Valorant-Analysis-Matchs

We will analyze player statistics, such as maps played, and complete missing information in order to build a standard predictive model

Valorant Competitive Analysis: Map-Level Performance and Match Outcome Prediction

Table of Contents

• Objective
• Dataset
• Tools
• Key Findings
• Visualizations
• Future Work

Limitations

• Player-level statistics are aggregated at the map level for simplicity.
• Agent composition is simplified and does not account for full team synergies.
• Economy and round-level data are not included in the current analysis.

These limitations were intentional to keep the analysis focused on map-level performance and model interpretability.

Objective

• Analyze player performance across maps
• Identify patterns in key performance metrics
• Handle missing and inconsistent real-world data

Dataset

• Source: VLR.gg (scraped / collected)
• Data includes player statistics per match
• Missing values present in some matches

Tools

• Python
• pandas, numpy
• matplotlib, seaborn
• scikit-learn (modeling, pipelines, evaluation)

Key Findings

• Player performance varies significantly by map, particularly in ACS and Rating.
• High-winrate teams show non-random loss patterns against specific opponents.
• Advanced performance metrics (KAST%, ADR, ACS) are more predictive of map outcomes than raw kill counts alone.
• Missing values required careful filtering to avoid bias and data leakage.

Opponent-based loss patterns (advanced analysis)

This visualization compares map losses by opponent for high-winrate teams. It highlights potential counter-play styles and matchup dependencies, and serves as a deeper exploratory analysis rather than a headline result.

Key insights:

• Certain opponents systematically dominate specific map pools.

• Loss distribution suggests preparation gaps rather than random variance.

• This pattern could be used to adjust veto strategies or training focus.

  Model performance (ROC curve)

The ROC curve evaluates the discriminative ability of the Random Forest model to predict map-level outcomes. The model achieves a strong ROC-AUC score, indicating good separation between wins and losses across different thresholds.

Feature importance (Random Forest)

This visualization shows the top 20 most important features used by the Random Forest model to predict map outcomes.

Key observations:

• Performance metrics such as Rating, ACS, KAST%, ADR and KD are the strongest contributors to map-level win probability.
• Contextual categorical variables (Map, Team, Opponent, Region) also play a significant role after one-hot encoding.
• The results align with competitive intuition: consistent round impact and efficiency matter more than raw kill counts alone.

This analysis improves model transparency and helps connect statistical results with in-game performance factors.

Future Work

• Incorporate agent composition and role-based features.
• Extend predictions to best-of series outcomes.