From Concept to Market: Ensemble Predictive Model for Research Project Crowdfunding Readiness

Abstract

This study introduces an ensemble model that integrates random forest, gradient boosting, and logistic regression to predict the success of crowdfunding campaigns. Our research develops a novel set of metrics that assess the developmental stage of research projects, facilitating the transition from concept to market-ready product. Utilizing data from multiple sources, including Kaggle’s dataset of Kickstarter and Indiegogo projects and a proprietary dataset tailored to our study, the model’s performance was evaluated against traditional implementations of random forest and gradient boosting. The results demonstrate the ensemble model’s superior performance, achieving an accuracy of 98.94% and an F1 score of 98.81%, significantly outperforming the individual models, showing the best accuracies of around 91% for random forest and lower scores for gradient boosting. This enhancement in predictive power allows for optimized resource allocation and strategic planning in project development, thereby increasing the likelihood of crowdfunding success. This approach streamlines the process of bringing innovative ideas to final products, while at the same time offering a methodologically advanced tool for stakeholders to enhance their campaign strategies effectively.

1. Introduction

The basis for this research lies in the well-known problem faced by academic research projects during their transformation from a concept into a product through crowdfunding platforms. In fact, success rates on these platforms vary greatly with specialized platforms for crowdfunding research projects, usually having success rates around 50% [1], whereas more general ones see success rates dropping to approximately 20% [2]. Therefore, it is absolutely essential to have an in-depth understanding of the factors that affect the outcome of crowdfunding campaigns in an academic setting, such as the field of study, choice of platform, and clarity and completeness in articulating project goals and presentation strategies.

Research works in technology, environmental science, and medical science tend to attract more attention and support because they are viewed as offering immediate benefits to society while also aligning with backers’ concerns. Additionally, selecting a particular crowdfunding platform may amplify or downplay the chances; sites that are dedicated exclusively to scientific funding will often have communities more prone to supporting such initiatives, hence the increasing odds for success as opposed to general versions. Furthermore, how effectively a campaign is run is determined by its presentation, like how clearly its goals and budgets are articulated or what methods are used for engagement and promotion. Engaging multimedia presentations, clear and exciting project descriptions, along with regular updates play a critical role in this regard. Moreover, tangible incentives that actually reward those who participate at the campaign level go a long way towards having backers feel involved, thus encouraging them to direct funds towards the project’s realization.

Against this backdrop, this paper seeks answers to the following research question: How might crowdfunding become a successful path for research projects thanks to careful campaign preparation? It aims at examining both potential successes and pitfalls leading to failures in crowdfunding campaigns within the context of research endeavors. Thus, the goal is creating an instrument that predicts whether or not research projects are ready for initiating their own crowdfunding.

To achieve this goal, this research critically examines current methods and techniques of evaluating such projects and incorporates them into new metrics to effectively determine the readiness level of innovative research initiatives. Furthermore, a combined machine learning model is adopted in order to boost the accuracy and reliability of the proposed readiness evaluation metrics. Through this multifaceted approach, the aim is to offer a comprehensive framework that aids young academic researchers in optimizing their crowdfunding campaigns, thereby bridging the gap between innovative ideas and their successful realization as products that contribute value to society. This aligns with the findings of Butticè and Ughetto, who emphasize the importance of understanding the various dimensions of crowdfunding research, including evaluation methods and metrics that can enhance project success [3].

For our study on the crowdfunding of research projects, we referenced a seminal work from five years ago [1], which, besides identifying the dedicated platforms available at the time, conducted an extensive analysis of data from over 700 campaigns on Experiment.com (accessed on 1 July 2024). This platform remains notable as the sole survivor among the eight independent platforms mentioned in the study, highlighting its significance and durability in the niche of research project crowdfunding. The disappearance of several platforms dedicated to crowdfunding research projects raises pertinent questions about the initial evaluation of the readiness of projects hosted on these platforms. It is this research gap that the present research aims to address by proposing a measurement tool based on machine learning to assess the readiness level of a research project for crowdfunding. This work seeks to contribute to the broader understanding of what key elements contribute to the success or failure of crowdfunding campaigns in the academic field, shedding light on how researchers can better prepare and position their projects for public funding.

Sauermann et al. [1] have been seen as laying down the foundation for investigating potential options of scientific projects’ funding through crowdfunding, which Daldrup et al. [2] have also examined closely for the German universities case. The latter paper suggested that young researchers’ innovation projects can be supported through crowdfunding initiatives. By exploring this path, both papers contribute to a growing body of knowledge that recognizes the value of crowdfunding not just as an alternative financing method but as a means to democratize the funding landscape for scientific research. Particularly, this approach is important for young researchers who may find themselves locked out by traditional mechanisms of funding their ideas in science. Their consistency with crowdfunding is an essential element in fostering innovative environment where all stakeholders, including wider backers, participate.

In order to launch innovative research projects via crowdfunding, much more than just new ideas are necessary; it calls for strategic planning and adequate preparation. The way these projects are prepared at their foundational level matters so much given that they are going into “business” as far as reaching out for financial support and public recognition via crowdfunding is concerned. This type of financing is a great opportunity for advancing investigations and bringing new findings to a wider audience; however, the success factors are numerous, and their careful examination should be undertaken to avoid the risk of failure.

Uncovering the reasons behind unsuccessful crowdfunding campaigns during evaluation enables overcoming them. Common problems are lack of market analysis, weak project presentation, uncertain value propositions, and a poor project plan. To tackle these tasks more effectively, our approach includes multidimensional evaluation built on parameters such as technological readiness (TRC), marketing readiness (MRC), commercial readiness (CRC), and management team readiness (MTRC). These dimensions reflect different aspects of the project’s readiness, making it both innovative in thinking and positioned strategically towards meeting market demands and expectations from possible supporters. Additionally, there are four more parameters, namely, the platform–campaign alignment index (PCA), financial efficiency index (FEI), campaign engagement index (CEI), and project success rate index (PSRI), that consider whether or not the campaign meets the specificity of a crowdfunding platform.

In addition to considering the current readiness level of the project, this approach also proposes metrics for analyzing and assessing innovative research projects in order to provide more profound insights into their strengths and weaknesses. These metrics take into account aspects such as uniqueness of innovation, how well the team can adapt to feedback and setbacks, and how congruent the project is with emerging market trends.

To demonstrate that our approach is effective, an experiment that involves developing an ensemble machine learning classification model using data from prior successful or unsuccessful crowdfunding campaigns of projects was carried out. This experiment not only shows how our assessment framework can be practically applied but also explains how different levels of readiness across the abovementioned eight parameters may affect the crowdfunding success of a project. This study provides the means to identify areas that still need to be worked on in order to increase acceptance by those who fundraise online.

Building on the foundational research described, our study was significantly influenced by feedback received from investors during project presentations at various national and international business events and/or accelerators like How to Web, Innovation Labs, and ADRVest Accelerator. These investors pointed out that the projects developed by student teams were often not fully prepared to meet the stringent demands of venture capital funding. They suggested that engaging in donation-based or reward-based crowdfunding campaigns could serve as a crucial preparatory step. This would allow for project teams to refine their projects based on public feedback and prove their concepts in a less risky environment before proceeding to more formal investment discussions or equity-based crowdfunding. This pragmatic advice from seasoned investors guided our focus towards understanding and analyzing the backer’s perspective, which notably differs from that of traditional investors. In response to this insightful feedback, we have tailored our research to prioritize project readiness from the proponent’s viewpoint, which is directly influenced by potential backers’ perceptions and requirements.

Such an extensive evaluation can assist in navigating the complex process associated with launching an innovative research project into a new market; this is precisely what our investigation is aimed at achieving with regard to crowdfunding readiness. By highlighting critical dimensions of readiness and introducing novel metrics, this approach helps to appreciate every single project’s potential for attracting funds from crowdfunding while providing grounds for further growth and acceptance within markets.

2. Literature Review

2.1. Crowdfunding—An Overview

Contemporary fundraising strategies include crowdfunding, which relies on internet platforms and social media to gain financial assistance from a large number of people, often in the form of small contributions [4]. This technique supports individuals, entrepreneurs, and institutions in learning environments to raise money for various projects, activities, or philanthropic causes. In contrast to conventional means of obtaining finance, crowdfunding is a method that allows for creators—including those from academic settings—to interact directly with potential supporters through the use of the internet. It comes in different forms: donation-based, rewards-based, equity-based and debt-based, each with unique incentives for contributors [5]. In addition, this approach bridges the divide between academia and the market by not just ensuring that innovative research projects are taken into account but also by educating project initiators about how products can be marketed as well as how communities may be rallied behind them. Crowdfunding plays an important role in democratizing access to capital while fostering communal support mechanisms necessary in realizing educational goals, thus improving both financial viability and making educational improvement possible for innovative research projects.

2.2. Statistics and Facts on Crowdfunding

This subsection highlights how active and vibrant the crowdfunding market is. Recent data indicate that things are improving in this sector, with a prediction that its total value could reach USD 300 billion by 2030 [6], showing the rise in its importance as a source for financing projects. Moreover, it estimates that over 12 million new crowdfunding campaigns will be launched by 2028 [7]. This shows that many individuals and organizations have increasingly found this approach useful in attaining their goals.

There has been massive growth in crowdfunding activity worldwide in recent years, with approximately 6,455,080 campaigns that have taken place all over the world. This raises awareness about both the wide adoption of this method of fundraising and the democratization of funding. It offers channels through which persons can receive endorsement and obtain funds for their businesses [8].

For successful crowdfunding efforts, backers play an important role. There are generally around 47 contributors per campaign on average, but some fully funded ones can count more than 300 backers, indicating their critical support from the crowdfunding community. The figures are important to success in these processes [9]. The projected compound annual growth rate (CAGR) from 2024 through to 2028 for transaction value is estimated at 1.43%, amounting to USD 1.27 billion to be realized by then.

There are several factors that determine how effective or ineffective a particular crowdfunding effort can be, such as platform selection and strategies employed to engage potential supporters among others. Successful outreach efforts made before launching any campaign strongly affect whether one will achieve their goal or not. In terms of donations from email shares, 53% reach campaign targets, while only every eighth share on social media platforms adds money to them [8].

Moreover, crowdfunding plays a large role when it comes to providing finances to numerous innovative ideas across various industries and sectors globally. It was estimated that the transaction value of the crowdfunding market would reach USD 1089.7 million by 2021, while the projected compound annual growth rate was expected to be around 2.63% until 2025 [10]. On top of this, social media interactions not only enhance customer engagement but also drive project development, since user involvement helps with adding new features and making improvements.

According to recent demographic studies, those aged between 24 and 35 are more likely than people over 45 to participate in crowdfunding campaigns [11].

2.3. The Reasons for Crowdfunding Campaign Failure

It is common for crowdfunding campaigns to fail due to a variety of obstacles that impede their ability to meet funding targets and achieve overall success [12]. The main reason is a project launching without enough market research, which leads to products or services not attracting the market [13]. Additionally, many startups have no coherent plan or strategy, thereby making their efforts disorganized as well as creating unclear communication. Failure to provide an attractive value proposition can severely diminish backer interest, underpinning the requirement for lucid, succinct, and persuasive communication to effectively engage potential supporters.

Meanwhile, backers may be put off if they perceive unrealistically ambitious target figures as unattainable [14].

Most of the campaigns’ information is disseminated through the marketing strategies employed; thus, marketing has a huge impact on the success of the crowdfunding campaigns. When there is a lack of marketing, such as poor-quality social media accounts, the visibility of the campaign, and hence the funds that the campaigns can receive, is reduced. Research shows that social media helps in increasing a campaign’s chances of success, particularly for small and medium-sized enterprises [15]. Good social media strategies not only create visibility but also make it possible for potential backers to interact with the project, creating confidence in community members.

Furthermore, information asymmetry is another issue that the campaigns face, and this has a negative impact on bringing in backers. There are different aspects, like incomplete information, ambiguous project timeframes, or not enough information regarding the project team, that make potential supporters doubtful about investing their money [16].

Lasting accomplishments, industry background, and third-party references are signals that may augment the credibility of a project and of the entrepreneur’s good faith [16]. It has been demonstrated that when endorsements by third parties are available, there is an increase in trust and credibility, which are fundamental attributes for tackling backer hesitation [17]. The absence of these elements may lead to a loss of trust during campaigns, which implies low levels of funding and engagement.

In addition, temporal elements with respect to crowdfunding campaigns, which include the provision and updating of information, are also very important to backers’ attention and commitment [18]. Entrepreneurs must navigate these temporal commitments carefully to sustain momentum and avoid the pitfalls of entrepreneurial hype, which can lead to disillusionment among backers if expectations are not met. Thus, a comprehensive approach that combines effective marketing, credible information dissemination, and strategic timing is essential for the success of crowdfunding campaigns.

Moreover, launching a campaign at less opportune times like holidays or major events can also negatively affect its visibility and support [6]. Little interest arises from products that do not match with market requirements or address specific customer pain points. Similarly, backer rewards that are poorly designed may fail to attract contributors.

Furthermore, highly competitive markets where there is low engagement with prospective backers such as slow responses to questions or comments by entrepreneurs may also lead to reduced success rates. Before embarking on a campaign, it is important that one has all loose ends tied up; otherwise, initial traction will be needed [11]. Backers might become discontented upon encountering over-promising during campaigns or during delivery of the product or service due to unforeseen difficulties.

Additionally, non-compliance with legal standards can oftentimes result in termination of the campaign process itself, along with intellectual property rights violation, thus bringing about lawsuits [19]. Moreover, internal conflicts within the team can expose a lack of experience or skills, as examples of some of the risks that seriously affect the success of the campaigns.

Regional factors, such as local economic conditions, technological infrastructure, and community support, play a crucial role in shaping the outcomes of crowdfunding campaigns. By analyzing data from various crowdfunding platforms, Ref. [20] reveals that regions with robust technological ecosystems tend to foster more successful high-tech crowdfunding projects and contribute to understanding how geographical context influences entrepreneurial finance, emphasizing the need for tailored strategies that consider regional dynamics in crowdfunding efforts, the results being significant for policymakers and entrepreneurs aiming to enhance the effectiveness of crowdfunding in high-tech sectors.

Finally, external factors such as economic recession, legal changes, and global crises can all highly impact a campaign, pointing to the fact that running a successful crowdfunding initiative is complex and multifaceted.

In addition to the challenges specific to individual crowdfunding campaigns, there are significant issues associated with the platforms themselves. Research presented by Cumming et al. [21] examines the widespread occurrence and detrimental impact of fraud within these platforms. The study reveals how fraudulent activities not only dissuade investors but also cultivate a pervasive mistrust that threatens the platforms’ long-term viability. It strongly advocates for the establishment of stricter regulatory frameworks and the enhancement of vetting processes. Such measures are essential for mitigating risks and boosting investor confidence, thereby bolstering the overall security and reliability of crowdfunding platforms.

2.4. Particularities of Crowdfunding for Research Projects

Crowdfunding emerged as a feasible alternative source of funding for research projects, especially in German Public Research Organizations (PROs) and universities [2]. This change brings unique opportunities and challenges, differentiating it from traditional funding approaches.

Moreover, crowdfunding offers not only financial resources but also enhances scientific communication, knowledge transfer, and technology. This is particularly advantageous for research projects that may not attract conventional funding due to their novelty, scale, or risk level. For instance, Experiment.com in the U.S. [1] and Sciencestarter in Germany provide dedicated environments for scientific projects to acquire public support without going through the usual bureaucratic and highly competitive traditional motions of obtaining funds. These platforms allow for researchers to directly engage with the public, thereby making science more accessible to the broader public. Furthermore, the “scientific cooperative crowdfunding” model suggested in the literature postulates the strategic integration of crowdfunding into PROs’ and universities’ financing mix, which could enhance typically underfunded knowledge/technology transfer steps.

However, despite these opportunities, there are various challenges that face crowdfunding for research purposes. The major concern is about the relatively small amounts collected through these channels, which cannot fully finance large, resource-intensive projects. Also, successful campaigns require significant effort in terms of communication and engagement, which diverts scientists from doing their primary scientific work.

The success of crowdfunding within science depends on how well researchers can explain to non-specialists what they have done and its significance beyond academia—not only clear presentation but ongoing dialogue with backers too. Those projects that manage to grab public imagination while demonstrating tangible benefits are closer to reaching their goals for the amount of funding needed.

Additionally, there are significant disadvantages associated with regulatory and bureaucratic aspects. Therefore, crowdfunding adoption has been slow in Germany due partly to regulatory barriers among other challenges together with poor incentives as well as lack of experience among PROs/Universities management regarding campaign management over such an initiative. It is also difficult to align crowdfunding with the strict ethical and operational norms that guide academic research work, which makes the direct use of this model difficult. This applies to all EU countries.

The integration of crowdfunding into the broader funding landscape represents a significant yet complex development in financing academic research projects. This alternative funding mechanism is particularly beneficial for projects that are too unconventional or at too early a stage to attract traditional capital investments, such as those from venture capital funds or university incubators, which typically select projects with high growth potential and clear pathways to substantial returns. In contrast, crowdfunding platforms deal with a wider array of innovative efforts, often embracing projects that defy traditional investment criteria and may not yet demonstrate their full commercial viability. This form of funding serves as a platform for validating public interest and potential market engagement, offering researchers preliminary feedback that can be important for further development and refinement. As crowdfunding becomes more embedded in the funding ecosystem, it is imperative for educational and research institutions to develop strategies that mitigate associated risks while capitalizing on the unique opportunities it presents. This involves adapting to the nuances of crowdfunding and fostering a shift towards more open, interactive approaches to scientific communication and public engagement, thereby enhancing the overall readiness of projects to succeed in this dynamic environment.

2.5. AI in Crowdfunding: Enhancing Prediction and Engagement

The use of artificial intelligence (AI) on crowdfunding platforms has changed the way campaigns are developed, executed, and assessed in significant ways. Studies carried out at different times through various authors show that AI is essential to predictive analytics as well as engagement strategies that enhance successful crowdfunding.

Machine learning models have increasingly been used in predicting the outcomes of crowdfunding campaigns. Predictive modeling techniques such as logistic regression analysis, the random forest algorithm, and extreme gradient boosting (XGBoost) have been applied on large datasets to forecast campaign results with considerable levels of accuracy [22,23]. Principal component analysis (PCA) is one of the advanced methodologies employed by these studies for dimensionality reduction—among others, which include log transformation—aimed optimizing the performance of predictive models [22]. By doing this, it allows for campaigners to tune their strategies using strong knowledge that is supported by substantial data.

AI-driven tools are both predictive and prescriptive; they recommend better ways for campaigns to interact with potential backers. Artificial intelligence allows for personalized communication strategies and effective audience segmentation using backer behavior and preferences during the campaign lifecycle [24,25]. Additionally, AI can optimize the timing and content of updates, making interactions more impactful and targeted at specific audiences’ interests [26,27].

Despite its advantages, AI application in crowdfunding must be approached carefully to handle notable ethical as well practical issues. These concerns include data privacy and algorithmic bias, as well as transparency in AI decision-making procedures that must be addressed properly to preserve trust and fairness within crowdfunding platforms.

The future role of AI in crowdfunding may incorporate blockchain technology, among others, to improve its transparency and security [28,29]. Furthermore, ongoing developments in machine learning might position AI more centrally in dynamic real-time campaign management.

The deployment of artificial intelligence technologies in crowdfunding signifies a revolution in contrast to traditional campaigning approaches. The possibilities for an open revolution caused by AI range from “Predictive analytics enhance campaign planning” to “Backer interaction is boosted by AI-driven engagement strategies”.

3. Materials and Methods

This section provides the step-by-step procedure that we followed in assessing if a research project is ready for crowdfunding. Such an assessment requires a comprehensive evaluation of several key factors, like the maturity level of technology, market preparedness, and commercial viability [30,31]. The various elements in this regard can be analyzed in detail by different stakeholders to single out what the strengths and weaknesses are, as well as risks that may arise in connection with such projects.

To perform this valuation, we employ both quantitative as well as qualitative methods so that a more inclusive view of the potential of each individual scheme is provided. For assessing technology maturity, data collection takes place through surveys, interviews, and market analysis, from which the degree of technological advancement and readiness for public use is determined. When evaluating market readiness, it is required to observe issues like customer receptiveness in a given target market as well as competitive landscape. Finally, financial forecasts and business model assessments are used to determine commercial viability.

Once gathered, the data are then processed and analyzed so as to realize what state any individual project is in regarding readiness. Through such rigorous evaluation processes, informed decision-making as well as strategic planning are facilitated, hence ensuring seamless transition from the idea stage up to implementation. In summary these evaluations help ensure optimal utilization of resources while at the same time maximizing outcomes from projects and achieving objectives in relation to technology development and innovation. This systematic approach ensures that only proposals with a high chance of success find their way onto crowdfunding platforms.

3.1. Techniques for Assessing Readiness

When analyzing research projects’ eligibility for equity crowdfunding, researchers make an elaborate examination starting from the idea’s birth until it becomes mature enough for sale, thus avoiding possible failures. One of essential tools applied during this process includes technology readiness levels (TRLs), which were originally introduced by NASA. TRLs serve as a standard way to measure technological maturity at different levels ranging from one to nine [32]. They are designed to serve governance purposes across technology, tailoring its applicability to engineering terms in the process of technology maturity. These levels provide a pathway through the research, development, and deployment of technology from concept to market, with universities typically focused on TRLs 1–4 and businesses on TRLs 7–9.

While TRLs mainly evaluate technological aspects, the commercial readiness level (CRL) was introduced to assess market and commercial factors, aiding in the commercialization of technology by evaluating commercial viability on a scale of one to nine [33,34].

The Cloverleaf framework combines historically separate methods for measuring TRLs and CRLs [35]. This device was created to ensure that no one factor dominates an evaluation and that there is balanced assessment across various areas. The Cloverleaf model contains critical dimensions like readiness for change, which identifies individuals who are willing to adopt or have adopted new technologies both at home as well as professionally. Design issues also influence this level of readiness [36,37,38].

Market readiness entails preparing a product for launch, engaging with potential users to ensure the product meets their needs and enhances task efficiency. It characterizes technologies that offer discernible advantages over competitors, highlighting the importance of a robust support structure for growth and stability [39].

Commercial readiness mirrors the developmental stage of the commercialization business model, focusing on customer satisfaction—a prerequisite for mainstream market transition and growth. This dimension encapsulates market evaluation outcomes, considering competitor products’ pricing and consumer attributes, and assesses the technology’s market penetration readiness [40,41,42].

The last point is management readiness, which involves evaluating the ability of the management team to lead and develop decent expectations; knowing which product type the organization commercializes more than others and patenting some; and being recognized in the industry [41]. It includes activities like organizing and integrating plans meant for project implementation.

Using the Cloverleaf framework for a comprehensive evaluation as such, projects are considered from various perspectives to determine their readiness for crowdfunding and subsequent introduction into the market. This systematic approach ensures that projects meet not only technical specifications but also have commercial viability and good management.

3.2. Key Characteristics for Research Project Crowdfunding Readiness Evaluation

A successful crowdfunding campaign is built on a combination of strategic storytelling, communication, and engagement techniques [19]. Then, these key attributes are employed in designing new metrics that can be used to gauge readiness for crowdfunding in research based on similar successes in past campaigns.

Clear and compelling story (CCS): A successful campaign begins with a gripping narrative that is clear-cut, concise, and emotionally engaging, highlighting what exactly drives it home. Such a manner of storytelling provides an emotional touch between potential supporters/backers, making them become part of its success story.

Transparent goals and planning (TGP): Trust is built upon transparency. Funding objectives should be stated clearly with an articulated budgetary plan specifying how the money will be used within project execution framework.

Engaging video and visuals (EVVs): A well-produced video acts as an effective tool towards enhancing successful campaigns since it becomes the first contact point for interested funders. Supportive images like high-quality pictures or informative graphics make this narration more applicable and understandable.

Defined and attractive rewards (DARs): Having different levels of rewards that are attractive enough to motivate backers may serve as incentives to support a crowdfunded research project. They should appeal to many individuals plus providing value while being economical for the project.

Active and transparent communication (ATC): Regular two-way communication with campaign backers is key. Responsiveness to inquiries, updates on progress, and involving them in the course of the project are all aimed at creating a “crowd” that trusts in and supports the project.

Strategic timing and duration (STD): This is to make the campaign successful when the target market is available. The period must be carefully calculated, taking into consideration how long it will maintain momentum without becoming a prospective backers’ nightmare.

Robust marketing and promotion (RBM): Successful marketing tactics are necessary to ensure that the campaign reaches as many people as possible. Social media sites can be used alongside press releases, influencer collaborations, or even using several adverts, which go beyond only raising money via crowdfunding platforms.

Authenticity and credibility (AC): Establishing credibility is important. Team competence, milestones achieved in previous instances, prototypes where these exist, endorsements from recipients, or news stories printed may give prospective funders confidence.

Flexibility and adaptability (FA): A campaign’s flexibility towards feedback acceptance makes it more effective. Changing what needs to be changed while running a campaign could help increase its viability.

These characteristics are found in successful crowdfunding campaigns that do not merely raise funds but also surpass expectations through an effective storyline, direct dialogue techniques, appropriate timing for soliciting contributions on a wider scale, and market presence established over time by engaging users proactively.

3.3. Project Readiness Characteristics

To evaluate crowdfunding readiness, we first define four key readiness characteristics: technological readiness for a crowdfunding campaign (TRC)—measures the technological preparedness of the project for crowdfunding; marketing readiness for a crowdfunding campaign (MRC)—assesses the readiness of marketing strategies and materials to support the crowdfunding campaign; commercial readiness for a crowdfunding campaign (CRC)—evaluates the market potential and commercial strategy of the project; and management team readiness for a crowdfunding campaign (MTRC)—looks at the readiness and capability of the team managing the crowdfunding project.

In our machine learning dataset, nine specific characteristics—clear and compelling story (CCS), transparent goals and planning (TGP), engaging video and visuals (EVVs), defined and attractive rewards (DARs), active and transparent communication (ATC), strategic timing and duration (STD), robust marketing and promotion (RBM), authenticity and credibility (AC), and flexibility and adaptability (FA)—were assessed. These characteristics were evaluated using a structured questionnaire, provided in Appendix A. Additionally, the four dimensions of crowdfunding readiness—technological readiness (TR), market readiness (MR), commercial readiness (CR), and management team readiness (MTR)—were analyzed using the Cloverleaf method, detailed in Appendix B.

Before introducing our formula, it is pertinent to state that the method developed by Heslop et al. quantifies judgments for each criterion within the four areas of the Cloverleaf framework. This method is designed to meticulously gauge the extent to which each condition is fulfilled, emphasizing the necessity of meeting all components across the framework’s four “leaves” to confirm a technological concept’s readiness for commercialization. Building upon this foundation, our study enhances this model by integrating an additional nine characteristics specifically tailored to the nuances of crowdfunding campaigns. This expansion enriches the differentiation among the four fundamental dimensions—technological readiness, market readiness, commercial readiness, and management team readiness. The following determination formulas for each characteristic quantitatively assess various aspects of a campaign, adjusted by specific readiness percentages and factors.

$$TRC=\underset{i=1..n}{\mathrm{avg}}{AC}_i·{\displaystyle \frac{TR}{100}}·{\displaystyle \frac{TRL}{10}}$$

TRC is calculated by averaging the authenticity and credibility scores (AC) of all respondents (i ranging from 1 to n), then scaling this average by the technological readiness percentage (TR) and the technological readiness level (TRL). This ensures that the credibility of the project team and their technological maturity are integrated into the campaign readiness measure.

$$MRC=\underset{i=1..n}{\mathrm{avg}}\left({CCS}_i+{EVV}_i+{DAR}_i+{RBM}_i\right)·{\displaystyle \frac{MR}{100}}$$

MRC combines scores from multiple characteristics: clear and compelling story (CCS), engaging video and visuals (EVVs), defined and attractive rewards (DARs), and robust marketing and promotion (RBM). The average score is then scaled by the marketing readiness percentage (MR), reflecting how well the campaign communicates value and engages backers.

$$CRC=\underset{i=1..n}{\mathrm{avg}}\left({TGP}_i+{STD}_i\right)·{\displaystyle \frac{CR}{100}}$$

CRC averages transparent goals and planning (TGP) and strategic timing and duration (STD) scores for all respondents, then multiplies this by the commercial readiness percentage (CR). This metric captures how well-defined goals and optimal timing contribute to the project’s market viability.

$$MTRC=\underset{i=1..n}{\mathrm{avg}}\left({ATC}_i+{FA}_i\right)·{\displaystyle \frac{MTR}{100}}$$

MTRC averages active and transparent communication (ATC) and flexibility and adaptability (FA) scores, which reflect the team’s responsiveness, openness to feedback, and adaptability. This average is scaled by the management team readiness percentage (MTR) to evaluate their preparedness to handle backer interactions and adapt as needed during the campaign.

Each characteristic is normalized on a scale similar in range for consistent integration into the model. The number of respondents, n, for the questionnaire provides the base data for these calculations.

Also, the importance of selecting the appropriate crowdfunding platform is explored in the context of maximizing the success potential of a project. The choice of platform is influenced by several dimensions evaluated using specific criteria by the following coefficients: i₁ =PCAI (platform–campaign alignment index), i₂ = FEI (financial efficiency index), i₃ = CEI (campaign engagement index), and i₄ = PSRI (project success rate index).

Calculating the values for each of these coefficients involves selecting the most advantageous project classification from the two prominent crowdfunding platforms, Kickstarter and Indiegogo. This decision process is pivotal because each platform has distinct characteristics and audiences that may be more conducive to the success of certain types of projects. For instance, Kickstarter is renowned for its strong community support for innovative and creative projects, particularly in technology, design, and the arts. Indiegogo, on the other hand, offers a more flexible funding model that can be beneficial for a wide array of projects, including those that are ongoing or require access to immediate funds regardless of whether the full funding goal is met.

The i₁ index measures how closely the theme and type of the campaign align with the typical projects hosted and successfully funded on a platform (Kickstarter or Indiegogo, as for our study), showing results from the analysis of the platform’s types of projects that are commonly successful. Each category is scored on the platform based on the degree of alignment, according to past successful campaigns and thematic relevance.

$$i_1={\displaystyle \frac{No.success project per category}{Total no. of projects in category}}$$

The values from the above equation are taken from Appendix C (Figure A1 and Figure A2).

Index i₂ assesses the cost-effectiveness of launching a campaign on a platform considering the typical fundraising success rate for campaigns of similar category. It is calculated based on the average net funds raised for similar campaigns and comparing across platforms, taking into account average donation size for similar campaigns and the success rates of these campaigns.

i₂ = average value of category/project value (upper limited at 1)

$$i_2=\left\{\begin{array}{c}{1, if V}_{project}\le V_{avg}\\ {\displaystyle \frac{V_{avg}}{V_{project}}}, else\end{array}\right.$$

where V_project = the project’s proposed value

V_avg = average of target values for each category’s successful projects (extracted from Appendix C, Figure A3 and Figure A4).

The i₃ index measures the level of engagement that campaigns generate, which can be indicative of a platform’s ability to attract and retain active backers who contribute meaningfully to campaigns. Average contribution of backers per category in successful campaigns, which reflects their engagement on the platform.

$$i_3=\left\{\begin{array}{c}{1, if V}_{reward project}\le {\displaystyle \frac{V_{avg\_raised}}{{no.backers}_{avg}}}\\ {\displaystyle \frac{\frac{V_{avg\_raised}}{{no.backers}_{avg}}}{V_{reward project}}}, else\end{array}\right.$$

where

• V_{reward project} = value proposed for the reward;
• V_{avg_raised} = average of raised funds of successful projects from each category;
• no. backers_avg = average number of backers of successful projects from each category.
• ${\displaystyle \frac{V_{avg\_raised}}{{no.backers}_{avg}}}$ values are in Appendix C (Figure A5 and Figure A6).

The i₄ index evaluates the success rate of projects on the platform, giving an indication of how effective the platform is in helping projects reach their funding goals. It reflects the amount exceeded of the funding goals for successful projects per category.

$$i_4=\left\{\begin{array}{c}{1, if V}_{project}-V_{avg}\le {\left(V_{raised}-V_{goal}\right)}_{avg}\\ {\displaystyle \frac{{\left(V_{raised}-V_{goal}\right)}_{avg}}{V_{project}-V_{avg}}}, else\end{array}\right.$$

where

${\left(V_{raised}-V_{goal}\right)}_{avg}$ = average of amounts exceeded for successful projects from each category (values available in Appendix C, Figure A7 and Figure A8), V_raised = value raised of each successful project, and V_goal = value targeted of each successful project.

The readiness characteristics and platform analysis coefficients are integrated as inputs into the machine learning dataset.

3.4. Data Preparation and Model Development for Crowdfunding Readiness Assessment

According to the research methodology, Figure 1 is an all-inclusive representation of how we developed the model for predicting whether a crowdfunding project will be successful or not. It is based on machine learning algorithms and includes such steps as literature analysis, model training, and testing on new data.

The first part of this research delves into the details of the literature review, helping to establish the features behind successful crowdfunding. This information helps in shaping subsequent methods used for collecting data, which are supplemented with qualitative and quantitative approaches. The selection process for Indiegogo and Kickstarter projects is careful enough to cover sufficient cases in order to obtain a complex and useful dataset. It narrows down to specific features that matter in view of scholar insights as well as practical observations, evaluated based on the nine success characteristics, the four dimensions of crowdfunding platforms from Cloverleaf model, and the four relevance coefficients of platform suitability. However, it is important also to let different students’ projects pass through the same filters in order to derive new data for model testing.

Once we gathered data from Indiegogo and Kickstarter campaigns, the data underwent extensive preprocessing to ensure their quality and consistency, which are crucial for effective model training. This step is essential, as any errors or inconsistencies in the data can introduce bias, potentially leading to overfitting in our predictive modeling. The statistical analysis performed at this stage helps in refining the data further, revealing meaningful patterns and relationships that are instrumental in selecting the most appropriate model for our analysis. This thorough preparation of the data ensures that the insights derived are based on reliable and accurately processed information, enhancing the robustness and reliability of our predictive model.

The main body of methodology lies in developing ensemble machine learning model. Random forest provides robustness against overfitting along with its ability to work well in high-dimensional spaces. To decrease bias while enhancing learning from previous errors, gradient boosting is applied. Logistic regression supplements probabilistic outputs thereby permitting classification of projects into three predefined categories: not ready, needs improvement, or fully prepared.

The suite of features selected for incorporating them into the model include technological readiness (TRC), market connection (MRC), commercial readiness (CRC), and management team readiness (MTRC), together with an additional 4 metrics derived for platform suitability evaluation: platform–campaign alignment index (PCAI), financial efficiency index (FEI), campaign engagement index (CEI), and project success rate index (PSRI). Both empirical findings and academic literature review informed these features, which mean they are relevant besides being highly important to the predictive accuracy of the model.

After the model had been trained, its performance was evaluated on new data collected from student project evaluations, which helped in generalizing this model’s efficiency to real-world scenarios. Precision, recall rates, as well as F1 score, among others, are some of the key metrics used to assess how accurately the model classifies new projects according to their readiness for crowdfunding.

This kind of a holistic approach not only makes for a strong predictive tool but also adds to understanding the dynamics in crowdfunding. Consequently, it will be useful for stakeholders who intend to start or manage crowdfunding campaigns; as such, these practices may improve the chance of success in a tough market.

4. Experiment

The experiment consisted of developing a dataset and a machine learning model for predicting the crowdfunding readiness level of projects resulting from academic research. The dataset consisted of two parts, one used for training the model and the other for testing on new, unseen data.

4.1. Training Dataset

The experiment from this study encompasses a detailed data collection procedure focusing on crowdfunding projects from Kickstarter and Indiegogo. Comprehensive datasets were sourced from Kaggle, specifically the “kickstarter_projects.csv” for projects spanning 2009 to 2018 and the “Crowdfunding_Indiegogo_Dataset.csv” for projects from 2010 to 2023. These datasets have enough information for analysis based on the metrics defined earlier (TRC, MRC, CRC, and MTRC), enabling an examination of these projects using a metrics system that was built upon previously mentioned formulas focusing on key success characteristics and the dimensions of the Cloverleaf method.

The analysis and evaluation covered 6864 projects, selected from the above datasets. The entire process spanned over two academic years, from 2022 to 2024. This extensive evaluation involved seven groups of students, each consisting of 15 undergraduate and master’s students enrolled in Entrepreneurship and Project Management subjects, coordinated by 2 doctoral students per group. This strategy was adopted to ensure uniformity and objectivity in the data collection process. Notably, there were no self-evaluations involved. The doctoral students applied the Cloverleaf method and determined the coefficients i₁ through i₄ for each project.

The outcomes of the evaluations conducted by the seven student groups and the doctoral students were integrated into the dataset as eight distinct features. Additionally, a ninth column representing the label was added to each project in the dataset, based on the success rate characteristics derived from the two Kaggle datasets for the evaluated projects.

Before utilizing the resulting dataset for training machine learning models, the first step involved identifying and removing erroneous data, specifically those outside the acceptable ranges for each feature. This process led to the elimination of 572 data points, resulting in a cleaned dataset of 6292 samples, which was subsequently analyzed to determine suitable learning algorithms based on the structure and distribution of the data. This methodological framework ensures a robust and reliable foundation for further analysis and model development.

The analysis of distribution within our datasets showed that the data did not conform to normal distributions, as resulted from a Kolmogorov–Smirnov (KS) test. The statistical analysis results for the dataset, particularly the normality tests conducted using the KS test, indicate that none of the features conform to a normal distribution. This conclusion is drawn from the p-values associated with each feature, which are significantly lower than any conventional alpha level (e.g., 0.05, 0.01), suggesting a rejection of the null hypothesis that the data are drawn from a normally distributed population.

The KS statistic values for each feature range from 0.092 to 0.108, indicating a moderate deviation from the normal distribution. The exceptionally small p-values, often approaching zero (e.g., 2.07 × 10⁻⁶³ for Feature 1), further underscore the non-normality of the data across all features. Such findings are important, as they inform the selection of appropriate machine learning algorithms that do not assume normality in data distribution.

The histograms from Figure 2 provide a visual distribution of eight different features analyzed from our dataset, showing a range of frequency distributions for each feature across the specified value range. Notably, the histograms for features TRC, MRC, CRC, and MTRC demonstrate fairly uniform distributions across several value intervals, with the feature MRC displaying a notable decrease in frequency at higher values. This indicates less variability in the data for these specific features and a uniform distribution of these characteristics among the projects analyzed.

Features i₁ through i₄ show more variability in their distributions. Features i₁ and i₃ exhibit particularly interesting patterns with significant peaks at specific intervals, suggesting thresholds that are significant for interpreting project success. Features i₂ and i₄ also show varied distributions but are somewhat more evenly spread than i₁ and i₃, although with visible peaks. These visual insights are critical for understanding the underlying characteristics of the crowdfunding projects in our study. The variations in the distributions are indicative of differing levels of impact these features have on the success of crowdfunding campaigns.

The class distribution in the dataset reveals a slight imbalance among the three classes, with Class 2 having the highest count of 2396, followed by Class 3 with 2041, and Class 1 with 1855 (Figure 3). This imbalance could potentially influence the performance of machine learning models, particularly those sensitive to class distribution such as logistic regression or linear discriminant analysis.

Given the non-normality of the dataset and the class imbalance, we used machine learning models that can handle such data characteristics effectively. For instance, tree-based models such as decision trees, random forests, or gradient boosting machines are more appropriate, as they do not require the data to be normally distributed and are less sensitive to an imbalance in the dataset.

The next step meant determining the suitability of the data for learning by evaluating whether the features selected were significant predictors of the outcome. This step was important to prove that all variables are impactful, thus ensuring the learning efficiency and prediction accuracy of the models. Through this evaluation, suitable algorithms were identified, which align with findings from two research studies [22,23]. These studies underscore the effectiveness of the random forest and gradient boost algorithms in contexts similar to ours, particularly emphasizing the robust performance of these methods in scenarios involving complex, nonlinear data relationships.

The scatter matrix in Figure 4 provides a visualization that allows for a detailed examination of the relationships between the eight features, divided into two groups: TRC, MRC, CRC, and MTRC (representing different readiness components), and i₁, i₂, i₃, and i₄ (indices indicating other critical platform related metrics). The first four features—TRC, MRC, CRC, and MTRC—reflect various readiness aspects of crowdfunding projects and show relatively uniform distributions as evidenced by the histograms diagonally along the matrix. This uniformity suggests that these features, while important, do not exhibit extreme variability or skewness, indicating that they capture fundamental aspects of project readiness that do not highly vary across the dataset.

In contrast, the indices i₁ through i₄ show more varied and multimodal distributions, which imply that these features capture more specific attributes of the projects that are sensitive to the project’s context and the crowdfunding platform. These features’ distributions suggest that there are distinct subgroups within the dataset, corresponding to different types of projects or varying strategies within the crowdfunding campaigns. For instance, peaks in the histograms for features like i₁ and i₃ indicate common strategies or attributes shared by groups of projects, which either succeed or fail together.

Looking at the off-diagonal elements of the scatter matrix, the pairwise scatter plots between TRC, MRC, CRC, and MTRC demonstrate little to no obvious correlation, suggesting that these readiness components contribute independently to the overall project readiness without duplicating information. This independence is important for model robustness, as it implies that each feature provides unique information about the project’s state, enhancing the predictive power of a model that includes these variables.

However, the scatter plots involving the indices (i₁ through i₄) with each other and with the first four features (TRC, MRC, CRC, and MTRC) reveal more complex patterns. Notably, certain combinations of indices show clustering, which could indicate interaction effects where the combination of specific project characteristics leads to distinctly higher or lower chances of crowdfunding success. These interactions are particularly evident in plots involving i₁ and i₃, which show more defined clustering patterns, suggesting that these features interact in a way that significantly impacts project outcomes.

The scatter matrix’s insights underscore the need for clever modeling techniques that can capture both the independent contributions of readiness components and the complex interactions among more nuanced project indices. Models such as random forests or gradient boosting machines, which are capable of handling complex and nonlinear relationships, are particularly well suited for this task.

So, the scatter matrix reaffirms the relevance of each feature in predicting crowdfunding success while highlighting the importance of considering feature interactions in the modeling process. This analysis provides a solid foundation for developing a predictive model that is both detailed and robust, capable of accurately forecasting crowdfunding outcomes based on comprehensive project data.

4.2. Testing Dataset

In order to test the resulting model on new unseen data, an additional assessment was carried out on projects developed by student teams at the University of Petroșani, which started as early as 2018 and reached the minimum viable product (MVP) stage by 2022. Initially, 145 projects were considered and tracked, of which 24 reached the MVP stage. These projects were randomly assigned to the same working groups for analysis. The evaluators, who were students, did not participate in the development teams of the projects they assessed. During the evaluation process, each student within the evaluation group categorized the project into one of three classes: 1, 2, or 3. The label assigned to the project was determined as the rounded average of the votes from the 15 evaluators. This approach resulted in the completely new test dataset.

The structure of the unseen dataset designed for testing the developed model on crowdfunding readiness level reflects a diverse and well-distributed array of projects spanning various categories and technological domains. This dataset is carefully curated to ensure a broad representation of types and complexities associated with crowdfunding campaigns, like the ones on Indiegogo and Kickstarter. The projects are categorized according to the nature of the project and the platform specific technological domain it belongs to. The best-fitted domain was chosen from the two platforms for each project analyses. The projects fitted for Indiegogo were as follows: in the domain “Transportation”: five automotive projects, specifically, two small electric vehicles, one electric SUV, one electric kart, and one attachable device for wheels to assist movement on muddy or snowy terrain; in the “Wellness” domain: three projects focusing on health and wellness technologies, including two upper-body exoskeletons and a fall detection device for the elderly; and in the “Video Games” domain: a single project, a mobile app designed to encourage physical exercise, blending the worlds of fitness and gaming. The projects fitted for Kickstarter were as follows: in the domain “Wearables”: three projects, including a stress estimation bracelet, a tremor management bracelet, and a navigation device for the blind; in “Robots”: four projects involving two interactive robots and two mobile robots with legs; in the domain “Flight”: two drone projects focused on agriculture and wildfire detection; in “Gadgets”: three projects including interactive gaming devices and a hologram display device; in “Software”: a project management app; and in the “Experimental” domain: two highly innovative projects—a thematic park and an underground VLC (visible light communication) system. The projects are labeled according to their assessed readiness level for crowdfunding: Class 1 (not ready): fourteen projects; Class 2 (needs improvement): five projects; Class 3 (fully prepared): five projects. These labels are instrumental for testing the model in recognizing varying degrees of readiness, which directly impacts the potential success of crowdfunding campaigns. The diversity in project types and the comprehensive range of categories ensure a robust test of the model’s generalizability and accuracy across different sectors and project scales, as will be presented in Section 5. This structured approach to compiling the unseen dataset allows for a rigorous evaluation of the model’s predictive capabilities. By including a wide range of project types and readiness levels, the dataset challenges the model across different dimensions of complexity and mirrors the multifaceted nature of real-world crowdfunding environments.

5. Results and Discussion

The employment of an integrated ensemble model that uses random forest and gradient boosting techniques in our research on predicting academic research project readiness for crowdfunding was backed up by an extensive review of literature as well as comparative performance analysis findings presented in the most recent scholarly articles. Some of these articles emphasized the strength and efficiency of these models in dealing with complex datasets common to crowdfunding platforms, where factors leading to success are multiple, mixed, and interrelated.

Random forest outperforms other classifiers through the bagging technique [23], which performs a majority vote among many decision trees to come up with a final prediction, thus substantially reducing the risk of overfitting—a problem often encountered when developing predictive models. This learning algorithm is very good at handling large datasets with many input variables like those found in crowdfunding data from different categories of projects. The fact that this method can handle high-dimensional data without significantly compromising its performance makes it ideal for our prediction task.

Contrarily, gradient boosting offers a powerful sequential improvement mechanism by building one tree at a time, whereby each new tree helps correct mistakes made by previous ones [22]. It has been found more useful than other methods under various conditions. In particular, feature selection techniques like principal component analysis (PCA) and transformations aimed at normalizing input data have been extensively used to demonstrate its effectiveness along with gradient boosting. Such changes are important for dealing with the scale variations among inputs that could lead to less robust models seen in crowdfunding data.

The combination of random forest and gradient boosting within an ensemble model brings together their strengths: reducing variance by random forest and lowering bias through gradient boosting. By so doing, it improves accuracy and ensures generalizability across different types of crowdfunding projects. Thus, this approach is consistent with the dynamics characteristic of crowdfunding platforms that attract projects from various sectors.

Strategic integration of these two algorithms based on insights gained from current studies ensure that our model is able to predict the success of crowdfunded projects with high reliability. This approach also aligns well with the larger objective of our research, which seeks to create a predictive tool that can act as a dependable decision support system for would-be campaigners, thereby optimizing their strategies and increasing their chances of getting funded.

5.1. Building the Ensemble Prediction Model

Creating an ensemble model that combines random forest and gradient boosting methodologies involves several steps designed to leverage the strengths of both approaches, thereby enhancing the overall predictive power of the model. This process follows a structured pipeline to ensure that each component contributes optimally to the final outcome. These steps are thoroughly presented in this section to ensure model reproducibility.

The ensemble modeling process begins with data preparation. The dataset is loaded from a CSV file, “cleaned_dataset.csv”, ensuring that all necessary preprocessing steps, such as handling missing values or normalizing data, are already completed. This dataset has eight feature columns and a single label column. The data are then split into training (80%) and test (20%) sets, to avoid overfitting and to ensure that the model is tested on unseen data, providing a reliable assessment of its predictive capabilities.

Once the data are prepared, the next step involves training two distinct models: a random forest and a gradient boosting classifier. Random forest is chosen for its ability to handle a large number of input variables and for its robustness against overfitting. It operates by constructing a multitude of decision trees during training and outputting the class that is the mode of the classes of the individual trees, providing a high level of accuracy. On the other hand, gradient boosting is employed for its proficiency in producing a predictive model in the form of an ensemble of weak prediction models, typically decision trees. It builds the model in stages like other boosting methods do, and it generalizes them by allowing for optimization of an arbitrary differentiable loss function.

In MATLAB, the random forest model is implemented using “TreeBagger”, which uses bootstrap aggregation (bagging) to form an ensemble of decision trees that vote for the most popular class. The gradient boosting model is set up using “fitcensemble” with the “AdaBoostM2” method, as the original “LogitBoost” does not support multi-class classification directly.

After the base models are trained, their predictions (or probability scores) on the test data are used as meta-features. This step essentially captures the initial predictions from both models, serving as input for the next level of modeling. This stacking technique leverages the distinct patterns recognized by each base model, aiming to correct the first level predictions in the subsequent model.

The meta-features are then used to train another model, this time a meta-model, which is aimed at effectively combining the predictions of the base models. In this case, a logistic regression model is used, implemented via “fitclinear” in MATLAB, which is suitable for binary classification. Given the complexity of multi-class categorization, “fitcecoc” is used instead to extend logistic regression for multi-class problems using the one-vs-all strategy.

Finally, the ensemble model’s performance is evaluated on the test set. Performance metrics such as accuracy, precision, recall, and F1 score are calculated to provide insights into the effectiveness of the model across different classes. These metrics are important for understanding the overall effectiveness of the model and how well it performs for each class, helping identify any biases or weaknesses in the model.

This systematic approach to building an ensemble model harnesses the combined strengths of random forest and gradient boosting, supplemented by logistic regression, to provide a robust predictive tool. By carefully integrating these techniques, the model can achieve higher accuracy and better generalization capabilities than any single model could on its own.

5.2. Ensemble Model Optimization

The efficacy of models like random forest and gradient boosting in the machine learning field are highly dependent on their respective hyperparameters. In this analysis, we chose two key parameters to improve the overall model performance: the number of trees in the random forest model and the number of learning cycles in the gradient boosting model.

The random forest algorithm is improved by altering the number of trees, a feature that has a great effect on generalization and overfitting. This parameter was tested with eight different values ranging from 100 to 600. These diverse values give an opportunity to investigate how smaller or larger forest sizes affect performance, thereby balancing between accuracy at training time and training duration.

In a similar vein, determining how well the gradient boosting model uses its training data depends on the value given to this hyperparameter. This parameter was varied over six values from 50 to 300 cycles. The depth of learning is directly influenced by this parameter; depending upon complexity levels of the data, it can help reduce problems such as underfitting at low values and overfitting at high ones.

Thus, optimal settings that provide maximum accuracy by systematically training ensemble models through every combination of these two parameters were identified. The resulted surface plot from Figure 5 offers a visualization of how different configurations of random forest and gradient boosting models influence the accuracy of the ensemble classifier.

A distinct highest accuracy point was reached at 200 trees and 200 learning cycles. At this point, the best model with accuracy of 98.94% was achieved, which means that a trade-off between the robustness of the random forest and precision of the gradient boosting is made at these settings.

The number of trees has an important role in capturing variance in data by a random forest. By increasing the number of trees, overfitting can be avoided, and generalization capability can be improved. However, after reaching a certain level, the improvement in accuracy becomes marginal, which may lead to unnecessary computational cost without any significant improvements in performance.

Similarly, the number of learning cycles in gradient boosting indicates how many sequential models are built, each one focusing on the errors of the previous models. More cycles lead to a better model but also increase the risk of overfitting, particularly if the number of cycles is too high compared to the complexity of the data.

At 200 trees and 200 learning cycles, an optimum point where both model complexity and performance are balanced effectively between overfitting and maximized predictive accuracy was encountered. This finding emphasizes that ensemble parameters should be adjusted sensitively according to specific properties of data so that both random forest and gradient boosting individually have high effectiveness as well as being complimentary when combined together as an ensemble.

This optimal configuration provides a valuable benchmark for deploying the model in real-world crowdfunding success predictions, where both accuracy and computational efficiency are important. It reflects a well-adjusted model, capable of delivering high predictive performance while maintaining manageable computational demands making it ideally suited for practical applications in predicting readiness level of crowdfunding projects.

Confusion Matrix

The Figure 6 presents the confusion matrix, for the above chosen parameters, that offers a detailed look at the performance of the classification model.

The major diagonal cells (584 for Class 1, 688 for Class 2, and 592 for Class 3) show the number of correct predictions made by the model. These values are substantially higher than the off-diagonal values, indicating a strong predictive performance. The percentages adjacent to these counts show exceptionally high accuracy rates—98.5% for Class 1, 98.4% for Class 2, and 99.5% for Class 3—illustrating that the model is highly effective at correctly identifying the majority of instances in each class.

There are nine instances where Class 1 was incorrectly predicted as Class 2, and very few (1.5%) were predicted as Class 3. This suggests a minor confusion between Classes 1 and 2. Class 2 saw a similarly low level of misclassification, with five instances mistaken for Class 1 and six for Class 3, translating to minimal confusion with the other classes. Notably, Class 3 shows the highest predictive accuracy and the lowest misclassification, with only three instances incorrectly classified as Class 1 and six as Class 2. The error rate here is also minimal, which suggests a clear distinction by the model between Class 3 and the others.

The bottom row of the figure presents the precision for each predicted class, showing high values across all three classes: 99.2% for Class 1, 98.3% for Class 2, and 99.0% for Class 3. This indicates that when a class label is predicted, it is very likely to be correct. The right column shows the recall for each class, which is similarly high, confirming that the model has a high rate of successfully predicting each class when it is the true class.

The confusion matrix indicates a well-performing model with high accuracy and precision across all classes, alongside excellent recall rates. The minor misclassifications between Classes 1 and 2 suggest some areas for potential improvement, possibly by further adjusting model parameters to better differentiate between these classes. Overall, the model demonstrates robust classification capabilities, effectively distinguishing between the classes with minimal errors, making it highly suitable for practical applications where accurate class predictions are essential.

5.3. Metrics of Each Class

The Figure 7 showcases the recall, precision, and F1 score for the classification model across the three distinct classes.

The first graph illustrates the recall for each class, showing extremely high values: 98.65% for Class 1, 98.28% for Class 2, and 99.5% for Class 3. Recall, or sensitivity, measures the model’s ability to correctly identify all relevant instances within each class. The near-perfect recall rates indicate that the model is exceptionally capable of capturing almost all true positives for each class without missing many actual instances.

The second graph presents the precision scores, which are similarly impressive: 98.98% for Classes 1 and 99% for Class 3, and 98.42% for Class 2. Precision reflects the accuracy of the positive predictions made by the model, signifying that nearly all instances predicted by the model as belonging to a given class were correctly classified. High precision is essential in our study, where false positives are costly or undesirable; this ensures that the model’s predictions are reliable and trustworthy.

The final graph shows the F1 scores, which are the harmonic mean of precision and recall, capturing the balance between the two. The scores are exceedingly high across all classes: 98.82% for Classes 1 and 99.25% for Class 3, and 98.35% for Class 2. The F1 score is particularly useful as a single metric to evaluate the model’s overall accuracy when dealing with datasets where class imbalances might affect the mean of precision and recall. The consistency in high F1 scores across all classes suggests that the model performs well even if the class distribution is uneven, as in our case.

The uniformity in performance across the metrics points to a well-tuned model, where the risk of overfitting or underfitting is minimized, ensuring that the model is robust and can generalize well on unseen data.

Our ensemble model demonstrates exceptional performance, with accuracy, recall, and F1 scores all exceeding 98%. This performance surpasses that of similar studies referenced in [22,23], which report a maximum accuracy of 91%. These results underscore the efficacy of our model in predicting the crowdfunding readiness of academic research projects, establishing a new benchmark for this field.

5.4. Testing on Unseen Data

Testing the developed model for crowdfunding readiness on unseen data yielded promising results, underscoring the model’s robust predictive capabilities. The accuracy achieved was remarkably high at 91.667%, also surpassing the accuracy using unseen data reported in [22,23], indicating that the model is effective at classifying projects into their appropriate readiness levels across the diverse set of scenarios represented in the dataset. This high level of accuracy demonstrates the model’s general utility in real-world applications, where it can reliably assess the likelihood of crowdfunding success.

The model also achieved a macro-averaged precision of 87.778%. This metric reflects the model’s ability to deliver relevant results, reducing the likelihood of false positives—important for stakeholders who rely on the model’s predictions to make informed decisions about which projects to pursue or invest in. The macro recall, at 90.952%, suggests that the model successfully identifies a high proportion of relevant instances across all classes, which is crucial for not overlooking potentially successful projects. Together, these figures indicate a balanced model that neither sacrifices recall to improve precision nor vice versa.

Moreover, the macro F1 score, combining precision and recall, stood at 89.337%, providing a harmonized view of the model’s performance. This score is particularly valuable in contexts like crowdfunding, where both the precision of predictions and the ability to capture all relevant cases are crucial for maximizing funding opportunities and strategic planning. This comprehensive performance suggests that the model is well calibrated and can be trusted to provide a dependable foundation for decision-making in crowdfunding campaigns. The results from this testing phase on unseen data confirm the model’s readiness for deployment, potentially offering significant benefits to platforms and creators by enhancing the predictability of campaign success.

5.4.1. Confusion Matrix

The confusion matrix from Figure 8 offers insightful details into the performance of our machine learning model, which was developed to predict the crowdfunding readiness of various projects using unseen data. This matrix categorizes project predictions into the three distinct readiness levels described above.

For Class 1, which comprises 14 instances in the dataset, the model displayed a high degree of accuracy, correctly predicting 13 of these as “Not Ready”. This suggests that the model is adept at identifying the signs of projects that are not yet ready for crowdfunding, an essential capability for preventing premature campaign launches. However, there was one instance where the model misclassified a Class 1 project as Class 2, indicating a minor confusion between projects at the cusp of “Not Ready” and “Needs Improvement”.

Class 2, represented by five instances in the dataset, posed a greater challenge for the model, which correctly identified only four of these. The misclassification involved one project incorrectly predicted as Class 3. This indicates some difficulty in distinguishing between projects that are at “Needs Improvement” readiness level and those that are “Fully Prepared” for crowdfunding, suggesting that attributes defining the medium readiness might overlap significantly with those of higher-readiness projects.

Class 3 also consisted of five instances, all of which were accurately classified by the model. This flawless prediction of “Fully Prepared” projects is particularly commendable, as it directly supports the goal of recognizing highly viable projects for successful crowdfunding.

Overall, the model performs robustly in recognizing both the lower and higher extremes of project readiness. The slight challenge in accurately classifying medium readiness projects could be attributed to the inherent subtlety in feature differences between the readiness levels, particularly between medium and high. To enhance the model’s performance, future efforts will focus on adjusting the model parameters to better delineate these subtleties. Enhancing the model’s discrimination between Class 2 and Class 3 could lead to more precise support for decision-making in crowdfunding campaign management, ensuring resources are allocated to projects with the appropriate level of readiness.

5.4.2. Metrics of Each Class

The performance metrics displayed in Figure 9 capture the effectiveness of the prediction model when applied to an unseen dataset, evaluating its precision, recall, and F1 score across the three distinct classes.

The precision metric, which reflects the accuracy of positive predictions, shows that the model performs exceptionally well for Class 1 with a perfect score of 1.00. This indicates that every project the model identified as Class 1 was indeed correctly classified, suggesting a strong ability to recognize characteristics typical of this class. Class 3 also demonstrates good precision at 0.83; although slightly lower, it still represents a high degree of accuracy in predictions. Class 2, however, shows a precision of 0.80, which is the lowest among the three classes, indicating some challenges the model faces in accurately identifying projects that truly belong to this medium readiness category.

In terms of recall, which measures the model’s ability to identify all relevant instances within a class, Class 3 achieves a perfect score of 1.00. This suggests that the model successfully captured all instances that should be classified under Class 3, underscoring its effectiveness at recognizing the readiest projects for crowdfunding. Class 1 has a recall of 0.93, which is commendably high, demonstrating that the model can effectively detect most projects that belong to this lower readiness level. Class 2’s recall is at 0.80, echoing the precision score and highlighting a relative difficulty in consistently identifying all projects that fit into this category.

The F1 score, a balanced measure that combines precision and recall, is presented with high values across the board, reinforcing the model’s robustness. Class 1 has an F1 score of 0.96, nearly perfect, reflecting both high precision and recall, making it a reliable class for prediction. Class 3 follows closely with an F1 score of 0.91, emphasizing strong overall performance in recognizing fully ready crowdfunding projects. The F1 score for Class 2 stands at 0.80, which, while lower than the other classes, still indicates a reasonable level of balance between precision and recall.

Overall, these metrics elucidate a model that is highly capable of identifying projects ready for crowdfunding, particularly for the extremes of readiness (Classes 1 and 3). The slight dip in metrics for Class 2 points towards potential areas for improvement, possibly requiring adjustments in model training to enhance the model’s sensitivity to characteristics of medium-readiness projects.

For the two projects that were misclassified by the model, a detailed analysis of the data, collected from the student group evaluations and forming the resulting dataset, offers some insightful revelations.

The first project, initially labeled as Class 2 but classified by the model into Class 3, is currently at a technological readiness level (TRL) of 6. This indicates a relatively advanced stage of development, where technologies have already been validated in a relevant environment. Interestingly, the average score derived from the evaluations of 15 student reviewers for this project was 2.46. This score suggests a borderline classification where the project teeters between needing improvements and being nearly ready for crowdfunding. The misclassification might stem from the model’s interpretation of the project’s readiness as more developed due to its higher TRL, overlooking nuanced evaluative feedback that indicates lingering needs for improvement.

The second project, which was categorized by the model as Class 2, while it was labeled as Class 1 in the dataset, resides at a lower TRL of 3. This stage is typical for projects where the proof of concept has been established but not yet validated in a laboratory setting. Among projects at this TRL, there is a mix of both Class 1 and Class 2, indicating varying degrees of development maturity and market readiness. The average evaluation score for this project from the 15 evaluators was 1.46, reflecting a consensus leaning towards it being “Not Ready” for crowdfunding. A notable feature of this project is the presence of a potential client within its portfolio. This aspect introduces a strategic crossroads: whether to further develop the project tailored specifically to the needs of this client or to pursue a broader commercial path suitable for crowdfunding.

5.5. Simulation of Improvement Scenarios

In order to demonstrate the additional utility of the proposed model, an illustrative example was selected from among the 24 projects that comprise the new unseen dataset. The chosen project focuses on developing a navigation solution for the visually impaired, which utilizes area-mapping sensors and vibration alerts integrated into a glove to warn of nearby obstacles. Initially evaluated at a technological readiness level (TRL) of 3, this solution indicates proof of concept in a laboratory setting that is still in its early stages.

The project was initially classified in Class 1, indicating it is “Not Ready” for crowdfunding. The evaluation kept the last four features (i₁, i₂, i₃, and i₄), which represent the project’s suitability within the most appropriate crowdfunding platform, constant. The simulation involved altering the other four features from their current starting values—TRC = 0.12; MRC = 0.03; CRC = 0.01; and MTRC = 0.79—up to their maximum potential values, as determined by predefined computational relations mentioned in Section 3, with an increment of 0.1.

Graphical representations and analysis of the predictions for two-feature combinations were conducted. The first simulation involving TRC and MTRC, which represent the technological level and the team’s capabilities, showed that increasing these features without altering the others does not elevate the project beyond the “Not Ready” class. This is depicted in Figure 10, emphasizing the multidimensional requirements for readiness improvement.

In a subsequent analysis combining TRC and MRC (technological level and market connection), illustrated in Figure 11, it was observed that unless the MRC exceeds 1.5, the project remains, at best, classified as “Needs Improvement”, regardless of technological advancements. This supports the notion that a successful project requires not just advanced technology but also a viable market.

Exploring the combination of TRC and CRC (technological and commercial levels), depicted in Figure 12, the analysis revealed that transitioning from “Not Ready” to “Needs Improvement’ necessitates industry contact and access to venture capital. However, these factors alone are insufficient to elevate the project to “Fully Prepared” status without improvements in market connectivity.

The final simulation, shown in Figure 13, used MRC and CRC as features, focusing on market and commercial levels while excluding technological and management changes. This scenario displayed a sharp transition from “Not Ready” to “Fully Prepared”, underscoring the significant impact of commercial readiness on project classification.

These simulations collectively illustrate that while individual improvements can shift project readiness to higher categories, a holistic enhancement across all dimensions is essential for a project to be considered ‘Fully Prepared’. The findings underscore the complexity of preparing for crowdfunding, highlighting that success is contingent upon a balanced advancement across technological, market, and commercial aspects.

Given the insights derived from the simulation of improvement scenarios on the four combinations of features (TRC, MRC, CRC, and MTRC), it is clear that targeted strategies are necessary for each feature combination to effectively transition a project from one readiness class to another, elevating the readiness level of a project, particularly focusing on moving from “Not Ready” to “Needs Improvement” and ultimately to “Fully Prepared”:

• Market engagement: Engage potential customers and industry stakeholders early in the development process to gather feedback and foster market interest. Conduct extensive market research to understand customer needs and validate the market potential of the technology being developed.
• Commercial strategy development: Develop a clear commercial strategy that includes pricing, distribution, and key value propositions tailored to target markets. Establish partnerships with industry players to gain commercial insights and access to market channels.
• Scalable business models: Design business models that are easily scalable and adaptable to changes in market conditions and consumer preferences. Invest in marketing and promotional activities to build brand awareness and create demand before the product hits the market.

While each combination has its own strategy, projects often require simultaneous enhancements across multiple areas. For instance, improving TRC often goes hand in hand with enhancements in MTRC to handle the increased complexity of advanced technologies, but it is not enough to take the great leap towards a crowdfunding campaign.

Thus, the prediction model can be used to regularly assess the impact of implemented strategies on project readiness. Moreover, it can be used to continuously assess the readiness level based on changes in any of the raw inputs, including the characteristics evaluated through the two questionnaires, as shown in Figure 14.

Figure 14 illustrates the flow of inputs and interactions within our proposed ensemble prediction model, which assesses crowdfunding project readiness.

The ensemble model operates by synthesizing the quantifiable metrics derived from these factors into a cohesive readiness assessment. Inputs from the questionnaires provide a detailed view of each project’s strengths and weaknesses across the evaluated characteristics, allowing for the model to deliver a nuanced prediction of the crowdfunding readiness level. This systematic approach not only enhances the accuracy of readiness assessments but also enables continuous monitoring and updating of the readiness status as project dynamics evolve. This adaptability makes it a powerful tool for project managers and stakeholders to pinpoint areas needing improvement and effectively direct resources to optimize their crowdfunding strategies.

Several practical applications of the predictive model highlight its utility in managing and optimizing crowdfunding campaigns:

• Scenario simulation: The model can simulate the impact of one or more specific actions—reflected through changes in model inputs—on the project’s readiness level. By adjusting various input values that represent different aspects of the project, such as marketing efforts or technological advancements, users can foresee how these changes could potentially elevate the project’s preparedness for a crowdfunding campaign. This capability allows for strategic planning and foresight, enabling project managers to make informed decisions about where to allocate resources for maximum effect.
• Impact analysis: The model is instrumental in identifying which actions have the most significant influence on enhancing the project’s readiness while consuming the least resources. This feature is crucial for efficiently managing limited resources while striving to improve critical aspects of the project that directly influence its likelihood of crowdfunding success. By pinpointing the most impactful actions, the model helps in focusing efforts and investments on areas that yield the highest returns in readiness improvement.
• Action prioritization: Beyond identifying impactful actions, the model aids in prioritizing these actions within the project’s timeline. It helps project teams to structure their tasks and milestones based on the urgency and impact of each action on the project’s overall readiness. This structured approach ensures that critical readiness factors are addressed at the right time, enhancing the project’s overall appeal and readiness by the time it is launched on a crowdfunding platform.

Through these scenarios, the predictive model serves not just as a diagnostic tool but also as a strategic asset in guiding the developmental trajectory of crowdfunding projects, ensuring they are well prepared and positioned for success.

6. Conclusions

This research introduces an innovative framework, comprising a comprehensive set of metrics to evaluate the development stage of projects alongside a predictive model that synergistically combines random forest, gradient boosting, and logistic regression. The fact that it outperforms the separate models on all fronts makes this ensemble model a major milestone in research on predicting success of crowdfunding campaigns, which is confirmed by its relative performance against similar implementations of random forest and gradient boosting on various data.

The ensemble model performs significantly better than each individual model independently. A comparison between different datasets shows that while random forest model [23] applied to the Kaggle Kickstarter dataset has strong performance with 91% accuracy and F1 score, the gradient boosting model [22] in a different set of Kickstarter data has relatively low metrics across the board with 86% accuracy and 67% F1 score. In sharp contrast, our custom dataset witnessed an impressive increase for the ensemble model accuracy to 98.94% and for its F1 score to 98.81%. This shows a dramatic precision–recall improvement where both hit rates go up to 98.80%.

This paper is useful for optimizing resource allocation when passing from concept state to market-ready products within scientific research projects. By correctly predicting if a project is good enough for crowdfunding, stakeholders will make well-informed decisions on resource investment, aligning it with the development needs and market potentialities of a particular project. Thus, this approach increases the chances of successful crowdfunding while speeding up the product creation process through efficient nurturing of innovative ideas into viable products.

Overall, the integration of a multifaceted predictive model within the crowdfunding domain offers a strategic advantage, enabling a more nuanced and data-driven approach to project management and funding. The enhanced accuracy and detailed project readiness assessment provided by the ensemble model pave the way for more strategic project planning and funding acquisition, ultimately contributing to the greater success rate of crowdfunding campaigns.

Our approach is structured to support project creators in refining their initiatives to align more closely with crowdfunding criteria and public expectations. While acknowledging that risk-adjusted returns are vital for investors, our model is principally designed to furnish project developers with critical insights and metrics that boost their projects’ appeal and crowdfunding viability. Recognizing the importance of broader financial outcomes for investors, we propose that future research could extend into analyzing post-campaign performance, offering a deeper exploration of financial returns and investor impact, thereby enhancing the comprehensiveness of crowdfunding evaluations.