Business Problem, Insights, and Methodology Report for Mobile Games Sales
Prediction
1. Business Problem Description
1.1 Context
The mobile gaming industry is a highly competitive and lucrative market, with
thousands of games launched annually across various platforms. Companies, including
game developers and publishers, face the challenge of accurately predicting a game’s
total sales (sales_total) to optimize resource allocation, marketing strategies, and
development efforts. The dataset, comprising 16,598 mobile game records, includes
features such as title_name, device_type, launch_year, game_genre, publisher_name,
and regional sales (sales_usa, sales_europe, sales_asia, sales_misc). The primary
business problem is to develop a predictive model that estimates total sales based on
these features, enabling stakeholders to make data-driven decisions.
1.2 Problem Statement
The objective is to build a robust machine learning model to predict sales_total for
mobile games with high accuracy, targeting an R² score of at least 0.80 (explaining 80%
of the variance in sales). The current RandomForestRegressor model achieves an R² of
0.70, indicating a gap in predictive power. Accurate predictions will help:
• Publishers: Prioritize marketing budgets for high-potential games.
• Developers: Identify successful game genres or platforms.
• Investors: Assess the viability of funding specific game projects. Key challenges
include handling skewed sales data, missing values (271 in launch_year, 58 in
publisher_name), high-cardinality categorical features (e.g., 11,493 unique
title_name values), and capturing non-linear relationships.
1.3 Business Objectives
• Achieve an R² score of 0.80 or higher on the test set to ensure reliable
predictions.
• Identify key features driving sales to provide actionable insights for stakeholders.
• Develop a scalable model pipeline that can handle new data for future
predictions.
• Minimize prediction errors (e.g., RMSE) to support confident decision-making.
2. Insights from Data Exploration
2.1 Dataset Overview
• Size: 16,598 entries, 10 columns.
• Numerical Features: launch_year, sales_usa, sales_europe, sales_asia,
sales_misc, sales_total.
o Skewed distributions: sales_total mean: 0.5374, max: 82.74; regional
sales show similar skewness (e.g., sales_usa max: 41.49, 75th percentile:
0.24).
o Missing values: 271 in launch_year.
• Categorical Features: title_name (11,493 unique), device_type (31 unique),
game_genre (12 unique), publisher_name (578 unique).
o Dominant categories: game_genre (Action, 3,316 entries), device_type
(DS, 2,163 entries), publisher_name (Electronic Arts, 1,351 entries).
o Missing values: 58 in publisher_name.
• Target Variable: sales_total, highly skewed, suggesting the need for
transformation or robust modeling techniques.
2.2 Key Insights
• Skewness and Outliers: Sales columns exhibit extreme outliers (e.g., top games
like Wii Sports with sales_total of 82.74), which may distort model predictions
unless addressed through transformations (e.g., log) or robust scaling.
• Feature Correlations: Regional sales (sales_usa, sales_europe) likely have
strong correlations with sales_total, as they are components of the target. A
correlation heatmap would confirm this.
• Categorical Impact: High-cardinality features like publisher_name and
title_name suggest potential overfitting if encoded directly. game_genre and
device_type may capture market trends (e.g., Action games or DS platform
dominance).
• Temporal Trends: launch_year (mean
