Esports-Oracle: An Esports Match Prediction System Based on Self-Training Logistic Regression and DeepSeek Inference Agent

Esports-Oracle is an automatically updated top-tier esports match prediction platform that supports daily win rate prediction and automated deployment for CS2 and LoL matches. Its core innovation lies in combining a rolling window self-training logistic regression model with DeepSeek inference agent to build a self-updating and self-evaluating system. The project uses GitHub Actions scheduled tasks to drive data crawling and model training, and deploys the results to the Vercel platform, providing the latest match insights for esports enthusiasts and analysts.

The project's data mainly comes from Oracle's Elixir's public match CSV files (including player performance, team data, match metadata, etc.). The data partitioning adopts a time-series aware strategy: training data is historical data from 2024-2025, validation data is the latest data from 2026, and local cache stores data from 2024-2026 to ensure fair model evaluation and avoid data leakage.

The core prediction engine is a logistic regression model with the formula p(blue win) = sigmoid(intercept + sum(weight_i * zscore(feature_i))), which has strong interpretability. Features include team-level (Elo difference, recent form, side preference), player-level (recent form, hero proficiency, team hero pool), and version/meta-game level (position hero strength, version experience). The system also supports converting single-game win rate to BO1/BO3/BO5 series win probability.

The DeepSeek inference agent plays three main roles: 1. Analyze unstructured data (player interviews, version update notes, etc.) and extract structured features; 2. Generate natural language analysis reports to explain prediction logic; 3. Detect deviations between predictions and actual results and suggest directions for feature engineering improvement.

The project implements automation through GitHub Actions: scheduled execution of fetch_data.py (to obtain data from 2024-2026) and build_dataset.py to construct datasets. Prediction results are deployed to the Vercel platform, providing low-latency global access and automatic HTTPS. Local development can preview the interface via python3 -m http.server 4173 --directory app.

The project's highlights include: 1. Integration of traditional ML (logistic regression) and LLM (DeepSeek), balancing interpretability and semantic processing capabilities; 2. Best practices for automated operation and maintenance (scheduled data updates, version control, continuous deployment); 3. Feature engineering incorporating esports domain knowledge (Elo rating, hero proficiency, etc.).

The project has two limitations: 1. Data acquisition must comply with source terms (e.g., prohibiting automatic crawling of HLTV); 2. Esports matches are affected by unpredictable factors such as on-the-spot form, so predictions cannot be 100% accurate, and users should use them as a reference rather than an absolute basis.

Esports-Oracle does not replace human analysts but provides support through automated processes and interpretable models. For developers, the project offers valuable references in data pipeline design, feature engineering, and automated deployment, making it a typical case of AI application in vertical fields.