012 - CS2 Match Prediction Engine | Susquehanna International Group

This project builds a predictive analytics engine for Counter-Strike 2 that estimates the probability of each team winning the current round, the current map, and the overall match while a game is in progress. Historical professional demo files are parsed into structured game-state data, including score, economy, player health and equipment, utility, bomb state, player positions, and round time. Live game state comes from a third-party CSTV provider. Both sources are converted into the same standardized game-state format before features are generated, so past and live matches are handled the same way, and every model input is available during live play.

Machine learning models are trained at three levels. The round model works from the state of the active round. The map model builds on the round prediction, the score and recent round history. The match model accounts for BO1, BO3 and BO5 series formats. Rule-based classifiers also label game situations such as buy type (full buy, force buy, eco or save) to give more context about the state of play.

Users interact with the system through a web application which includes two modes. Replay Mode loads historical matches with full playback controls, including skipping between rounds, stepping through time, jumping to kills and changing playback speed. Live Mode streams current game state and updates predictions in real time. Both modes include a scoreboard, a 2D radar of player positions and bomb status, player inventory panels, a kill feed, and charts that show how round, map and match win probabilities change as the game goes on.