Guide
What Does This Tool Do?
This tool learns "which player wins more on which hero" from your past match data. Every match you record adds to the data, improving recommendation accuracy over time.
The flow is simple
- Record matches (hero compositions and results)
- The tool learns patterns from your data
- Get winning compositions suggested for your next match
It starts with recommendations based on general data (official per-map win rates and composition-type matchups), and as data accumulates, it increasingly reflects individual hero strengths, synergistic pairs, and matchups against the enemy.
Getting Started
Four steps to start using OW Advisor.
Create Account
Register with a username and password. You can also sign in with Google.
Register Party Members
Add your regular teammates on the Group management page. You can set their main roles and hero pools.
Record Matches
Enter map, hero compositions, and match results on the Record page. More data means better recommendations.
Get Recommendations
Visit the Recommend page to get optimal hero composition suggestions for your next match.
Using Recommendations
The Recommend page suggests optimal hero compositions based on your accumulated match data.
Basic Steps
- Select map and side (Attack/Defense)
- Select participating party members
- Optionally enter known enemy heroes
- Click the "Get Recommendations" button
Enemy Information
Enemy hero input is optional, but entering them adds matchups against the enemy (counters) to the recommendation. The enemy tank counts even on its own, and with four or more enemies entered, matchups against the composition type (dive, poke, etc.) are added too. Try entering any enemy heroes you spot at the start of a match.
Reading Results
Results are displayed in score order. Each composition shows a score (%) and a data level (Lv1-3), and the bar above the results shows how much the recommendation rests on your own records (share from your records). Higher scores indicate compositions predicted to win more often, based on official data and your records. Choose from multiple suggestions based on your team's preferences and the situation.
Statistics Dashboard
The Statistics page shows analysis of your accumulated match data across five tabs.
- Overview — Total matches, win rate, recent win/loss trends
- Player — Each member's hero usage and win rates
- Hero — Per-hero usage count and win rate analysis
- Map — Win rates by map, identifying strengths and weaknesses
- Trend — Win rate over time and streak analysis
Some tabs may be unavailable when data is limited. More detailed analysis becomes available as you record more matches.
Meta Analysis
View hero aptitude by map and counter relationships against enemy archetypes (dive, poke, brawl, etc.). This page visualizes the game knowledge (prior distributions) used internally by the recommendation engine.
View Meta Analysis →Using Groups
Groups let you share data with your party members. Match records, players, and statistics are all managed per group.
Creating a Group
A default group is automatically created when you register. You can create additional groups from the Group management page.
Inviting Members
Each group has an invite code. Share this code with friends so they can join your group. You can find the invite code on the Group management page.
Switching Groups
If you belong to multiple groups, you can switch the active group using the dropdown in the navbar. Switching changes the displayed data accordingly.
Data Isolation
Group data is completely isolated. Match records from Group A are invisible to Group B. This lets you manage separate records for different parties.
Understanding Recommendation Scores
The recommendation score is a prediction that starts from official per-map win rates and adds three adjustments estimated from your own records, all in one statistical model.
- Individual Performance — How much better or worse the player does on that hero than the general win rate
- Pair Synergy — Whether two specific players' heroes win more together than their individual strength predicts
- Counter — Whether the player's hero does better than usual against the enemy tank or composition
There is no fixed percentage split: the more records behind a component, the more it influences the recommendation.
How to read scores
Scores are shown as percentages, and 50% is the neutral baseline. For example, "53.0% (+3.0%)" means an outcome 3 points better than an even matchup is predicted. This is a model prediction, not a win rate as such (on Escort/Hybrid maps it predicts "phase success", which includes attack and defense progress). Use it as an index for comparing compositions.
Data Levels
The data level shows how far the recommendation is backed by your own records ("effective games" count older games as less than one).
- Lv1 — Few of your own records yet (mostly individual strengths)
- Lv2 — Every member's hero and every pair in the composition have at least 3 effective games
- Lv3 — Lv2, plus at least 3 effective games of enemy-matchup records for every member (solo: no pair condition)
The bar above the results (share from your records) shows how much the recommendation rests on your own records rather than on official data.
What Happens With Limited Data?
With limited data, recommendations follow official per-map win rates (general data). This is a Bayesian estimation feature that prevents unreliable recommendations when data is scarce, using general data as a starting point and gradually adjusting it with your own records.
When your records start to count
Records take effect component by component. The numbers below are the "effective games" (older games count as less than one) at which your records and general data weigh equally.
- Personal strength — about 16 games of that player on that hero
- Teammate synergy — about 44 games of the same two players on the same heroes
- Enemy matchups — about 44 games of that player's hero against the same enemy tank
Recommendation Stages
Above the results, the page shows how much the recommendation rests on your own records (share from your records) and a stage. The stage depends on that share, not on a match count, and the calculation is the same at every stage.
- Mostly general data (share below 25%)
- Learning from your games (25–60%)
- Mostly your own games (60% or more)
How to Record Matches
Record the hero picked at match start. Even if you swapped mid-match, record the initial pick.
Why initial picks?
This tool optimizes "what to pick first in the next match". Mid-match swap decisions depend on the situation, so we focus on the initial selection.
- Select map, side, and party members
- Choose each member's initial hero pick
- Record the result and save
The impact of swaps on match outcomes is indirectly learned through the counter (enemy matchup) estimates (see FAQ below for details).
FAQ
Is it useful with limited data?
Because we use Bayesian estimation, even with limited data the system provides estimates based on official per-map win rates. It never gives random recommendations, and naturally transitions to personalized recommendations as data grows.
As a guideline, your records start to matter after a dozen or so games on a hero for personal strength, and after a few dozen games of the same pairing for synergy. Even if it is only a "rough guide" at first, the more you use it, the more your own records are reflected.
Is it OK to ignore hero swaps?
We don't record swap details, but their impact is indirectly reflected in the data.
For example, if Player A picks Reinhardt first, the enemy runs dive, they swap but still lose — recording this creates a "Reinhardt × Enemy Dive × Loss" data point. If this pattern continues, Reinhardt's recommendation score naturally drops against dive comps.
In other words, matches where swaps occurred are learned as "signals that the initial pick wasn't optimal." The tool's goal is to recommend optimal initial picks that minimize the need for swapping.
Are scores win rates?
No. A score is the predicted outcome of that composition, not a win rate as such.
It starts from official per-map win rates and adds adjustments for Individual Performance (hero-specific win tendency), Pair Synergy (compatibility with allies), and Counter (advantage against enemy comp), all computed together in one statistical model (on Escort/Hybrid maps it includes attack and defense progress).
50% is the baseline — above it means the composition is predicted to do better than an even matchup, below it worse. Think of it as a comparative index for compositions rather than a probability like "60% win rate."
Do I need to enter enemy info?
Recommendations work without enemy info, but accuracy improves the more you provide.
- No enemy info — Recommendations use individual performance + pair synergy only
- Enemy tank only — Counters are added, reflecting tank matchups
- Four or more enemies — Matchups against the enemy composition type (shares of dive, poke, etc.) are added too
If you can see the enemy team at match start, entering what you can is recommended.
What data does counter analysis use?
No special input needed. Every match record (ally heroes + enemy heroes + result) automatically feeds the counter estimates. Besides matchups against the enemy tank, it also learns matchups against the composition type (dive, poke, etc.) derived from the five enemies' sub-roles.
For example, "Player A's Reinhardt × Enemy Winston × Loss" and "Player B's Genji × Enemy Winston × Win" — these combinations accumulate with each match.
When enough data accumulates, the system automatically learns matchup patterns like "D.Va is better than Reinhardt against Winston." Just keep recording your regular matches.
How is pair synergy determined?
Pair synergy is how much better (or worse) two players do together than their individual strengths predict.
For example, if Player A (Reinhardt) + Player B (Genji) win more than their individual strengths would predict, that's positive synergy. If less, it's negative synergy. Two strong players who simply win a lot together have zero synergy.
When pair data is limited, synergy influence is automatically reduced, growing stronger as data accumulates. Bayesian estimation naturally adjusts the weight based on data volume.
What happens to old match data?
Old data isn't discarded entirely but gradually loses influence.
- Time decay — Exponential decay with a 180-day half-life. Recent matches are weighted more heavily
- Season boundary discount — Weight is multiplied by 0.85 for each season boundary (major patch) crossed
For example, a match from 90 days ago keeps about 71% of its weight from time decay alone; with one season boundary in between it keeps about 60%, with two about 51%. Meta changes come in through the official win rates, which are refreshed every season, while records of slowly changing personal skill and teamwork stay useful for longer.
Statistical Background
For those interested — an overview of the statistical methods behind this tool's recommendations.
Show details
Bayesian Logistic Regression (a Personalized Model Anchored on Official Win Rates)
A logistic regression that takes official per-map win rates as its starting point (offset) and jointly estimates player-hero strengths, pair synergies, and enemy matchups from your group's records. Every component has a Gaussian prior centered on "no deviation from the general data", and the estimate is the posterior mode (MAP). Pairs and counters are estimated as what individual strength cannot explain (interactions), so the same strength is never counted twice.
Bayesian Shrinkage
With limited data, estimates are pulled toward the general data (official win rates) (shrinkage). This provides stable estimates without being swayed by a few biased results. As sufficient data accumulates, individual performance is reflected more strongly.
Time Decay and Confidence
Older match data naturally loses weight through exponential decay (180-day half-life). Separately, data crossing season boundaries (major patches) receives an additional discount (×0.85). Because meta changes come in through the official win rates, refreshed every season, records of slowly changing personal skill and teamwork are kept useful for longer. A match's weight is shared between its attack and defense phases, so an Escort/Hybrid match counts as one match.
For detailed theory, see the Scoring Engine Theory page.