1 of 19

NUS IS5126 AY 2025-2026 Sem 1, Group 11

Movie Box ROI Prediction

NUS IS5126, AY25–26, G11

Cham Jin Wei�A0318721H	He Zean�A0297000N	Lian Jie Nicholas�A0108553H	Liang Junyi�A0314698R	Yang Yue�A0194569J

2 of 19

Content

Problem & Research

Define goals & review literature

1

Clean Data

Handle missing values & outliers

3

Method Development

Build & compare ML models

5

Web Demo

Interactive profit prediction tool

7

Data Explore

Analyze distributions & relationships

2

Feature Engineering

Transform & create new variables

4

Result/Business Analysis

Interpret results & business impact

6

Our project’s workflow also follows this flow

3 of 19

Problem & Gap

Current Challenge

Pre-release uncertainty

Prior work: macro / binary

Limited actionable insights

Industry struggles with pre-release forecasting

The Gap

Need film-level ROI regression

Granular insights required

Time-aware analysis

Specific ROI targets needed

Bridge: Film-level time-aware forecasting

4 of 19

ROI Definition & Formula

Return on Investment

ROI = (Revenue - Budget) / Budget

Express as percentage�for easier interpretation

ROI Interpretation

Loss

ROI < 0: Investment lost

-20%

Breakeven

ROI = 0: No gain, no loss

0%

Profit

ROI > 0: Investment gained

+50%

ROI enables decision thresholds for investment strategies

5 of 19

Initial data cleaning

Initial Feature Exclusion

Budget, revenue, release dates

Cast, crew, genres

High-Missing Text Handled

Language imputation

Genre classification

Duplicates Removed

Exact duplicates

Data Leakage

Profit

revenue

roi

Post-release

popularity

vote_count

6 of 19

Distributions & Log Transformation

Right-skewed Distributions

Budgets & revenues show right skew

Heavy tail on right side

Apply Log

Multiplicative → Additive

Log-Transformed Distributions

More symmetric distribution

Stabilizes variance

Benefits of Log Transformation

Makes models more robust

Stabilizes variance

Simplifies interpretation

7 of 19

Temporal Trend & Leakage

Revenue Trends

Revenues trend upward

Temporal dependence

Data Leakage

Avoid serial correlation

Use time-based splits

Mean & Median Revenue Over Years

Mean

Median

Timeline Split Diagram

8 of 19

Feature Correlation

Correlation Analysis

Correlation Pruning

Remove features with |r| ≥ 0.8

Keep higher variance one

Keep more unique one

Information Score = (unique_count, variance)

9 of 19

Advanced Feature Engineering

📝 Text Processing - overview

TF-IDF vectorisation
GPT summarisation

🏢 Production Company

⭐Cast

Lead Actor and Actress

Star power? Based on number of past movies

Top 4 billed Cast Members

Star power? Based on average number of past movies

📅Number of movie releases in the same month

10 of 19

Advanced Feature Engineering

Positive Effect	Some or Negative Effect
Romance x Comedy	Horror x Comedy
Fantasy x Adventure	Horror x Sci-Fi
Sci-Fi x Action	Action x Thriller
Drama x Comedy	Action x Comedy
	Action x Adventure

🔀Feature Interaction

Genre x Release Period
Release Period x Star Power
Genre with Lead Star vs Genre with Star Ensemble
Genre x Genre

11 of 19

Basic Machine Learning Models

Ridge Regression
Random Forest
XGBoost

I'll walk through our basic machine learning models and the transformer pipeline we built on top of the same time-aware splits from The Movies Dataset. We started with 38 cleaned numeric features, all standardized, and trained on Ridge Regression, Random Forest, and XGBoost. Every model is evaluated with RMSE, MAE, and R² on validation and test sets.

Among the baseline models, Ridge Regression achieved the best performance with a test RMSE of 1.378, and R² of 0.676. Random Forest achieved a test RMSE of 1.407 with R² of 0.662, approximately 2.1% worse than Ridge Regression. XGBoost performed the worst among the three models. The superior performance of Ridge suggests that the feature space may be relatively linear, or that Ridge's regularization effectively prevents overfitting that affects the tree-based models.

12 of 19

Basic Machine Learning Models

Random Forest (Tuned)

XGBoost(Tuned)

13 of 19

Transformer

Enriched dataset: LLM-based overview summaries, text length indicators, multilingual signals, and temporal roll-ups.

14 of 19

ROI Decision Making

15 of 19

Quick DEMO

16 of 19

English fuzzy match: confidence 75%

English AI fix + fuzzy match: confidence 100%

17 of 19

English fuzzy match: failed with confidence 63%

English AI fix + fuzzy match: confidence 100%

18 of 19

Chinese fuzzy match: failed

Chinese (even with wrong name) AI fix + fuzzy match: confidence 100%

19 of 19

Thank you for your attention!