Prediction methodology

1. Collect and check the match data

Our scraping services collect professional schedules, results and available match statistics. League of Legends coverage includes gol.gg; schedules and other game coverage also use sources such as Liquipedia and VSGG. Coverage depends on the competition and source. Missing data should not be interpreted as a team having no history or as an event being canceled.

The fixture hubs show the latest stored match-data update when available. That timestamp does not mean every fixture, model forecast or bookmaker quote refreshed together. Check the individual match details before relying on a schedule or price.

2. Build a picture of the matchup

Features summarize the historical evidence available to a model. They vary by game, prediction target and configuration. League of Legends features include rolling team form and head-to-head results, with player-role, champion and objective-tempo features where enabled and covered by the source data.

Historical features are constructed relative to the match date, using earlier games rather than the result being predicted. This point-in-time approach reduces future-data leakage. It does not eliminate source errors, incomplete histories or uncertainty about a newly announced roster.

The contextual guides on our match hubs explain useful analytical checks. A factor mentioned in a guide, such as a map veto or a patch change, is not necessarily an input to every deployed model. Editorial copy and model probabilities serve different purposes.

3. Train and evaluate the models

The training pipeline keeps matches in chronological order. Earlier data is used to fit models, with later slices used for validation and, where appropriate, probability calibration. Keeping time order helps test how a model behaves on newer matches rather than a randomly mixed sample of the same period.

Classification targets estimate event probabilities, such as a team winning. Regression targets estimate quantities, such as total kills or duration. These are different tasks: a predicted duration is a point estimate, not a probability or a promised finishing time.

Calibration adjusts probability estimates against held-out results. The pipeline can skip or reject a calibration adjustment when the sample is insufficient or the adjustment fails its safeguards. The method and available evaluation measures vary by game, target and model version.

4. Read a probability correctly

Illustrative example: a 65% win forecast for Team A leaves a 35% chance of Team A losing in a two-outcome model. It does not mean the team is certain to win, or that it will win exactly 65 of the next 100 matches. If a model is well calibrated, many comparable forecasts near 65% should win roughly that often over a sufficiently large sample.

Bookmaker odds are market prices, not model probabilities. Decimal odds of 2.00 imply a raw probability of 50% using 1 ÷ odds; the implied probabilities across a market can sum above 100% because of the bookmaker margin. A difference between a model and a market is not proof that either is correct or that a wager will be profitable.

Unavailable probabilities are labeled as unavailable or locked. A missing forecast is not a 50% prediction. Access restrictions do not change the uncertainty of an estimate.

5. Assess the published track record

Use the track record to inspect settled forecasts and their sample sizes. Historical backfill is excluded from public performance results; reconstructing an earlier forecast after the fact is different from recording a forecast prospectively. Training evaluation and the public track record should also be interpreted separately.

Accuracy counts how often the selected outcome was correct. Calibration compares predicted probabilities with observed outcomes. For example, a model that calls every favorite a 90% chance can get many winners right while still being overconfident. Look at both measures where available.

• Compare the same game and target over a stated date range; series winners and individual objectives are different prediction tasks.
• Check how many settled predictions support a percentage. A small sample can swing sharply after only a few results.
• Inspect recent results as well as longer-term trends. Patches, roster changes and tournament formats can change performance.
• Accuracy is not profitability. Odds, margins, price changes and the timing of a decision affect any market comparison.

Explore settled forecasts and performance

Dated evaluation summary: series winners

This fixed snapshot was retrieved from the public performance API on October 5, 2026. It summarizes series-winner predictions for each game across model versions; it is not an evaluation of a single current model. Other targets and regression estimates are outside this table. Follow the track record for updated figures.

The API excludes rows flagged as historical backfill. That flag is not an independent audit proving that every included forecast was recorded before kickoff. The public aggregates do not expose model-version breakdowns or exact forecast-generation and match-date ranges, so those details are not inferred here.

The date range shows the first and last UTC week-starts containing settled series-winner predictions in the trend response. It is a settlement-time range, not a training window or exact match-date range. The current week is partial. Separate API requests and caches can produce slightly different samples.

Calibration gap is the sample-weighted absolute difference between mean predicted probability and observed outcome rate across the API’s probability bins (decile ECE). Lower is better on this sample; it does not guarantee future calibration. The bins assess the reference team’s win probability, not the confidence of the selected pick. For Dota 2 best-of-two matches, this does not measure calibration of the full win/draw/loss distribution. The calibration sample is shown separately from the accuracy sample. These figures are descriptive, not confidence intervals or evidence of profitability.

Snapshot retrieved (UTC): 2026-10-05T09:41:02.248496+00:00

Public API sources (current responses may differ from this snapshot):

• League of Legends / series winner: accuracy · settlement trend · calibration bins
• CS2 (experimental) / series winner: accuracy · settlement trend · calibration bins
• Dota 2 (experimental) / series winner: accuracy · settlement trend · calibration bins

Settlement and corrections: what counts as a result

For classification predictions, settlement compares the stored predicted side with the resolved outcome. A prediction is correct when those sides match. Regression predictions store the resolved numeric value separately; classification accuracy does not describe regression error.

Settlement depends on the target’s required source data. A completed status alone may be insufficient: League of Legends series winners use enriched game results, while CS2 and Dota 2 series winners use a completed match and its recorded winner. Individual objectives and map or game targets require their corresponding result data.

• Canceled or postponed series without a completed, resolvable outcome stay unsettled. An unresolved prediction is not counted as an incorrect settled prediction. Independently resolved game or map targets can still settle.
• Missing results or incomplete statistics can delay settlement. There is no guaranteed settlement deadline, and the settlement timestamp can be later than the match itself.
• There is no universal forfeit override: the current outcome queries use the completed status and available result data. A forfeit can settle if it supplies the required outcome; absent that evidence, it remains unresolved. Platform scoring is separate from a bookmaker’s settlement rules.
• A completed Dota 2 best-of-two scored 1–1 resolves as a draw. For series-winner accuracy, the stored prediction is compared with that draw outcome. Associated two-sided value bets are voided with zero profit/loss when the outcome is outside their two-side vocabulary; prediction scoring and bet scoring are different.
• Already-settled rows are not automatically regraded when a source changes. Report a revised score or incorrect outcome with the match URL and supporting evidence. A correction needs investigation; automatic historical recalculation is not promised.

Where the evidence is weaker

New teams, roster substitutions, sparse regional coverage, patches and format changes can make historical form less representative. Some feature pipelines use baseline values when there is too little usable history. A displayed probability alone does not tell you how complete the underlying data is.

CS2 and Dota 2 models are marked experimental on the platform. Treat their track records separately from League of Legends, and check the sample size for the specific target. We do not claim independent certification, guaranteed returns or a universal accuracy rate.

Who maintains this page and how to report an issue

zest.win is operated by Zest Interactive LLC. This page describes the platform methodology; it is not an independent audit of each model. Forecasts are algorithmic, while explanatory guides and methodology copy provide editorial context.

Email hello@zest.win with the match URL, the incorrect field, the time you noticed it and a supporting source if available. Please distinguish a source-data error from a forecast that simply did not occur: an uncertain outcome going the other way is not itself a data error.

Email hello@zest.win

Responsible use

Models and statistics are for analytical and informational purposes only. They are not betting or financial advice, and no forecast guarantees a result. If you choose to gamble, meet the legal age requirement in your jurisdiction, follow local laws, set spending and time limits, and never chase losses. If gambling causes harm, stop and seek support from a qualified local service.