Automated Sport Prediction Analyzer

AI Sport Prediction Analyzer Documentation

Workflow: Sport Predictions Stats for upcoming Events By: BlockShield Systems

1. Overview

An AI-powered sports intelligence system that automatically discovers upcoming events across 8 sports, researches historical and real-time data, generates mathematically rigorous 3-layer predictions, and delivers results via 4 communication channels.

Core Pipeline:

1. Event Discovery
2. Parsing and Filtering
3. Parallel Research (Stats and News)
4. Data Fusion
5. 3-Layer AI Prediction
6. Multi-Channel Delivery (Email, Telegram, Notion, Archive)

Schedule (UTC): 0 1,7,13,19 * * * (01:00, 07:00, 13:00, 19:00 daily, each scanning an 8-hour event window).

2. Sports Covered

3. Architecture Phase 1: Event Discovery and Parsing

1. Event Discovery Agent: An input-mode AI agent browses for sports and web-searches for eSports and Boxing.
2. Event Parser and Filter: A code node enforces strict rules, requiring specific keywords, an ISO datetime, an 8-hour time window, filtering junk entries, and capping at 20 events. It extracts the sport, league, datetime, teams, venue, and importance.

4. Architecture Phase 2: Parallel Research

Both agents run in item mode with error continuation enabled so individual failures do not crash the pipeline.

1. Historical Stats Agent: Performs two searches per match for head-to-head results and standings/form. Outputs head-to-head summaries, recent form, sport-specific KPIs, and standings.
2. Injury and News Agent: Performs two searches per match for injuries, lineups, predictions, and odds. Outputs injury reports, expert sentiment, and external factors.
3. Merge, Clean, and Deduplicate: Combines branches per match, normalizes team names, applies freshness scoring, and deduplicates into unified match objects.

5. Architecture Phase 3: Prediction Engine

An item-mode AI agent performs pure mathematical reasoning on pre-researched data without using external tools. It outputs structured JSON per match. A cleanup code node then strips verbose reasoning fields before delivery to minimize payload size.

6. Architecture Phase 4: Multi-Channel Output

7. 3-Layer Prediction Methodology

Each match is independently analyzed through three layers, scoring between 0.00 and 1.00. Probabilities sum to exactly 100%.

1. Layer 1 Bayesian Analysis: Prior probability from head-to-head and team strength, adjusted by recent form and home/away multiplier.
2. Layer 2 ML Feature Weighting: Weighted sum of seven features including recent form, injury impact, and expert consensus.
3. Layer 3 Historical Pattern Matching: Cross-validates against historically similar scenarios to detect anomalies.

8. Interpretation of Scores

9. Connectors and Configuration

Connectors Used:

1. Microsoft Outlook (OAuth 2.0)
2. Telegram Bot API (API token)
3. Notion MCP (OAuth 2.0)

10. Key Architecture Decisions

1. Template Injected Prompts: The HTML Email Agent uses template syntax to inject actual match data directly into the prompt, preventing AI hallucination.
2. Expression Based Notion Pages: Uses expression mode for deterministic page creation.
3. No Tool Prediction: The Prediction Engine has zero tools, forcing pure mathematical reasoning on pre-researched data only.
4. Meta Cleanup: Strips verbose reasoning fields before output channels.
5. Parallel Research: Both research agents continue processing remaining matches even if individual calls fail.

11. Known Limitations

1. Dynamic Rendering: Some pages require browser mode and may not fully load. Mitigated by web search fallback.
2. Telegram Limits: Large reports get truncated at 4096 characters. Auto-truncation appends a note to check the email.
3. Research Data Gaps: Lower-tier leagues often lack head-to-head data. Handled via freshness scoring.
4. eSports Parsing: Events do not always follow standard naming formats; unparseable entries are filtered out.
5. No Error Monitoring Pipeline: Relies on continuation flags rather than a dedicated error-alert system.

12. Roadmap

1. Error monitoring pipeline
2. Google Sheets logging for prediction audit trails
3. Post-match result verification for accuracy measurement
4. Collection hygiene (auto-delete reports older than 30 days)
5. Multi-message Telegram support for large reports