How the Winio analytics platform works

Winio is a sports and esports analytics platform powered by proprietary machine learning models to predict match outcomes and other key events. Today, the platform covers Dota 2, CS2, and football, with plans to expand into additional esports titles and sports disciplines over time, offering predictions on who will win, how many maps or rounds a series may last, and many other game-specific events.

This document explains how the platform works: where the data comes from, how the models operate, and what technical decisions drive the stability and accuracy of the analytics.

For definitions of key terms used throughout this document, see the Glossary.

Winio does not predict exact match outcomes. The platform calculates probabilities, and is transparent about its confidence level.

Each match page has three tabs:

• 01

  Match overview

  Each team’s win probability as a percentage, alongside the live score.

• 02

  Prediction breakdown

  A written analysis covering what influenced the prediction, how the draft was assessed, and which factors carried the most weight.

• 03

  Analytical widgets

  Key insights at a glance: team form, player statistics, tournament context, head-to-head history, and other important metrics.

Predictions update automatically as new data arrives: first before the draft, then once heroes or maps are known, then continuously during the match. Everything happens on one screen.

This is the core of the platform. Winio’s proprietary multi-layer ML pipeline generates predictions from real match data collected over the past three years.

Three models run sequentially for each match.

• 01

  Pre-draft model

  Estimates win probability before heroes or maps are known. It draws on team form, head-to-head history, and player ratings.

• 02

  Post-draft model

  Activates once the draft is complete. It analyzes the specific hero or map composition and recalculates probabilities based on synergies, counterpicks, and map-specific statistics.

• 03

  Live model

  Runs in real time during the match. It factors in the current score and game state, continuously updating the prediction.

To give the model a meaningful understanding of heroes, each of the 163 heroes is assigned a numerical profile called an embedding. An embedding is a vector of numbers describing the hero through its actual behavior in matches: role, strength at different game stages, and interactions with allies and enemies.

The model trains embeddings on 90,000+ drafts, and updates them as the meta evolves with new patches and balance changes.

Winio maintains its own ratings, separately for teams and for players, across more than 8,000 teams in Dota 2 and CS2.

Player ratings account for individual performance adjusted for role and game context. The methodology is proprietary and central to the quality of Winio’s predictions. Ratings feed into every prediction on the platform.

Winio’s predictions are built on two types of data.

Over three years, the team collected and structured match history across Dota 2 and CS2, covering all patches and all major tournaments. A proprietary parser processes match replay files and extracts data in a structured format: picks and bans, items purchased by the second, kill counts, etc.

Most importantly, the parser captures XYZ coordinates for every hero at every second of the match. The model sees the full flow of the game: where heroes moved, when they used a smoke, and how ganks developed. For example, if the model sees five heroes converging on one point, it recognizes that a gank is going to happen, and can assess how that is likely to affect the rest of the match.

Live match data comes from official providers of tournament organizers, like PGL and ESL. It flows over a WebSocket at broadcast speed, with the same latency as a Twitch stream. The live model immediately sees and factors in everything happening in the game.

The platform consists of a microservices architecture: each component handles a specific part of the logic and can be updated independently of the others. This provides flexibility under load and resilience to individual service failures.

The system processes data asynchronously, and does not block during peak traffic. Queues and caching smooth out load spikes, allowing data processing to unfold in stages, without degrading performance.

The frontend uses server-side rendering (SSR). The server assembles pages and sends them ready to display, reducing load times and dependency on the user’s device. The interface is responsive across desktop and mobile.

Winio processes in-game events in streaming mode, factoring data into calculations as soon as it arrives. When inputs change, predictions recalculate automatically.

Services run in containers, which simplifies scaling and updating individual components without taking the platform offline. Automatic restarts recover from failures. Regular backups reduce the risk of data loss.

The platform continuously monitors infrastructure and surfaces problems before they affect users. A separate data quality layer runs automatically, detecting anomalies and inconsistencies that could compromise prediction accuracy.

The platform uses authentication and role-based access control tied to subscription levels. User sessions are protected by modern token management. Users verify their accounts to reduce the risk of unauthorized access.

Payment logic is decoupled from the analytics components. Winio processes subscriptions through external payment services. Once a payment is confirmed, the system automatically grants the corresponding level of access.

For details, see the Subscription page.