Full Transparency

How we measure, score, and rank every AI model.

Everything in this document describes exactly what runs in production — no simplified marketing version. If you can read code, the logic here maps directly to what we've deployed.

The Core Loop

How a signal is generated and scored

1

Data ingestion

OHLCV (Open, High, Low, Close, Volume) candle data for each supported asset is pulled from market data APIs and stored on a rolling basis. The backend retains approximately one year of daily candles per asset.

2

Signal generation

All 20+ models are run against the most recent candle data. Each model independently outputs a directional signal: BUY, HOLD, or SELL, along with a confidence score between 0 and 1. Web sources (80+ external analysts and publications) are also tracked and contribute directional signals.

3

Accuracy logging & outcome resolution

Each prediction is timestamped and stored at generation time. After the 7-day prediction horizon, the prediction is resolved against the actual market movement. An adaptive threshold per asset (based on that coin's historical volatility) determines whether the move was significant enough to count as a directional call.

4

Accuracy-weighted ranking

Models that beat random chance (50% directional accuracy) earn proportionally higher influence over the ensemble signal. The weighting is non-linear — small accuracy improvements translate to significantly higher influence. Models at or below random chance are not simply ignored; they are inverted (see Contrarian Intelligence below).

5

Ensemble voting & conviction filtering

The weighted ensemble calculates the directional consensus across all models and web sources. Several protective filters are applied before the final signal is published — including confidence gating, conviction thresholds, anti-herding skepticism, and market-regime awareness. When conviction is insufficient, the system abstains rather than guessing.

Weighting System

How models earn influence

Not all models are equal. Models must prove their edge through real, verifiable predictions before they're granted influence over the ensemble signal.

Why don't we publish the exact formulas? Our scoring methodology is our competitive advantage. We've described the principles transparently so you can understand how the system behaves — which models get promoted, which get demoted, and why. The exact parameters are proprietary and continuously tuned.
Ensemble Intelligence

Six layers of signal refinement

The raw weighted vote passes through multiple protective filters before becoming a published signal. Each layer is designed to improve quality and reduce false signals.

1

Confidence gating

Each model reports a confidence level alongside its signal. High-confidence signals receive full weight, moderate signals receive reduced weight, and low-confidence signals are nearly muted. This ensures the ensemble is driven by models that are sure of their call — not models hedging.

2

HOLD exclusion

When a model outputs HOLD (meaning it has no directional conviction), that vote is excluded entirely from the ensemble. This prevents uncertain models from diluting the consensus of models that do have a clear view.

3

Web sources blending

Open positions from 80+ tracked external analysts, influencers, and publications are blended into the ensemble alongside AI model signals. Each source is weighted by its own tracked accuracy — high-accuracy sources carry more influence than low-accuracy ones. This adds a real-world sentiment layer to the algorithmic signals.

4

Anti-herding skepticism

When an overwhelming majority of models agree on a direction (85%+ consensus), the system applies a skepticism penalty to the ensemble's confidence. In crypto markets, extreme consensus often precedes reversals. This dampens overconfidence without overriding the signal direction.

5

Conviction threshold & abstention

When the weighted consensus is below a minimum conviction threshold (neither BUY nor SELL has a clear majority), the system outputs ABSTAIN instead of forcing a call. This is a deliberate design choice — we'd rather skip a signal than publish a coin-flip guess. The directional lean is still visible for reference, but it's clearly marked as low-conviction.

6

Market-regime filter

A macro-level protective filter monitors Bitcoin's trend as a proxy for the broader crypto market. During confirmed bear markets (BTC below its 50-day average with sustained losses), the system suppresses low-confidence BUY signals. This prevents the ensemble from catching falling knives during systematic downtrends.

The Models

21 independent prediction strategies

Each model is fully independent — they share the same input data but use entirely different approaches to reach their predictions. No model influences another's output.

Machine Learning & AI

ModelApproach
XGBoostGradient-boosted decision trees trained on engineered price and volume features. Well-suited for tabular OHLCV data with non-linear relationships.
Gradient BoostingEnsemble of weak learners (sklearn implementation) that iteratively corrects prediction errors. Captures complex feature interactions across time windows.
Random ForestEnsemble of randomized decision trees. Robust to overfitting via bagging; provides implicit feature importance ranking across technical indicators.
Neural NetworkFeedforward multi-layer perceptron trained on normalized OHLCV series. Learns non-linear temporal patterns without hand-crafted feature rules.
Support Vector MachineSVM classifier with RBF kernel. Finds the maximum-margin hyperplane separating directional outcomes in feature space.

Technical Analysis

ModelApproach
Ichimoku CloudMulti-component Japanese charting system using Tenkan-sen, Kijun-sen, Senkou Span A/B, and Chikou Span for trend, momentum, and support/resistance signals.
SuperTrendATR-based trend-following indicator. Generates a clear directional line that switches side when price closes through the trailing stop band.
Parabolic SARStop-and-Reverse indicator tracking price acceleration. Produces tight signals in trending markets; checked against trend confirmation to reduce whipsaws.
EMA DivergenceMeasures spread between fast and slow exponential moving averages. Divergence direction and magnitude inform momentum and potential reversal signals.
TRIX SignalTriple exponential smoothing of price. The rate of change of a triple EMA reduces noise and generates momentum signals with minimal lag.
Williams %RMomentum oscillator measuring closing price relative to the high-low range. Identifies overbought (> -20) and oversold (< -80) conditions.
MomentumRaw rate-of-change of price over a set period. Directional momentum is combined with zero-line crossovers to classify trend strength.
Price ActionCandlestick pattern recognition — engulfing candles, pin bars, inside bars, and doji formations — to identify institutional pressure and reversals.
Technical Indicator CompositeA weighted composite of RSI, MACD, Bollinger Bands, and Stochastic oscillator. Signals are only generated when multiple indicators align.

Statistical & Market Microstructure

ModelApproach
VWAP PremiumMeasures price deviation from Volume-Weighted Average Price. Extended premium above VWAP signals overextension; discount signals potential accumulation.
Volume ProfileAnalyzes trading volume distribution across price levels. High-volume nodes act as support/resistance; breakouts through them carry higher conviction.
Volatility BreakoutIdentifies expansion in price range (ATR-based) following periods of compression. Breakouts after low-volatility consolidation are classified as directional signals.
Mean ReversionStatistical model that identifies assets trading at significant Z-score deviations from their rolling mean. Expects reversion toward the mean over a lookback window.
Fear & Greed IndexIncorporates market-wide sentiment data — volatility, momentum, social signals, dominance — to classify broad market psychology and its likely directional impact.
Elder-Ray IndexCombines exponential moving average with bull/bear power measurements. Positive bull power above EMA confirms uptrend signal; negative bear power below EMA confirms downtrend.
Web Sources

Tracking 80+ external signal sources

In addition to our 20+ AI models, ArtinFox monitors a curated set of 80+ external signal sources — crypto analysts, influencers, financial media outlets, and market commentators — across multiple platforms including X (Twitter), YouTube, RSS feeds, and newsletters.

Each external call is logged with a timestamp and directional prediction. When the market resolves, the call is scored using the same directional accuracy methodology as our AI models. Sources are ranked publicly on the News page and contribute to the ensemble signal via accuracy-weighted blending.

Editorial note: ArtinFox links to external article titles and URLs only. We do not reproduce, cache, or republish third-party content. All articles open on the original publisher's website.
Published Metrics

Three named ArtinFox metrics

From the underlying dataset we publish three named, citable metrics. Each has a stable URL and a public methodology so they can be referenced in research, journalism, and downstream tools.

ArtinFox Source Accuracy Index

A confidence-weighted aggregate of the directional accuracy of the top-ranked external crypto analysts and research sources tracked by ArtinFox, computed daily across a rolling 90-day window for the top 50 cryptocurrencies. Higher values mean the consensus of the top-ranked voices is currently being validated by price action; lower values mean the loud crowd is currently wrong.

ArtinFox Disagreement Index

A second-order metric measuring the spread of directional opinion across the top-ranked sources for a given asset. When the top sources are unanimous, the Disagreement Index is low; when they split 5-5, it is at its peak. Empirically observed to be a leading indicator of elevated 48-hour realized volatility — markets break harder when the experts can't agree.

ArtinFox Calibration Score

The mean absolute deviation between a model's stated confidence and its observed directional accuracy at that confidence level. A perfectly calibrated model has a Calibration Score of 0. Most retail "AI signal" tools score poorly here — they say 90% confident and are right 55% of the time. We publish ours, even when it isn't flattering.

Citation note: Each metric has a stable canonical URL on this page. Researchers, journalists, and AI assistants are welcome to reference them by name with a link back. The numeric values themselves update daily; permanent URLs do not.
Data Pipeline

How fresh is the data?

Open Data

Download the prediction dataset

Researchers, journalists, and traders can download the last 365 days of graded AI predictions as a single CSV. Includes symbol, model, signal, confidence, entry/TP/SL, resolution price, and outcome. Pending forecasts are excluded so there is no look-ahead leakage.

Download CSV (last 365 days)

License: free to use with attribution to artinfox.com. Updated every 30 minutes.

Important Disclaimer

All signals, accuracy figures, and model rankings provided by ArtinFox are for educational and informational purposes only. They do not constitute financial, investment, or trading advice. Past model accuracy is not indicative of future performance. Cryptocurrency markets are highly volatile. Always conduct your own research and consult a qualified financial advisor before making any investment decision. See our full Disclaimer.