Source-linked AI summary
Early Predictions of Movie Success: the Who, What, and When of Profitability
Michael T. Lash, Kang Zhao
TL;DR
Movie investment decisions require profitability predictions at the pre-production stage, when available information is limited. The paper proposes MIAS, which combines heterogeneous data and analytical methods to model cast, content, release timing, and their interactions. Experiments report better profitability prediction than benchmark methods, while the analysis highlights directors and other factors relevant to movie profitability.
Problem
Movie success is uncertain, and investors need profitability predictions based on information available during pre-production.
Method
MIAS combines machine learning, text mining, social network analysis, and heterogeneous data to extract cast, content, release, and feature-matching characteristics.
Results
The system achieves better movie-profitability prediction, and the analysis highlights directors as an important factor.
Takeaways & Limitations
The system provides practical decision support for investors, while its profitability analysis may inform research on team performance and creative-work success.
Takeaways & Limitations
Profit is calculated from production budget, so the profitability measure is constrained by that cost basis.
Abstract
from arXiv · showhide
This paper proposes a decision support system to aid movie investment decisions at the early stage of movie productions. The system predicts the success of a movie based on its profitability by leveraging historical data from various sources. Using social network analysis and text mining techniques, the system automatically extracts several groups of features, including "who" are on the cast, "what" a movie is about, "when" a movie will be released, as well as "hybrid" features that match "who" with "what", and "when" with "what". Experiment results with movies during an 11-year period showed that the system outperforms benchmark methods by a large margin in predicting movie profitability. Novel features we proposed also made great contributions to the prediction. In addition to designing a decision support system with practical utilities, our analysis of key factors for movie profitability may also have implications for theoretical research on team performance and the success of creative work.
1 Introduction
Movie profitability is difficult to predict early enough to guide investment, so the paper develops MIAS to automate pre-production profitability predictions from heterogeneous historical data.
- Motivation: Only 36% of U.S. movies produced from 2000 to 2010 had box-office revenue exceeding their production budgets.This motivates profitability-focused investment support rather than revenue-only prediction.
- Research objective: The paper defines movie success as profitability and aims to predict it automatically to support investors’ decisions.
- Research objective: Predictions must use information available during pre-production because later predictions arrive too late for meaningful investment decisions.
- System design: MIAS extracts “who,” “what,” “when,” and matching features from historical data, then applies machine-learning methods under different profitability criteria.
- Contributions: The system combines machine learning, text mining, and social network analytics to predict profitability with minimal human intervention.
- Contributions: Novel features include dynamic networks, plot topics, cross-group matches, and profit-based star-power measures, all of which contribute to prediction performance.
2 Related work
Prior movie-success research commonly predicts revenue, admissions, or audience reception, while this paper emphasizes profitability and richer representations of cast, content, release timing, and feature interactions.
- Success measures: Earlier studies commonly treated success as revenue, admissions, discrete categories, or continuous success values rather than profitability.
- Success measures: Revenues are not directly related to profits because revenue-based metrics ignore production costs.
- Success measures: Profitability, measured through profits or ROI, is presented as a more meaningful success measure for investment decisions.
- Feature categories: Prior work uses release timing, theater availability, seasons, holidays, and release competition as predictors of movie success.
- Cast features: Cast research emphasizes star power, but prior studies had not explored actors’ profitability records as indicators of movie profitability.
- Cast features: The paper examines directors’ effects on movie profitability, addressing limited prior connections between director star power and financial outcomes.
- Team features: Earlier teamwork measures were limited by simplistic experience and diversity proxies, small samples, and selection bias toward successful movies.
- Research response: This study uses larger datasets and develops more accurate measures of actor experience and team diversity related to profitability.
3 The system framework
MIAS acquires and standardizes movie data from complementary online archives, engineers four feature groups, and trains predictive models that investors can use to assess profitability from historical data.
- Data preparation: The system uses APIs and web scraping to retrieve data, then cleans, transforms, consolidates, and stores it in a database.Standardization removes extraneous title characters, applies stemming and stop-word removal to plot synopses, and prevents duplication across sources.
- Data acquisition: MIAS begins by acquiring movie information from IMDb and BoxOfficeMojo, which complement plot coverage with revenue and budget data.IMDb provides plot synopses, while BoxOfficeMojo provides more comprehensive revenue and budget data; their information can be used jointly.
- Feature engineering: Feature engineering classifies predictors into four groups: “what,” “who,” “when,” and “hybrid.”The hybrid group matches “who” with “what” and “what” with “when.”
- Prediction and use: MIAS trains predictive models and uses cross-validation to select parameters, compare prediction methods, and evaluate overall system performance.The resulting model can receive information about a potential investment and predict whether it would be profitable from historical data.
4 Feature engineering
Using historical online movie data, the study derives four feature groups that represent the movie’s cast, content, release timing, and cross-domain relationships.
- Feature groups: The study derives four feature groups from historical movie data: “who,” “what,” “when,” and “hybrid.”Hybrid features match “who” with “what” and “what” with “when.”
4.1 “Who” Features
The “who” features characterize cast and director experience, prior financial performance, collaboration structure, and network changes associated with a movie’s team.
- Star power and experience: Star-power features summarize cast members’ prior box-office revenue and profit records, alongside industry experience measured through tenure.The study calculates actor and director gross and profit measures, including total, average, and top values, before the predicted movie.
- Collaboration networks: The network representation treats actors as nodes and prior co-starring relationships as edges, with yearly networks aggregating earlier collaborations.Static features for a movie in year y use the cast network through year y−1.
- Static network features: Network heterogeneity measures average cosine similarity among cast members’ collaboration neighborhoods, where lower similarity indicates greater team diversity.The measure uses actor neighborhood vectors from the adjacency matrix and is intended to capture prior collaboration overlap.
- Static network features: Average degree and betweenness centrality capture collaborative expertise and the presence of brokers connecting otherwise less-interconnected groups.The study also considers actor-director collaboration frequency and the average profitability of their previous collaborations.
- Dynamic network features: Dynamic network features measure how a new movie changes collaboration-network structure to capture network-level spanning of structural holes.The proposed measures include decreases in clustering coefficient and average shortest path after incorporating the new movie’s collaborations.
- Dynamic network features: A static average clustering coefficient captures structural holes only at the ego level, motivating network-level dynamic measures.The paper distinguishes immediate-neighbor structure from changes across the broader collaboration network.
4.2 “What” Features
The “what” features represent a movie’s content through genre, rating, plot synopsis, topic distributions, and adaptation status, using information available before funding decisions.
- Content representation: The “what” features include genre, rating, and fine-grained plot-synopsis descriptions of movie content.Examples of genre include action, science fiction, and family; ratings include PG13 and R.
- Content representation: LDA converts plot-synopsis text into topic distributions that provide aggregated content features for predictive modeling.The model produces topics represented by probabilistic word distributions and assigns each synopsis a probabilistic distribution over topics.
- Content representation: Topic-distribution vectors summarize a movie’s plot at an aggregated level for use as predictive features.Words with high probability within a topic serve as representative keywords.
- Adaptations: Adaptation features indicate whether a movie’s plot comes from a comic, true story, or book/novel.The paper motivates these features by noting that adaptations may originate from previously successful sources.
4.3 “When” Features
The model captures release timing through industry profitability, release-date characteristics, and seasonal or holiday context. These features reflect changing market conditions, audience preferences, and competition around a movie’s planned release.
- Release timing matters because industry peaks can attract audiences but also increase marketing, distribution, and competitive pressure.A target release trajectory usually exists early in production even though the exact date may change.
- Average Annual Profit measures the average profitability of movies in the year before a planned release.It captures the overall profitability of the movie industry before release.
- Release-date features encode whether a movie is released during a holiday and which season—spring, summer, fall, or winter—it enters.Holiday and summer releases may attract more audience, while competitive conditions also vary by release period.
4.4 Hybrid Features
Hybrid features model how a movie’s genre interacts with its cast and release context. They capture cast–genre expertise, cast novelty, prior genre profitability, and nearby competition beyond standalone “who,” “what,” and “when” variables.
- Hybrid feature rationale: Hybrid features match movie content with cast characteristics and release conditions, including actor experience with the planned genre and genre popularity at release.The paper motivates these matches as potentially more informative than actor star power or genre information alone.
- “What” + “Who”: Average Genre Expertise measures the cast’s average experience with the movie’s genres using actor and movie genre vectors.Actor vectors record genre appearance proportions across 26 genres, while movie vectors indicate one or more genres.
- “What” + “Who”: Weighted Average Genre Expertise extends AGE by weighting genre familiarity with each cast member’s gross revenue, combining genre experience and star power.A high WAGE corresponds to a big star who is familiar with the movie’s genre.
- “What” + “Who”: Cast Novelty measures team diversity through star-power-weighted inverse genre experience, highlighting unexpected stars in a movie’s genre.Higher values indicate a more diverse cast in terms of unexpected genre participation.
- “What” + “When”: “What” + “When” features combine prior genre profitability, evolving genre preferences, and competition from other releases near the planned date.Competition is calculated from the average star power of movies released within one month, while AWPG weights prior-year profitability by genre similarity.
5 Experiments
The experiments evaluate profitability prediction on a filtered movie dataset using ROI-based success labels and multiple classifiers. Across decision boundaries, the proposed model outperforms benchmarks, while removing the novel features substantially reduces performance and increases mis-classification cost.
- 5.1 Dataset and Basic Statistics: The dataset is broadly representative by genre except for Foreign movies, for which budget and revenue data may be harder to obtain.This creates a documented sampling boundary for interpreting the experiments.
- 5.3 Prediction models: The system uses binary and multi-class classification with logistic regression, naive Bayes, SVM, MLP, decision trees, random forest, and LogitBoost, evaluated by 10-fold cross-validation.Performance is assessed with AUC, accuracy, precision, and recall, with higher values indicating better performance.
6 Discussions
The discussion uses regression alongside classification to identify factors associated with movie profitability and to predict continuous ROI. LASSO selected informative features, highlighting star power, genre expertise, collaboration history, diversity, release timing, and data limitations.
- Regression analysis: Regression complements classification by estimating feature influences and predicting continuous ROI values for investors.Because ROI is highly skewed, the analysis uses log(ROI + 1).
- Regression analysis: LASSO was the best algorithm for predicting numeric ROI among the six algorithms evaluated.The model’s coefficients were then used to examine factors behind movie success.
- Selected features: 48 of 120 LASSO features retained non-zero coefficients: 16 negative and 32 positive.The penalty parameter was increased until all attribute-wise variance inflation factors fell below 10, with λ = 0.0065.
- Selected features: The selected features covered every feature group, with more than half classified as novel and the What group contributing the most features.Twelve of the 30 LDA-derived topics were among the selected features.
- Factors associated with profits: Profitability was positively associated with genre expertise, prior collaboration success, cast diversity, and winter releases, while several What features had negative coefficients.Negative features included drama and foreign genres, an R rating, and plot topics related to wars and music.
7 Conclusions and Future Work
MIAS is an early-stage movie-profitability decision support system that combines automatically extracted “who,” “what,” “when,” and hybrid features. Experiments and regression analyses support its predictive utility, practical applications, and possible relevance to creative-work and team-performance research, while identifying several directions for future refinement.
- Conclusions: MIAS learns from early-available historical data and organizes classic and novel predictors into “who,” “what,” “when,” and hybrid feature groups.The groups cover cast and teamwork, movie content, release timing, and matches between content and cast or timing.
- Conclusions: Across 11 years of movies, MIAS outperformed two benchmark methods in predicting movie profitability.The system is intended to support early investment decisions about whether a proposed movie is worth funding.
- Conclusions: MIAS supports “what-if” analysis by allowing users to manipulate cast members, plot elements, and planned release time.The system is also presented as potentially useful to other movie-industry stakeholders, including cinemas deciding whether to air a movie.
- Conclusions: The proposed features made substantial contributions to profitability prediction, including improved measures of factors such as team expertise and diversity.Regression analyses also identified key profitability factors and highlighted the importance of directors.
- Conclusions: The framework and extracted features may be adapted to other creative works whose contributors, content, and timing can be represented.Examples named by the paper include research papers, grant proposals, and operas.
- Future Work: Future work includes matching “who” with “when,” modeling informal friendship networks, analyzing full scripts, and adding economic and seasonal-release features.These directions aim to capture actor or director career changes, cast relationships, finer-grained content, consumer spending power, and timing suitability.
8 Appendix
The appendix reports additional binary and multi-class classification results under alternative profitability decision boundaries.
- Appendix: Additional binary classification results use top 30% return on investment as the decision boundary.
- Appendix: Additional binary classification results use ROI ≥67% as the decision boundary.
- Appendix: Additional multi-class prediction results are reported for the first decision boundary.
- Appendix: Additional multi-class prediction results are reported for the second decision boundary.