Inside Our AI: Understanding the Features That Drive Better League Games

Want AI‑powered coaching that respects your privacy? Join the waitlist to get early access to ArcaThread AI Coach.

What goes into an AI coach?

If you’ve ever wondered how a virtual assistant can tell you when to contest dragon, recall safely, or buy anti‑heal, the answer isn’t “magic.” It’s a carefully selected set of signals that describe the game state. In ArcaThread’s case, we’ve verified three distinct models, each operating on its own feature space. Rather than one monolithic network, we use purpose‑built predictors for draft decisions, item choices and post‑game mistake detection.

We’ll explain the types of information these models use—without exposing proprietary training details or sensitive storage. The goal is to be transparent about what matters, not how to copy it.

Champion representations: the 20‑dimensional blueprint

All models begin with the same foundation: a champion embedding. Each champion is represented as a 20‑dimensional vector capturing role affinity, damage profile, combat style, mobility, utility, scaling and durability【893838282729953†L16-L27】. This embedding compresses a champion’s characteristics into a format a model can consume. For example, high mobility champions have larger values in the mobility dimensions, while durable tanks score high in durability. These embeddings feed into the features described below.

Draft decisions: 68 features in champion select

During champion select, our draft predictor weighs 68 separate features to estimate win probability. Each feature falls into one of several categories【893838282729953†L117-L122】:

• Champion embeddings (40 dims) – vectors for both your pick and your likely lane opponent【893838282729953†L124-L132】.
• Matchup and synergy priors (6 dims) – aggregated win rates and synergy scores based on rank and patch【893838282729953†L124-L132】.
• Role encoding (5 dims) – probabilities that each pick will play top, jungle, mid, bot or support【893838282729953†L124-L132】.
• Team composition gaps (12 dims) – measures of frontline, magic damage, engage and wave clear needs【608969380721136†L35-L41】.
• Historical context (1 dim) – smoothed win‑rate prior for the champion【893838282729953†L124-L132】.
• Damage‑type matchups (2 dims) – physical vs magic damage mix【893838282729953†L124-L132】.
• Scaling timeline (2 dims) – early‑ vs late‑game power curves【893838282729953†L124-L132】.

By capturing these aspects, the model can suggest which champions increase your team’s odds of winning. It doesn’t peek at private data; it works solely from champion identities, role probabilities and aggregated matchup statistics.

Item choices: 47 signals to guide your build

Once the match begins, your shop decisions depend on context. Our item optimizer operates on 47 features, including:

• Champion embedding (20 dims) – summarising your champion’s inherent strengths【893838282729953†L124-L132】.
• Role encoding (5 dims) – clarifying your lane position【893838282729953†L124-L132】.
• Current inventory (6 dims) – how many item slots are occupied, boots tier and progress on core items【608969380721136†L76-L79】.
• Enemy team profile (10 dims) – aggregated threat weights, damage mix and healing saturation【608969380721136†L62-L74】.
• Game tempo (5 dims) – time buckets, gold efficiency gap and resource pacing【608969380721136†L52-L61】.
• Efficiency indicator (1 dim) – a simple measure of KDA or performance to adjust recommendations【893838282729953†L134-L142】.

These features let the model decide if anti‑heal is urgent, whether armor or magic resist will help, and when to buy economy items. As with draft, the input signals come from in‑game statistics—no player credentials, no external data.

Mistake detection: 28 dimensions in post‑game analysis

After the match, our mistake classifier reviews performance using 28 features【893838282729953†L142-L148】:

• Champion characteristics (10 dims) – summarising the champion’s strengths and weaknesses【893838282729953†L142-L148】.
• Role encoding (5 dims) – understanding expected lane behaviour【893838282729953†L142-L148】.
• Performance metrics (8 dims) – kill participation, deaths per 10 minutes, CS efficiency and similar stats【893838282729953†L142-L148】.
• Vision control (2 dims) – number of wards placed and removed【893838282729953†L142-L148】.
• Objective control (1 dim) – track deficits in dragons, barons or tower plates【893838282729953†L142-L148】.
• Risk indicators (2 dims) – high‑variance plays and over‑extends【893838282729953†L142-L148】.

Rather than scolding you, this model surfaces categories of mistakes—positioning, objective timing or recall habits—so you know where to focus next time.

Live game insights: beyond static feature counts

The numbers above describe static inputs used by our models. In parallel, the frontend computes real‑time metrics to adapt recommendations while you play. These features include:

• Timing and phase – seconds, minute buckets and whether you’re in lane, mid or late game【608969380721136†L52-L56】.
• Economy and tempo – gold total, item value and gold efficiency gap【608969380721136†L57-L61】.
• Threat model – each enemy’s fedness z‑score, threat weight and team damage mix vector【608969380721136†L62-L66】.
• Counter triggers – healing, shielding and crowd‑control saturation; thresholds for anti‑heal, armor and magic resist【608969380721136†L67-L74】.
• Inventory progress – core item completion percentage, boots tier and flex slots【608969380721136†L75-L80】.
• Candidate scoring – for each potential item, calculations of survival gain, damage gain, spike timing and regret risk【608969380721136†L86-L93】.
• Anti‑heal analysis and snowball prediction – heuristic modules that spot healing threats and forecast snowballing【893838282729953†L41-L49】【893838282729953†L83-L96】.
• Objective and fight priorities – ranking dragon, baron, herald, tower, gank and farm opportunities【893838282729953†L45-L48】.
• Wave and recall advice – when to freeze, slow push, fast push or recall based on tempo and objective timing【893838282729953†L48-L50】.
• Jungle and vision recommendations – statistical priors on enemy jungle routes, gank timing and optimal ward placement【893838282729953†L50-L54】.
• Player behaviour profile – blending live metrics with historical tendencies to personalise guidance【893838282729953†L54-L56】.

These live insights aren’t fed into the three predictive models; they’re computed locally in your client. They update continuously as the match unfolds, so you see context‑aware recommendations without sending new data to a remote server.

Keeping details proprietary

We’ve verified these feature counts by examining internal documentation and source code, but we intentionally omit certain specifics. For example, we don’t list exactly how each feature is encoded or where model artifacts are stored. Our models are proprietary and remain the result of extensive experimentation. This is both for competitive reasons and to prevent misuse.

What we can share is that all inputs come from public game state and aggregated statistics. We do not collect personal data, chat logs or credentials. The recommendations you see are derived from champion identity, in‑match events and anonymised performance metrics. We also adhere to Riot’s third‑party policy, which prohibits automation and emphasises player agency.

Want to learn more about our architecture? Read How We Built an AI That Thinks While You Play for a high‑level overview.

Why multiple models instead of one?

Some might ask why we don’t simply feed all game data into a massive neural network. Splitting the problem into distinct models keeps things interpretable and robust. Each model focuses on a single decision: drafting, shopping or post‑game review. When a draft recommendation appears, you can trace it back to team composition gaps or damage mix. When an item suggestion pops up, you know it’s because enemy healing is high or your gold efficiency is lagging. If we combined all tasks into one black box, it would be harder to explain why a particular suggestion appeared and to debug when something went wrong.

The takeaway for players

By understanding the general categories of signals we use—champion traits, matchup data, team composition needs, item progress and real‑time threat metrics—you’ll see how the AI thinks about the game. And by knowing that there are separate models with 68, 47 and 28 input dimensions, you can appreciate the complexity without getting lost in proprietary details. The result is a coaching experience that’s both powerful and respectful: we share what matters and keep the rest under the hood.

Want to be first in line? Join the waitlist for early access, release updates and exclusive strategy breakdowns.

Image suggestions and alt text

• Feature breakdown diagram – a visual summarising the draft, item and mistake models with their respective 68/47/28 feature counts. ALT: "Diagram showing the number of input features used in ArcaThread’s draft, item and mistake models".
• Live dashboard screenshot – displaying objective priorities, snowball prediction and anti‑heal alerts. ALT: "ArcaThread dashboard highlighting objective priorities, snowball prediction and anti‑heal alerts".
• Champion embedding visualization – a graphic representing how a champion’s characteristics are encoded into a 20‑dimensional vector. ALT: "Conceptual visualization of a champion embedding used by ArcaThread".

Conclusion

The power of an AI coach comes from its inputs. By combining 20‑dimensional champion embeddings with purpose‑built feature sets for draft, itemisation and mistake detection, and layering real‑time analysis on top, ArcaThread delivers guidance that’s both precise and interpretable. We hope this overview gives you confidence that the system’s recommendations are grounded in verifiable signals—not guesswork, but without revealing proprietary secrets.

Ready to improve your decision‑making? Join the waitlist and be among the first to experience our AI coach in your own matches.