Project: Critical Factors Determining Movie Success

An OLS Regression analysis to identify and quantify the key factors influencing movie box office performance.

Overview

This project applies Ordinary Least Squares (OLS) Regression to predict movie view counts using a dataset of 232 films. By combining data preprocessing, statistical transformations, and diagnostic testing, the study identifies significant predictors of movie success and develops a parsimonious regression model.

Key Features

  • Applied OLS Regression on a curated dataset from UCI Machine Learning Repository.
  • Incorporated Box-Cox transformation and Stepwise Regression for model refinement.
  • Conducted comprehensive diagnostic tests for normality, constant variance, multicollinearity, and outlier detection.
  • Compared multiple models using ANOVA and Adjusted R-squared to select the most efficient predictor set.

Approach

  1. Data Preparation
    • Selected relevant variables: Views, Likes, Dislikes, Comments, Budget, Screens, Sequels, and Aggregate Followers.
    • Removed incomplete rows (NA entries) to ensure model reliability.
    • Chose Views as the primary response variable.
  2. Model Building
    • Fit initial OLS regression models to capture relationships.
    • Applied Box-Cox transformation to address non-normality in residuals.
    • Used Stepwise Regression for feature selection and model simplification.
  3. Model Diagnostics
    • Shapiro-Wilk Test: Checked residual normality.
    • Brown-Forsythe Test: Verified homoscedasticity.
    • Variance Inflation Factor (VIF): Assessed multicollinearity.
    • Cook’s Distance: Identified influential outliers.
  4. Model Comparison
    • Performed ANOVA tests to evaluate model improvements.
    • Selected the final model based on Adjusted R-squared and predictor significance.

Results & Evaluation

  • Significant predictors of movie views included audience engagement metrics (likes, dislikes, comments) and distribution factors (screens, budget).
  • The Box-Cox transformation improved residual normality and reduced model bias.
  • Stepwise Regression removed redundant variables, enhancing model interpretability.
  • The final model balanced predictive accuracy with simplicity, achieving a higher adjusted R-squared than the baseline.

Interesting Discoveries

  • Movies with larger distribution (more screens) and strong social media presence (high aggregate followers) consistently achieved higher views.
  • Sequels tended to outperform standalone films when controlling for other variables.
  • Budget had a positive effect, but with diminishing returns beyond a certain threshold.

Report

Full Report (PDF)

Sample Commands / Snippets

```r

Box–Cox (lambda via MLE)

library(MASS) bc <- boxcox(lm(Views ~ Likes + Dislikes + Budget + Screens, data=movies), lambda=seq(-2,2,0.1))

Stepwise selection (both directions)

library(MASS) final <- stepAIC(lm(Views ~ Likes + Dislikes + Comments + Budget + Screens, data=movies), direction=”both”)

Diagnostics

shapiro.test(residuals(final)) library(car) vif(final)

Influence

plot(cooks.distance(final), type=”h”)