I and II) or the execution of the symposium discussion and qualitative content analysis coding (Methodology Phase III), which can affect the credibility of the findings [47].

In terms of the (2) content of the framework, MECOVMA is strongly targeted through its focus on the thematic triad comprising (T₁) AI/ML/Statistical Learning based Prediction, (T₂) Marketing/Consumer Behavior, and (T₃) Volatility/Uncertainty/Macroeconomic Environment. While this clearly delineated thematic focus can enhance reliability, particularly in framework evaluation, it can compromise adequate attention to thematic areas, including ethical considerations such as fairness or bias avoidance [1,26,125] or the security and protection of data [126,127]. Additionally, a greater emphasis on addressing organizational aspects, such as data barriers between different departments of a company, may be necessary. The framework can also be extended to encompass different data hierarchies, including multi-level structures like multiple related marketing products. Additionally, updating the framework regularly is a substantial requirement because of the rapid development of state-of-the-art ML algorithms, marketing, and economic analysis [21,26,55].

Specific conditions are necessary for (3) the implementation of MECOVMA on a practical level, such as qualified expert personnel, state-of-the-art IT resources, and data that are comprehensive in terms of both quality and quantity [21,55,88,107]. Gennatas et al. [107] (p. 4571) highlight the following when it comes to the importance of data: “Lacking any general knowledge of the world, it is no surprise that current ML algorithms will often make mistakes that would appear trivial to a human”. Hence, by including a range of data sources and modeling techniques in MECOVMA, reinforced by iterative feedback loops, the quality of predictions can be improved, but it might make practical applications more complex. Therefore, the applicability of the framework is validated through feedback from experts (Methodology Phases III). Qualitative content analysis [113] is used to analyze the feedback comprehensively before it is included in MECOVMA to enhance its trustworthiness. In order to strengthen the external validity of the framework, however, it is necessary to validate it further empirically in real-world situations. Validation methods to ascertain its general applicability to a range of industries and contexts include testing by various technical and non-technical users such as practitioners and academics, as well as representative samples from different fields of application [37,47,51].

From a concrete application perspective, the MECOVMA framework can be recommended for use in various business scenarios within dynamic market environments, enabling further empirical validation. Fundamentally, the methodological framework guides predictions for typical marketing objectives, such as consumer decision-making and market developments, under conditions characterized by volatility. One example is predicting purchasing behavior in the consumer goods sector, in which ML models aim to anticipate changing consumer habits and—following the conceptual recommendations in P1_MECOVMA and P2_MECOVMA—incorporate forward-looking macro factors like sentiment indicators. Furthermore, the MECOVMA framework can guide forecast development for future vehicle demand, particularly under conditions of dynamic pricing or shifting economic circumstances that require appropriate modeling strategies and predictive algorithms (P3_MECOVMA). The framework is also recommended for identifying potential customer churn, such as on e-commerce platforms or streaming services. In these cases, the focus is on applying marketing strategies that leverage churn forecasts to develop targeted marketing activities and conduct broader business-oriented evaluations to validate practical benefits—such as reducing churn rates—in real-world operations (P4_MECOVMA).

6. Conclusions

In response to the recognized need for an adapted approach to dynamic marketing environments [55], this study introduces the MECOVMA framework and details its development. MECOVMA aims to integrate ML—described as “invaluable across disciplines” [107] (p. 4571)—specifically into marketing prediction under the recent macroeconomic conditions, and is an acronym derived from this focus. The MECOVMA framework intends to represent a bridging of relevant gaps in current research, as indicated in Section 4. A structured methodology that brings interdisciplinary theoretical and practical aspects together is enabled by synthesizing current findings in the fields of both computer science, which focuses on ML, and business administration, particularly marketing.

A three-tier methodology (Section 3) forms the basis for the MECOVMA framework, mainly inspired by McMeekin et al. [41]. Methodology Phase I, in which the steps are selected (Section 3.1), comprises the application of strict inclusion and exclusion criteria to the critical analysis of existing frameworks. This analysis flows into the development of the P_MECOVMA steps by enabling the evaluation of the degree of relevance of these frameworks in relation to the defined relevant thematic triad of this study. In content design (Section 3.2), representing Methodology Phase II, a systematic literature review that involves an analysis of N = 64 peer-reviewed studies provides the basis for the specification of framework process steps. In this phase, a sequential description of the individual process steps is conducted, including a complete visualization of MECOVMA. In the subsequent Methodology Phase III (Section 3.3), a qualitative content analysis is carried out on the basis of a presentation and feedback from prediction experts at an international symposium, in the interest of validating and refining the applicability of the MECOVMA framework that was developed before. This methodology, however, still requires additional empirical assessment in relation to MECOVMA’s thematic coverage, the resources that are required, and its applicability in practice, as detailed in Section 5.

The ultimate purpose of MECOVMA is to make it easier to effectively apply ML algorithms to make marketing predictions, even in the largely dynamic and challenging current circumstances. By integrating a variety of perspectives and an overall process, from problem definition to modeling to implementation and updating, a wider approach to the research problem is taken. All four of the respective P_MECOVMA steps include a continuous adjustment of ML models to the fluctuating macroeconomic climate through the integration of iterative features in MECOVMA, including feedback loops all along the line. MECOVMA is based on the notion that it provides important value even if its structured guidelines are only partially applied, for example, on account of limited resources, in line with the principle proposed by Cruz Rivera et al. [45]. As highlighted in recent studies [35,43], one of the key objectives of MECOVMA is to offer adaptability and flexibility while maintaining a structured framework that facilitates a systematic approach. It also addresses the challenge posed by the fact that “the inherent uncertainty of highly dynamic environmental conditions brings… less rational strategic decision-making” [55] (p. 29). The methodological framework hence forms the basis for further intensified research in this important field.