Cultural economics frames films as unique cultural goods shaped by consumer preferences, distribution strategies, and cultural value. Star power theory, emphasized by Lash and Zhao (2016) , highlights the role of well-known actors in attracting audiences and boosting box office performance. Chisholm et al. (2014) also underscore casting’s economic significance.

Traditional marketing channels like TV and print create broad exposure ( Scott, 2019 ). Digital marketing has emerged as a key strategy, leveraging social media platforms to target audiences effectively ( He & Hu, 2021 ). While large budgets typically correlate with higher revenues, diminishing returns occur when market saturation is reached ( Wisnefsky, 2023 ; Lash & Zhao, 2016 ).

Star power, measured by actors’ previous earnings and media presence, drives strong opening weekends ( Lash & Zhao, 2016 ). However, casting decisions must balance popularity with narrative quality to sustain success ( He & Hu, 2021 ). Excessive reliance on star power can also result in diminishing returns ( Wisnefsky, 2023 ).

Research on film success prediction frequently employs statistical and machine learning methods. Scott (2019) used regression analysis to examine factors like marketing spend, star power, and release timing on box office revenue. He and Hu (2021) utilized machine learning techniques to analyze large datasets, offering improved predictive accuracy. Social network analysis by Lash and Zhao (2016) highlighted relationships between actors, directors, and audiences as key profitability indicators.

Data inconsistencies for international films and independent productions present challenges ( Chisholm et al., 2014 ). Additionally, subjective metrics such as reviews and audience sentiment are difficult to quantify, complicating their integration into models ( Lash & Zhao, 2016 ). These issues can limit the generalizability of findings.

Films must adapt marketing strategies to cultural contexts. For instance, Chinese audiences prefer action-packed spectacles, while European audiences favor narrative-driven films ( Lash & Zhao, 2016 ). Additionally, streaming platforms like Netflix challenge traditional box office metrics while introducing new revenue opportunities ( He & Hu, 2021 ).

While existing research examines marketing spend, casting, and timing, gaps remain regarding the impact of streaming platforms and digital metrics on film success. Further studies should explore diminishing returns on marketing in the streaming era and the interplay between cultural factors, traditional, and digital marketing strategies.

The study addresses these gaps by analyzing traditional and digital marketing strategies, casting decisions, and other film attributes. By integrating prior findings, it provides actionable insights and a comprehensive perspective on film success, contributing to the evolving understanding of the industry’s economic dynamics.

This study examines factors influencing U.S. movie success, focusing on release timing (e.g., holiday seasons) and audience engagement, measured by the vote-to-popularity ratio, to analyze their correlation with return on investment (ROI).

Data was collected from two main sources: The Movie Database (TMDb), offering details like runtime, budget, revenue, and popularity, and the Open Movie Database (OMDb), providing additional metrics, including Rotten Tomatoes and IMDb ratings. Together, these sources created a comprehensive dataset integrating qualitative and quantitative film performance indicators.

Prior to analysis, as shown in Figure 1 , the dataset underwent thorough cleansing and feature engineering to ensure data integrity and enhance model performance:

1) Missing Data: Key fields like budgets and revenues were imputed or removed.

2) Feature Selection: Relevant features such as IMDb ratings and holiday releases were retained to reduce multicollinearity.

3) Balancing: SMOTE addressed imbalances between successful and non-successful films.

4) Splitting: Data was split 70:30 into training and testing sets ( Figure 2 ).

Initially, a movie was classified as profitable if its revenue exceeded its production budget. However, considering that additional costs such as marketing and distribution typically add 25% - 50% to the production budget, the ROI threshold was raised to 1.5. This adjustment ensures that an ROI around 1 likely only covers costs, allowing for a more realistic assessment of a film’s profitability (see Figure 3 ). Star power was measured by cumulative box office earnings, with actors surpassing $900 million classified as superstars.

Exploratory Data Analysis (EDA) was performed to analyze the dataset’s distribution, patterns, and outliers. Python and R were used for data manipulation and visualization, which informed handling missing data and refining the dataset.

Log-transformed revenues ( Figure 4 ) and budgets ( Figure 5 ) revealed clearer patterns, with most revenues between log values of 15 - 21 and budgets between 17.5 - 20. Figure 6 compares budgets of successful and non-successful films, showing overlap that suggests budget alone does not predict success. Figure 7 highlights the correlation between budgets and revenue, indicating higher budgets generally yield more revenue but are not guaranteed. The correlation matrix ( Figure 8 ) shows moderate positive correlations between IMDb/Rotten Tomatoes Ratings and success, while “Is a Holiday” and “Has Won Award” have weaker correlations.

To enhance model reliability, multicollinearity was addressed using:

The VIF values for the final set of predictor variables are presented in Table 1 . All variables have VIF values below the threshold, indicating acceptable levels of multicollinearity.

The correlation matrix ( Figure 9 ) indicates that the highest correlation between any two predictors is 0.751 (between IMDb.Rating and Rotten.Tomatoes.Rating). While this exceeds the common threshold of 0.7, it is essential to consider the context and the specific roles these variables play in the model. Both ratings provide distinct insights into film success—IMDb.Rating reflects audience ratings, whereas Rotten.Tomatoes.Rating aggregates critical reviews. Given their unique contributions and the overall low multicollinearity indicated by VIF values, both variables were retained in the final feature set to capture diverse aspects of film success.

As a result, the final feature set included: Has.Won.Award, Is.a.Superstar, IMDb Rating, Rotten Tomatoes Rating, and Runtime.

Three machine learning models were employed to predict film success as a binary outcome (0 = No, 1 = Yes):

1) Logistic Regression (LR): A fundamental statistical method valued for its simplicity and interpretability, LR is ideal for understanding relationships between predictor variables and binary outcomes. It serves as a robust baseline for comparing the performance of more complex models.

2) Random Forest (RF): An ensemble method combining multiple decision trees to handle non-linearity, assess feature importance, and reduce overfitting, making it well-suited for datasets with diverse features.

3) Support Vector Machine (SVM): Effective in high-dimensional spaces, SVM is robust against outliers and excels in classification tasks, making it beneficial for distinguishing between successful and non-successful films.

Cross-Validation ensured robust performance by dividing the dataset into training and testing subsets, enabling reliable model comparisons and minimizing overfitting risks.

Logistic Regression: Provides clear interpretability and serves as a baseline model to benchmark against more complex approaches.

Random Forest: Handles non-linear relationships effectively, includes mechanisms for evaluating feature importance, and is robust against overfitting due to ensemble learning.

Support Vector Machine: Excels in handling high-dimensional datasets, demonstrates robustness to outliers, and performs strongly in binary classification tasks.

Decision Trees: While intuitive and easy to implement, single decision trees are prone to overfitting and lack the predictive power of ensemble approaches like Random Forest. Their performance limitations make them less suitable for this study’s objective of robustly predicting film success.

The Random Forest Model outperformed other models with an accuracy of 81.85% and a sensitivity of 79.41% for class 0. Its strength lies in its use of multiple decision trees, t bootstrapping, and random feature selection, which enhance predictive power and reduce feature correlation. This ensemble approach effectively handles complex relationships between variables, making it a valuable tool for stakeholders seeking data-driven decision-making.

The Support Vector Machine (SVM) Model achieved an accuracy of 74.52% and a sensitivity of 74.26%. While slightly less accurate than the Random Forest, SVMs excel at defining clear decision boundaries, providing valuable classification insights. However, ensemble methods like Random Forest demonstrate superior predictive capacity when managing a broader set of features.

Regularization techniques (Ridge L2 and Lasso L1) achieved similar accuracy scores of 71.81% and sensitivities between 70% - 72%. The minimal improvement suggests that the dataset’s pre-selected features were already optimized, rendering additional regularization less impactful.

These findings have significant implications for stakeholders in the film industry, particularly when planning marketing strategies and making casting decisions. The analysis revealed that key features like “Runtime”, “IMDb Rating”, and “Rotten Tomatoes Rating” were highly influential, as reflected in their “MeanDecreaseGini” scores. High values for these features suggest that focusing on these elements can enhance the predictive modeling of a movie’s potential success. Additionally, “Has Won Award” underscores the role of award recognition in boosting a film’s credibility and audience interest, suggesting that strategic targeting of awards can enhance promotional campaigns.

Combining star power (“Is a Superstar”) with positive ratings and optimal runtime maximizes a film’s appeal. These findings emphasize the importance of leveraging established actors and directors to attract audiences and drive financial performance.

These results have several implications for film industry stakeholders:

1) Marketing Strategies: Focus on elements that significantly influence success, such as securing high ratings and targeting award nominations. Tailored, cost-effective marketing strategies are preferred over solely increasing budgets, aligning with the concept of diminishing returns on excessive marketing spend.

2) Casting Decisions: Prioritize hiring superstars and renowned directors, as their presence enhances marketability and success, especially during opening weekends.

3) Production Planning: Optimize film runtime and leverage high ratings to improve box office performance and long-term profitability.

The findings align with Lash and Zhao’s (2016) Star Power Theory, which emphasizes the significant impact of well-known actors on a film’s commercial success. The study confirms that star power attracts larger audiences and enhances financial performance, supporting previous research that established actors and directors increase a film’s marketability and success.

The results also echo Wisnefsky’s research, which demonstrated that films with larger marketing budgets typically perform better ( Wisnefsky, 2023 ). However, the study highlights potential diminishing returns on excessive marketing spend, reinforcing the need for strategic and targeted marketing efforts rather than focusing solely on budget size.

Furthermore, He and Hu’s findings on the importance of social media engagement and digital marketing as modern predictors of film performance are indirectly supported. The emphasis on ratings and award recognitions, which often influence online discussions and viewer interest ( He & Hu, 2021 ), suggesting that targeted digital campaigns and leveraging award credibility are essential components of modern film marketing.

These findings contribute to the understanding of how machine learning can predict box office outcomes and inform strategic decisions. The strong performance of the Random Forest Model suggests that ensemble learning methods are particularly well-suited for handling the multifaceted nature of film success prediction, where multiple factors interact in complex ways.

However, the study’s applicability may vary across different cultural contexts. For instance, in markets like China, spectacle-driven films and holiday releases may hold greater appeal, whereas in regions like France, local storytelling and regional star power might be more influential. Future research should explore these predictors across diverse cultural landscapes to provide broader insights for optimizing international film success.

The study faced data extraction issues, resulting in missing information for several films, such as Avatar: The Way of Water, Spider-Man: Far from Home, and The Conjuring: Last Rites. These gaps could affect the dataset’s completeness and the predictive models’ accuracy, potentially limiting the generalizability of the findings. Future studies should aim for more comprehensive data collection to mitigate these limitations.

Future studies could expand on this research by incorporating data from streaming platforms and analyzing how their metrics influence long-term success. The study’s findings provide a foundational understanding of the relationship between movie features and financial success, serving as a guide for industry professionals seeking to optimize their strategies in an ever-changing entertainment landscape.