Machine Learning
Prediction Engine
Two game-specific ML models generate every prop line on the platform. Each outputs a predicted stat line with a confidence score.
Why RandomForest?
Valorant stats are low-variance with strong categorical predictors (agent, map, role). RandomForest handles these well without overfitting on smaller esports datasets. The ensemble of 200 decision trees provides natural confidence scoring through inter-tree variance.
RandomForestRegressorValorant
Evaluated on holdout test set from 2024–2025 seasons
MAE (Kills)
2.31
R² Score
0.71
RMSE
3.14
Training Maps
8.2K
Feature Importance
Top features ranked by model importance for Valorant
avg_kills_last_5
92%
avg_kills_last_10
85%
map_encoded
72%
agent_encoded
68%
team_win_rate_last_10
61%
opponent_strength
54%
avg_deaths_last_5
45%
avg_first_bloods_last_5
38%
Feature Engineering
All input features for Valorant model
avg_kills_last_5avg_kills_last_10std_kills_last_10agent_encodedmap_encodedrole_categoryteam_win_rate_last_10opponent_strengthevent_tieravg_deaths_last_5avg_assists_last_5avg_first_bloods_last_5avg_headshot_pct_last_5
Confidence Scoring
Variance across 200 decision trees — low inter-tree disagreement means high confidence.
80%+ High
65–79% Medium
<65% Low
Inference Pipeline
End-to-end flow from data ingestion to prop line generation
1 PandaScore API
→
2 Feature Extraction
→
3 RandomForest Model
→
4 Confidence Score
→
5 Prop Line + Direction
Tech Stack
Tools powering the prediction engine
scikit-learnpandasPandaScore APISupabaseNext.jsVercelPython 3.12RandomForestGradientBoosting
Roadmap
Completed milestones and planned improvements
✓RandomForest baseline model (Valorant kills)
✓GradientBoosting model (CoD kills)
✓Automated PandaScore data pipeline
✓Map-scoped prop types (guaranteed maps only)
✓Confidence scoring system
XGBoost ensemble for improved CoD accuracy
Additional props: ACS, Headshot %, First Bloods
Live odds adjustments mid-series
Agent/map interaction features
SHAP explanations per prediction
BoardLineupsRanksEngine