Core engines

The Five Prediction Engines

Every Arvio prediction comes from one of five engines. Each engine has a different lens on the market, but all of them share the same discipline: they log every call, score it against reality, and feed the result back into calibration. The engines below are listed in the order they mature — Radar and Early Signals first, Token Analysis newest.

Radar Conviction

Discovery

What it looks for

Under-the-radar, high-quality tokens — real projects (not memes or degen plays) ranked roughly 50–400 by market cap with strong fundamentals, active development, or emerging narratives that most people aren't paying attention to yet.

How conviction is formed

An LLM with live web context scores each candidate 0–100 on how well-supported the thesis is, producing a directional conviction (bullish), a risk tier (Low / Medium / High), and a set of signed evidence reasons. Curated ~hourly; the active set is replaced each run.

Signal fingerprint

Fundamentals + narrative fitRisk tier per pickSigned evidence reasonsHourly refresh, deduped

See it live: Opportunities View measured track record

Early Signals

Momentum

What it looks for

Emerging on-chain and social momentum before it becomes obvious — tokens surfacing through volume acceleration, liquidity buildup, social attention, or breakout setups.

How conviction is formed

Classifies each candidate into one of five signal types, then scores conviction 0–100. Each signal carries a calibrated probability and a status (active → faded or broken out) so you can see whether the setup is still live.

Signal fingerprint

Volume accelerationLiquidity buildupSocial surfacingNarrative emergingBreakout setup

See it live: Watchlist View measured track record

Catalysts

Event-driven

What it looks for

Event-driven directional setups — upcoming or active catalysts (launches, unlocks, partnerships, upgrades) that imply a probable price move in a specific direction.

How conviction is formed

Each cached catalyst carries a headline, timing, direction (bullish / bearish), and a confidence score. Predictions are logged per token so the catalyst engine enters the resolve → calibrate loop like every other engine.

Signal fingerprint

Directional (bull/bear)Timing-awarePer-token cachedConfidence-scored

See it live: Catalyst radar View measured track record

Narratives

Thematic

What it looks for

Theme-level conviction and rotation — which narratives are heating up or cooling down, measured across capital flow, developer activity, social momentum, token performance, and liquidity quality.

How conviction is formed

Each narrative carries a composite score (0–100), momentum, and direction (up / down / flat). A representative lead token anchors each narrative so predictions resolve against a real, tradeable asset — never a vague theme.

Signal fingerprint

Composite momentum scoreCapital flow factorDirection (up/down/flat)Lead-token anchored

See it live: Narratives View measured track record

Token Analysis

Per-token

What it looks for

On-demand deep analysis of any token — a multi-section report covering fundamentals, momentum, liquidity, social, catalysts, risk, and a bull/bear case, synthesized into an overall conviction.

How conviction is formed

The deep analysis produces an overall conviction (0–100), a risk level, and section-level scores. The newest engine: its calibrated probability (powered by the fitted token_analysis model) surfaces directly on the Token Report, alongside the raw conviction.

Signal fingerprint

Multi-section reportOverall + risk levelBull/bear caseCalibrated probability on report

See it live: Any token reportView measured track record

All five engines feed the same loop: every conviction is logged as an open prediction, resolved against the real price move at its horizon, and the outcomes fit the calibration models that turn raw conviction into realized probability. Two more engines run the same loop in parallel — the Wallet Health engine (its predictions resolve against the frozen held basket, covered in its own section) and the token-authenticity audit engine (verdicts resolved against token survival). That layer is covered next.

Continue to Calibrated Confidence

Edited on July 18, 2026