Part 1Marc de Batlle
0%
Part 1Marc de Batlle
All work09 / 11
Ten

Valorant Match PredictorForecasting esports matches with 180+ features.

09
Period2025 · Nov
RoleSolo build
StatusOpen source
Chapter concepts
  1. 1Role: Solo build.
  2. 2Status: Open source.
  3. 3Stack: Python, scikit-learn, XGBoost, LightGBM, Neural nets.
  4. 4Key figure: 180+ engineered features.

An ensemble ML system that predicts professional Valorant match outcomes from temporal form, team chemistry and economy.

The idea

Esports markets misprice matches constantly. If you can model a match better than the crowd, you have an edge. This is my second pass at doing exactly that for professional Valorant.

Exhibit 10-2
Data pipeline
  1. 1.Collect (augment.gg exports)
  2. 2.Convert to CSV
  3. 3.Build team-perspective rows
  4. 4.Engineer 180+ features
  5. 5.Train + cross-validate
  6. 6.Predict with confidence
180+ engineered features
  1. 1Temporal weighting: recent form counts more
  2. 2Opponent-adjusted: strength of schedule
  3. 3Team chemistry: player synergy metrics
  4. 4Map-specific meta: per-map performance
  5. 5Clutch analytics: high-pressure rounds
  6. 6Economic efficiency: spend → impact

The model

No single model wins, so I blend several. A voting ensemble of Random Forest, XGBoost and LightGBM captures the tabular signal; a regularized neural network catches what they miss. The final prediction is a 70/30 blend of ensemble and network.

Exhibit 10-5
Key figures
70/30
ensemble · NN blend
3
boosted models voting
180+
features
At a glance
LanguagePython 3.8+
ModelsRF · XGBoost · LightGBM · NN
Dataaugment.gg match exports
InterfaceJupyter notebooks
Continue
Previous chapterPolymarket Arbitrage Bot
Next chapterVLR.gg Scraper