itaimorag/Video-Games-Sales-EDA

🎮 Video Game Sales — Exploratory Data Analysis (EDA)

Video Presentation

If I didn’t cover everything it’s because I didn’t have enough time

Executive Summary

The video game industry is a multi‑billion dollar market characterized by extreme unpredictability—a single "mega‑hit" can generate more revenue than thousands of average games combined. This project analyzes historical video game sales (1980–2016) to uncover the main patterns behind commercial success.

Key findings: Action and Sports genres have the highest historical sales volumes, but the market is highly volatile, era‑dependent, and dominated by a small number of outlier hits. Higher critic scores raise the ceiling of potential sales and are more predictive than user scores, but they do not guarantee commercial success, and regional tastes differ substantially.

Methodology extras (notebook): Extreme sellers are named (top titles) and the long tail is summarized with Tukey IQR fences on Global_Sales—many rows above the upper fence are expected for hit‑driven sales, and those rows are kept as legitimate data. A Mann–Whitney U test (median split on Critic_Score) provides a non‑parametric check on long‑tailed sales; it supports a rank‑based association, not causality.

Engineered fields & missing metadata: The notebook adds Is_Hit (top quartile of Global_Sales), Year_missing (flag for unknown release year), and fills Publisher / Developer missing values with Unknown (no fake company names). Review scores stay NaN where unknown. A short sensitivity table compares a median‑based hit rule vs the 75th‑percentile default.

---

Navigation

• Executive Summary
• 1. Dataset Description
• 2. Intended Uses & Audience
• 3. Dataset Structure
• 4. Data Integrity & Cleaning
• 5. Exploratory Data Analysis Highlights
  • 5.1 Market Size & Global Trends
  • 5.2 Demographics & Audience
  • 5.3 Quality vs Commercial Success
  • 5.4 Hit label, missing metadata & sensitivity
• 6. Machine Learning Readiness
• 7. Strategic Takeaways
• 8. Limitations
• 9. Notebook & Libraries
• 10. Author

---

1. Dataset Description

This dataset contains historical sales data and review scores for video games released between 1980 and 2016.
Each row represents a single game release.

1.1 Source

The data is based on public video game sales and review aggregators, cleaned and repackaged here for EDA and ML tasks.

• Upstream dataset (Kaggle): Video Game Sales with Ratings (curator: rush4ratio).

1.2 Features

• Identity
  • Name, Platform, Year_of_Release, Publisher
• Categorization
  • Genre, Rating (ESRB)
• Financials (in millions of units)
  • NA_Sales, EU_Sales, JP_Sales, Other_Sales, Global_Sales
• Reception
  • Critic_Score, Critic_Count, User_Score, User_Count
• Additional
  • Developer

Engineered in the notebook (after cleaning) — same rows as the cleaned table, extra columns for analysis and ML prep:

Column	Meaning
Is_Hit	1 if Global_Sales ≥ 75th percentile on the cleaned data, else 0.
Year_missing	1 if Year_of_Release is NaN, else 0.
Publisher / Developer	Missing values replaced with the literal label Unknown.

1.3 Dataset card & research framing

• Unit of analysis: One row per game title with aggregated regional and global unit sales (not weekly time series).
• Research focus: How genre, region, ESRB rating, and review scores relate to global sales, and how genre‑level demand changed over 1980–2016.
• Modeling note: Global_Sales is a natural regression target; Genre, Platform, Publisher, Rating, and score/count fields are natural predictors for later supervised learning.

---

2. Intended Uses & Audience

• Publishers & investors – estimate market size and historical performance by genre, platform, and publisher.
• Indie developers & analysts – understand genre saturation, realistic sales expectations, and promising niches.
• Data scientists & ML practitioners – build models for:
  • sales prediction,
  • hit vs non‑hit classification,
  • portfolio and trend analysis.

Out-of-scope use

• Not for post-2016 forecasting without new data; trends, platforms, and digital share changed after the coverage window.
• Not for causal business decisions based on EDA alone (e.g. “publish Action → profit”); associations in the notebook are exploratory, not proven causal effects.
• Not a substitute for revenue data; units sold are not price-adjusted and omit many modern monetization models.

---

3. Dataset Structure

• Rows: ~16.7k games (after cleaning; exact count in notebook output)
• Columns: 16 fields from the source file after cleaning, plus Year_missing and Is_Hit (18 columns in the main analysis DataFrame). Publisher / Developer are still those two columns, with missing values recoded to Unknown.
• Split: single train split (users create their own train/val/test).

3.1 Sample Rows

Before cleaning

---

4. Data Integrity & Cleaning

Raw game sales data is noisy and inconsistent. These steps were taken to make it analysis‑ready.

4.1 Initial State

• Missing identifiers and partial metadata for older titles
• A few implausible Year_of_Release values
• User_Score as strings with "tbd", forcing an object dtype
• Strong numerical outliers in Global_Sales (e.g., Wii Sports)

Missingness reporting: Both raw counts and percentage of rows per column were printed so sparse columns (e.g. scores on older titles) are easy to compare at a glance.

Pre‑cleaning snapshot:

4.2 Cleaning Decisions

• Removed games with Year_of_Release after 2016 to avoid partially observed recent years.
• Converted "tbd" in User_Score to NaN and cast back to float.
• Left missing Critic_Score / User_Score as NaN to avoid fabricating scores for eras without aggregators.
• Publisher / Developer: missing → Unknown (explicit category for groupbys; not a real company name).
• Year_of_Release: still NaN when unknown; Year_missing = 1 on those rows (no median year imputation).
• Is_Hit: binary label, 1 if Global_Sales ≥ 75th percentile (see §5.4 and notebook).
• Retained extreme best‑sellers; handled their influence via log scaling (or y‑axis caps) in plots rather than dropping them.
• Categorical profiling: After cleaning, object / string columns were summarized with describe(include=['object']) (counts, uniques, top category) to spot sparse labels and typos before plotting. After Developer.fillna("Unknown"), top may show Unknown if it is the mode—compare freq to count (see notebook missing‑data policy).

Post‑cleaning snapshot:

4.3 Summary Statistics After Cleaning

4.4 Sanity checks (domain rules)

Automated plausibility checks on the cleaned data:

• All regional and global sales columns are non‑negative.
• Year_of_Release lies in a sensible range for this table (and > 2016 rows were already removed).
• Critic_Score, where present, lies in 0–100.

These catch scrape errors, wrong units, or bad merges before trusting aggregate charts.

4.5 Outlier documentation (Global_Sales)

• Top titles table: The notebook lists the top 10 games by Global_Sales so mega‑hits (Wii Sports, etc.) are explicit, not only visible as scatter extremes.
• Tukey IQR fences: Lower fence = Q1 − 1.5×IQR, upper fence = Q3 + 1.5×IQR. For heavily right‑skewed game sales, many rows can exceed the upper fence by expectation; that is interpreted as structural hit‑driven skew, not automatic grounds to delete rows.
• Decision: Keep those rows as real sales; use log scales, caps, or robust methods in models as needed.

---

5. Exploratory Data Analysis Highlights

5.1 Market Size & Global Trends

A. Total Global Sales by Genre

Question: What are the most profitable genres of all time (by total units sold)?

• Insight: Action and Sports dominate absolute sales volumes, with Shooters also crossing the billion‑unit mark.

B. Share of Global Sales by Genre (%)

Question: What fraction of all global sales in the dataset does each genre represent (1980–2016)?

Totals (panel A) show scale; percentages show market composition (mix), which is how analysts often report structure alongside magnitude.

• Insight: The same leaders tend to dominate both totals and shares, but the percentage view makes relative weighting explicit for storytelling and slides.

C. Genre Popularity Over Time

Static trend of top genres

• Insight: Platformers and Puzzle titles were strong early; Action and Shooters rise sharply in the 2000s. Genre viability is era‑dependent.

D. The Titans of the Industry

Question: Which publishers dominate lifetime global sales?

• Insight: A small number of publishers (Nintendo, EA, Activision, etc.) control a large share of total units sold.

---

5.2 Demographics & Audience

A. Regional Taste Differences

Question: Do North America, Europe, and Japan prefer different genres?

• Insight: NA and EU lean toward Action/Sports/Shooters; Japan strongly prefers Role‑Playing games and contributes little to Shooters.

B. ESRB Age Ratings and Sales

Question: Does restricting a game to a mature audience limit its sales potential?

• Insight: Medians are similar across E, T, M, E10+, but E and M have the highest outliers. Both family‑friendly and mature games can reach very high sales.

C. Sales Distribution by Genre (Log Scale)

Question: What does typical performance look like within each genre once we control for outliers?

• Insight: The median game in almost any genre sells well under one million units; huge totals are driven by a few extreme hits.

---

5.3 Quality vs Commercial Success

A. Feature Correlations

Question: How do review scores relate to sales numerically?

• Insight:
  • Regional sales correlate strongly with Global_Sales (by construction—they are components of global totals in this dataset).
  • Critic_Score has a moderate positive correlation (~0.24–0.25) with Global_Sales.
  • User_Score shows a weaker correlation with Global_Sales.
  • Critic_Score and User_Score correlate moderately with each other.
  • Is_Hit vs continuous columns are point‑biserial correlations; Is_Hit vs Global_Sales is near‑perfect by construction (label derived from global sales).
  • Year_missing vs Year_of_Release: pairwise correlation is a degenerate diagnostic (missing year aligns with the flag); read the notebook note, not as discovery.
• Interpretation: Professional reviews are a better linear predictor of sales than user scores, but still far from deterministic.

B. Distribution of Professional Critic Scores

Question: How are critic scores distributed overall?

• Insight: Scores are slightly left‑skewed and heavily clustered between 65 and 85. In practice, ~70+ behaves like the “average” functioning game.

C. Critic Scores vs Global Sales

Question: Do higher critic scores actually translate into more sales?

• Insight: Very high‑selling games almost all have critic scores above ~80, but many highly rated games still sell modestly. A high score raises the ceiling more than it guarantees a result.

D. Critic vs User Alignment

Question: Do critics and players agree on quality?

• Insight: Points follow an overall upward trend but with wide dispersion above and below the 45° line, indicating both agreement and strong disagreements (e.g., cult classics or controversial titles).

E. Stratified view (top genres) & Mann–Whitney U check

Why facet: The global scatter (panel C) mixes all genres. The notebook adds faceted critic‑vs‑sales plots for the three genres with the highest lifetime global sales, with the same y‑axis cap (~30M) as the main scatter so the bulk of the distribution stays visible.

Mann–Whitney U: Sales are long‑tailed, so a simple t‑test is a poor default. The notebook splits games at the median Critic_Score (among rows with non‑missing score and sales) and runs mannwhitneyu(..., alternative='greater'): it asks whether the high‑score group has stochastically larger sales than the low‑score group (rank‑based). Interpretation: A small p‑value supports an association in this non‑parametric sense; it does not prove causality (genre, IP, marketing, and platform still dominate outcomes)

---

5.4 Hit label, missing metadata & sensitivity

The notebook defines a commercial hit as Is_Hit = 1 when Global_Sales is at or above the dataset 75th percentile (≈25% positives). It also compares this to a median‑threshold rule (50% positives) in a small sensitivity table—useful for discussing how sensitive conclusions are to the cutoff.

1. Final feature summary
   

2. Hit rate by genre

3. Hit rate by platform

4. Critic score by Is_Hit (box plot)

5. Mann–Whitney on Critic_Score by Is_Hit — text output with statistic and p‑value.
   

6. Sensitivity table — the display(sens) table (median vs 75th percentile cutoffs, n_hit, hit_rate).
   

---

6. Machine Learning Readiness

• Scaling / normalization
  • Global_Sales (millions), Critic_Score (0–100), and User_Score (0–10) live on different scales.
  • Use StandardScaler or MinMaxScaler for distance‑based or gradient‑based models.
• Multicollinearity & target leakage
  • NA_Sales, EU_Sales, JP_Sales, and Other_Sales are not independent of Global_Sales in this table—they are pieces of the global total. Do not treat all region columns plus Global_Sales as unconstrained features and target without a clear design (e.g., predicting regions separately, or predicting global while dropping regions, or using shares with care). Otherwise you risk redundancy or leakage depending on how the target is defined.
• Targets
  • Regression: Global_Sales or log(Global_Sales + ε) to reduce skew.
  • Classification: use the notebook’s Is_Hit (75th percentile of Global_Sales) or define your own cutoff—document it. Do not use Is_Hit as a feature when predicting Global_Sales (leakage).
• Useful features
  • Categorical: Genre, Platform, Publisher (includes Unknown), Developer (includes Unknown), Rating, binned Year_of_Release where year exists
  • Numerical: Critic_Score, User_Score, Critic_Count, User_Count, Year_missing
  • Engineered: critic–user gap, regional shares (if not leaking the target), log‑sales, time‑decayed features; Is_Hit only as a label, not a regressor for sales.

---

7. Strategic Takeaways

1. Genre shapes opportunity size.
   Action, Sports, and Shooters reach the largest audiences but are also the most saturated genres.
2. Most games are modest sellers.
   The market is strongly hit‑driven: medians are low, and a handful of blockbusters carry totals.
3. Reviews help, but aren’t everything.
   Higher critic scores are associated with higher potential sales, but IP strength, marketing, and platform reach remain crucial.
4. Market power is concentrated.
   Major publishers start with a much higher expected baseline than small studios.
5. Audience targeting matters.
   Both E and M‑rated games can perform extremely well if genre, platform, and marketing match the intended demographic and region.

---

8. Limitations

• Dataset coverage effectively ends in 2016 and does not fully capture the shift to digital‑only, live‑service, or mobile ecosystems.
• Many older titles lack Critic_Score / User_Score due to the absence of historical aggregators.
• All relationships are correlational, not causal (including the Mann–Whitney result: it is a distributional comparison, not proof that “better scores cause sales”).
• Hit-driven tails: IQR-style outlier counts on Global_Sales can be large without implying bad data—interpret alongside the top‑title table and domain knowledge.
• Representation & coverage: Sales figures reflect the retail / physical‑heavy era and source reporting practices; NA, EU, and JP are coarse regions and do not represent the full global market. Critic and user scores favor titles and periods covered by major aggregators, with potential English / Western tilt and survivorship (only games present in the source appear at all).

---

9. Notebook & Libraries

The analysis was performed in Google Colab using Python.
Main libraries used:

• Data manipulation: pandas, numpy
• Visualization: matplotlib, seaborn
• Utilities (download/export): google.colab.files

Full analysis with code and interactive plots: Google Colab

10. Author

Itay Morag

---