The online gaming industry experiences unpredictable changes in sales and performance metrics due to the continuous advent and improvement of video games. This project takes a deep dive into historical data on video game sales to:
- Understand what the gaming industry market in the last three decades has been like
- Identify key features influencing the performance of these games globally and regionally
- Predict global sales based on relevant features
- Provide a solution to the question: What game feature combinations will turn in high or low sales?
The project workflow below seeks to address these pain points.
- Data Collection
- Data preparation
- Exploratory data analysis
- Impact of features
- Prediction of global sales
- Classifier for sales category
- Model deployment and hosting
- Recommendations
- Future work
Some tools and packages used in the course of this project include:
- Pandas, and NumPy
- KNNImputer
- Seaborn, and matplotlib
- Scikit
- Joblib
- Streamlit
The data used here was obtained from Kaggle. It contains information about video game sales worldwide, including factors such as critic and user reviews, genre, platform, and more. Note that sales are in millions.
| Column Name | Description |
|---|---|
| Name | The name of the video game. |
| Platform | The platform on which the game was released. |
| Year_of_Release | The year in which the game was released. |
| Genre | The genre of the video game. |
| Publisher | The company responsible for publishing the game. |
| NA_Sales | The sales of the game in North America. |
| EU_Sales | The sales of the game in Europe. |
| JP_Sales | The sales of the game in Japan. |
| Other_Sales | The sales of the game in other regions. |
| Global_Sales | The total sales of the game across the world. |
| Critic_Score | The average score given to the game by professional critics. |
| Critic_Count | The number of critics who reviewed the game. |
| User_Score | The average score given to the game by users. |
| User_Count | The number of users who reviewed the game. |
| Developer | The company responsible for developing the game. |
| Rating | The rating of the game. |
All accompanying codes for the next steps are contained in the respective notebooks to be linked. This is done to ensure the brevity and conciseness of this readme. The proceeding steps below will contain the thought process for them, relevant results, and codes if necessary.
Following best practices, a copy of the dataset was made after importation, and all analyses were carried out on that copy. Conducting an initial exploratory data analysis revealed:
- No duplicate rows
- Inappropriate data types for some columns
- Missing data as seen below
- Summary statistics for numerical columns
- Unique values for categorical columns
The data types were handled as needed, the summary statistics provided information on the spread and range of the numerical data, and the categorical columns provided an overview of the unique values present.
Upon further analysis, three categories of missing data were observed. Each was treated differently.
1. Missing Completely at Random(MCAR) Where the probability of a data point missing is entirely unrelated to any other observed/unobserved data. The name and genre columns fell into this category. Since the number of missing values for this was negligible, they were therefore dropped from the dataset
2. Categorical columns like Publisher, rating, etc The NaN values were replaced with "missing" to indicate the unavailability of relevant data.
3. Missing at Random(MAR) This applied to the missing values in the numerical columns where missingness is not completely random but can be explained by some other known information. These rows cannot dropped as that will lead to gross information loss thereby impacting the efficiency of our model and analysis in the future. To handle this, I used a a multivariate approach - the KNNImputer with k=5 nearest neighbors which allows the imputer to find the 5 most similar rows in the dataset and make imputations.
After proper handling of all the cases aforementioned, we have this:
This involved creating new features based on already available information. I created a new feature called release_era that groups the release year of games into three eras - pre-2000s, 2000-2010, and post-2010. This was created to enable me to perform some group-level analysis during the EDA process.
The notebook for a detailed breakdown of the EDA process including univariate and bivariate analysis can be found here. Here, I present some questions I answered and insights revealed from this stage in my analysis.
1. What have the sales through the years been like regionally?
INSIGHT
Generally, sales spiked for all regions in 1995. In North America, there was a wave of fluctuations from 1980 till 1995 when it picked up and rose steadily with a few dips here and there. This reached peak sales of 350 million copies. This was not sustained though, the sales began dwindling through the years till 2016. The sales in all regions were low compared to a decade before that. Sales in Europe and Japan follow a similar pattern. For sales in "other" regions, we see a relatively steady growth from less than one million sales to its highest point around 70 million and then a decline.
2. Is there a correlation between critic scores and user scores? Do they tend to agree or disagree?
INSIGHT
Based on the correlation plot and a heatmap created during the analysis, there is a moderate positive linear association (0.5) between the user score and the critic score. What does this mean?
This means that there is a noticeable trend between the critic scores and the user scores, even though it is not a perfect relationship. It also tells us that professional critics and users tend to agree to some extent on their assessment of video games. Therefore, stakeholders need to consider both critic and user scores when making decisions about game development and marketing.
Want to see more? Check out the complete EDA process here.
The next stage of my analysis sought to answer the question: What effect do the number of critics and users and their review scores have on the sales of video games in North America, Europe, Japan, and "Other" regions?
Find the notebook for this section here.
For this, I conducted an initial examination to find out if the relationship between the variables was linear.
There was no discernable linear relationship between the variables. This was backed further by the results of the R² when I used multiple linear regression. After performing transformations to the data, the relationships between the variables provided no hint of linearity. To answer this question, I therefore used a non-linear model - RandomForest - to determine the feature importance of the variables in regional sales. The results can be seen below.
NORTH AMERICA
EUROPE
JAPAN
"OTHER" regions
INSIGHT
In all the regions, we see that the number of users accounts for a greater percentage of the effect on the sales of video games. This indicates the popularity of games amongst individuals where the more people use the game, the higher the sales in the region can increase.
Second to this is the number of professional critics who review the games. The least contributing factors are the user and critic scores. Though the user and critic score is an important metric to track during and after game development, it is however more important that the game becomes a favorite among the audience while keeping an eye on the user score which provides an understanding of the overall user satisfaction or impression of the game on the audience and critics.
The focus here was to create a model that can predict the global sales. The step-by-step process can be found here.
The steps here included:
- Removal of outliers
- Feature selection: see the selected columns below
- One-hot encoding using
pd.dummiesto transform categorical data into a set of binary columns - Data splitting
- Data rescaling on the train and test data differently to prevent data leakage
- Defining the models/algorithms
##define models
dt_model = tree.DecisionTreeRegressor(max_depth=200)
svr_model = svm.SVR(kernel='rbf')
rf_model = RandomForestRegressor(n_estimators=100, random_state=42)
bayes_model = BayesianRidge(compute_score=True)
lass_model = Lasso(alpha=0.1)
xgb_model = XGBRegressor(n_estimators=100, learning_rate=0.1, random_state=42)
linear_model = LinearRegression()- Training the algorithm
- Tests
- Model evaluation and conclusion The model was evaluated based on the R2 and RSME as seen below:
INSIGHT
The Support Vector regression model is the best-performing model in this comparison, with the lowest RMSE and the highest R² value. It best explains the variability in global sales. The RF and XGB models also perform well, with relatively low RMSE and decent R² values. The DT model however performs the worst as indicated by its high RMSE and negative R² values.
This classifier predicts whether a video game will have high (>= 1 million), or low (< 1 million) global sales based on the inputs given by the user such as:
- Platform: The platform on which the game will be released (e.g., PS2, X360, PS3, Wii, etc.).
- Year of Release: The year in which the game will be released.
- Genre: The genre of the video game (e.g., Action, Adventure, Sports, etc.).
- Publisher: The company responsible for publishing the game.
- Developer: The company responsible for developing the game.
- Critic Score and counts: The average score given and the number of professional critics who review the games.
- User Score and counts: The average score given and the number of users.
- Rating: The rating of the game (e.g., E, T, M, etc.).
This classification model is useful to simulate what a game with certain features/characteristics would sell for, that is if it would have reasonable sales globally.
This notebook contains details on building the classifier model. The steps here included:
- Feature engineering: creating the
global_sales_catbinary feature indicating high or low sales - Feature selection as seen below:
- One-hot encoding using
pd.dummiesto transform categorical data into a set of binary columns - Data splitting
- Data rescaling on the train and test data differently to prevent data leakage
- Training the algorithm and handling data imbalance using
class_weight="balanced" - Tests
- Model evaluation The model was evaluated based on the training score, test score, average cv score, and standard deviation of the cv scores as seen below:
Choosing the best model to use
The significant difference in the training and testing accuracy as can be observed for most models indicate cases of overfit.
The Logistic Regression (LR) model has similar training and testing accuracy with an average cross-validation score of 74% and a low standard deviation (0.14) indicating consistent and reliable performance across different data splits. LR is therefore the best-performing model.
This best-performing model is integrated with an app to give users real-time classification results. The app is built using Streamlit. Find the code here.
The interface for the app looks something like this for dark-themed screens:
You can also interact with the app here
Where do we go from here? What can we do with this information? For stakeholders looking to bring a new game into the market:
1. North America is a major player in the gaming industry. To capture and drive more sales globally, we need to penetrate the market there and garner as much user/fan base and positive reviews as needed to skyrocket sales.
2. Choice of game genre
- Action games are considered the most popular genre of all time, globally. In Japan, however, we may want to focus on role-playing games to leverage Japan's preference for this genre to drive sales from that region.
- Based on reviews, critics and users prefer Role-playing and fighting games, however, action, shooter, and role-playing games have a larger fan base.
For a new game release, we can consider building a game that belongs to any of the following categories - Action, or Role-playing to drive more sales globally.
3. Game acceptability (ratings) It is recommended that games should be designed for either E(Everyone), T(Teens ==> 13 years and older), or M ( Mature ==> 17 years and older) to reach more users. This would also impact the return on sales as more people can access the games released by virtue of the game's suitability with the audience.
4. Platform PS2, X360, and PS3 are household names known to gamers and users globally. In the past four decades, these platforms have contributed the most to global sales. It is recommended that a new game to be released, use any of these platforms to leverage the familiarity of these platforms with the users to drive sales.
The results from this work can be extended to improve its functionality and effectiveness by:
- Investigating the distribution of users in each region to gauge the receptivity of these games in the respective regions
- Carrying out a comparative analysis on the performance of games from the point of release till date.