Source-linked AI summary
Early Prediction of Movie Box Office Success based on Wikipedia Activity Big Data
Márton Mestyán, Taha Yasseri, János Kertész
TL;DR
The paper asks whether collective online activity can support early prediction of movie popularity despite the difficulty of forecasting box-office success before release. It measures Wikipedia editor and reader activity for movies and applies linear regression, finding reasonable predictions about one month before release. The approach demonstrates that Wikipedia activity can provide an alternative signal for collective reactions to cultural products, while accuracy is weaker for less successful movies.
Problem
Predicting a movie’s financial success before release remains challenging, despite prior approaches using market, quality, popularity, and other data.
Method
The study combines Wikipedia reader page views and editor activity measures for 312 U.S. movies released in 2010 in linear regression models predicting first-weekend revenue.
Results
R2 > 0.925 is reached as early as one month before release, while the full activity-based predictor set reaches R2 = 0.77 around that time.
Takeaways & Limitations
Wikipedia and similar mass-collaboration platforms can provide practical signals for collective reactions and may generalize to other products or fields.
Takeaways & Limitations
Prediction is more precise for successful movies and less accurate for less successful movies, where related data volume is smaller.
Abstract
from arXiv · showhide
Use of socially generated "big data" to access information about collective states of the minds in human societies has become a new paradigm in the emerging field of computational social science. A natural application of this would be the prediction of the society's reaction to a new product in the sense of popularity and adoption rate. However, bridging the gap between "real time monitoring" and "early predicting" remains a big challenge. Here we report on an endeavor to build a minimalistic predictive model for the financial success of movies based on collective activity data of online users. We show that the popularity of a movie can be predicted much before its release by measuring and analyzing the activity level of editors and viewers of the corresponding entry to the movie in Wikipedia, the well-known online encyclopedia.
Introduction
Digital activity traces and user-generated content offer data for monitoring and predicting collective reactions, but forecasting a movie’s financial success before release remains challenging. This work addresses that gap using Wikipedia editor and reader activity.
- Motivation: Digital environments record passive activity traces and actively shared opinions, moods, and views that can support analysis of individual and social phenomena.The paper frames large-scale digital data as useful for predicting, monitoring, and coping with events.
- Related work: Prior studies found correlations between online activity measures and offline outcomes, including citations, stock transactions, and economic indicators.Examples include Twitter mentions with arXiv downloads and Google searches with company transactions or GDP-related behavior.
- Open problem: Despite extensive statistical and machine-learning approaches, predicting a movie’s financial success remains a challenging open problem.Earlier work examined movie-income distributions, novelty, and pre-release quality or popularity variables.
- Related work: Wikipedia research has examined its growth, coverage, editorial conflict, patterns, linguistic features, and responses to external events.Using Wikipedia to detect or predict real-world events had previously received limited attention.
- Study contribution: The study uses editor activity and reader page views for 312 U.S. movies released in 2010, applying linear regression to forecast first-weekend revenue.The approach reports reasonable predictions as early as one month before release and avoids language-based analysis.
Results
The study measures Wikipedia activity before release and relates it to movie revenue through correlations and multivariate linear regression. Activity-based predictors become more informative near release, with the full predictor set reaching R2 = 0.77 about one month beforehand.
- Data and measures: The dataset contains 312 tracked Wikipedia articles corresponding to U.S. movies released in 2010, whose histories often begin before theatrical release.This earlier article creation permits monitoring movie popularity well in advance.
- Data and measures: Four accumulated activity measures are used: views V, human editors U, human edits E, and collaborative rigor R.Movie time sets release as t = 0, and measures accumulate from article inception to prediction time.
- Correlation analysis: For all activity-based predictors, Pearson correlation with revenue generally increases toward release and rises sharply around release.Among these measures, page views V have the highest pre-release correlation with revenue.
- Prediction model: The authors fit multivariate linear regressions with different predictor combinations and evaluate them using 10-fold cross-validated R2(t).The model includes time-varying coefficients, a constant, and a noise term.
- Prediction results: R2 = 0.77 is achieved by the {V, U, R, E, T} predictor set around one month before release.The {V} model performs roughly as well as the market-information benchmark, while {V, T} and {U, T} perform above it.
Discussion
The study argues that Wikipedia activity can support early movie box-office prediction, including reasonable accuracy one month before release. Its simple model performs best for popular movies and leaves room for more sophisticated methods and broader data.
- Wikipedia user-generated data can enhance prediction of society’s collective reaction to cultural products and may support marketing applications.The authors also suggest possible generalization to other fields, languages, and local markets.
- The model excludes article creation-to-release time and article length because these activity measures showed no significant correlation with box-office revenue.
- R² reaches 0.94 a few days before release for the Wikipedia model on the 24-movie comparison sample, versus 0.98 for Twitter at release night.The Wikipedia model nevertheless provides reasonable accuracy, R² > 0.925, as early as one month before release.
- At t = −30 days, predicted revenue is more precise for successful movies, while deviations from perfect prediction increase for less successful movies.The authors attribute this difference to the larger volume of related data for more popular movies.
- More sophisticated methods, including neural networks and detailed content-related parameters such as article controversy, could potentially improve prediction efficiency.
- Figures compare predictor correlations, regression R², Twitter results, and predicted versus actual revenue across the movie sample.Figure 1 summarizes variables for 312 movies; Figures 2–5 show temporal correlations, model fit, cross-method comparison, and predictions at t = −30 days.
Methods
The study combines financial records for 312 U.S. movies released in 2010 with Wikipedia activity measures aligned to a common movie-time coordinate. It uses human editing and viewing histories, accumulated through prediction time, and evaluates predictions with 10-fold cross-validation.
- Data collection: The dataset covers 312 of 535 movies screened in the United States during 2010, matched to corresponding Wikipedia pages.Movie pages were located using Wikipedia’s 2009 films and 2010 films categories.
- Temporal alignment: Movie time measures temporal variables relative to each movie’s U.S. release, while activity measures accumulate from article inception through prediction time.This coordinate allows activity histories for different movies to be compared at equivalent stages around release.
- Predictors: The activity measures include human users, edits, collaborative rigor, and page views.Collaborative rigor counts subsequent edits by the same user as one edit, reducing the effect of rapid repeated corrections.
- Predictors: Wikipedia revision histories are used to calculate editor measures, with automated bot contributions excluded from human-activity counts.The revision databases register information about page modifications.
- Evaluation: The coefficient of determination is estimated by randomly partitioning the 2010 movie sample into 10 subsets, training on nine and testing on the remaining subset across all permutations.The reported coefficient is the average over the 10 validation permutations.
- Predictors: Figure 6 represents article-edit activity as ticks, with each tick marking a page modification and tick styles identifying different users.The schematic depicts how editor activity variables are characterized over an article’s history.