Methodology

AI Price Forecasts

Forecasts come from an XGBoost + LSTM ensemble, walk-forward validated with a 2-day embargo period to prevent look-ahead bias. Confidence scores are Isotonic-calibrated against resolved predictions, and every forecast — win or loss — is logged publicly on the track record page.

Only coins with a trained model use the full ensemble. Coins without one yet fall back to a momentum-based heuristic, which is labeled as an estimate rather than an AI forecast wherever it's shown.

A market regime detector (trending / sideways / high-volatility) automatically dampens displayed confidence when direction models are historically less reliable, rather than showing a misleadingly high number.

Risk & Quality Score

The risk screener answers a different question than a forecast: not "which direction will the price move," but "how risky is this coin today," using factual, checkable metrics rather than a prediction about the future.

Every score is computed automatically from published market data, with no human judgment, editorial opinion, or manual override. Every coin — regardless of name recognition or team reputation — is scored by the identical formula below.

The overall score (0–100, Low/Medium/High/Critical) is a weighted average of up to six factors, each scored 0–1 where higher means more risk:

• Volatility (20%) — ATR, Bollinger Band width, and short-term price swings
• Drawdown history (20%) — the largest peak-to-trough decline over the trailing 90 days
• Liquidity (20%) — market capitalization and trading-volume turnover; thinner markets are harder to exit without moving the price
• Exchange risk (15%) — not yet live; requires a per-exchange data integration still in progress
• Volume authenticity (15%) — not yet live; wash-trading detection is planned but unbuilt
• Holder concentration (10%) — not yet live; requires on-chain data and only applies to token-standard coins, not native coins like BTC

Exchange risk, volume authenticity, and holder concentration are not yet available and are not part of the current score. They are excluded, not defaulted to zero — see the reweighting note below.

When a factor isn't available for a coin, the remaining factors are reweighted proportionally rather than treating the missing factor as zero risk — a coin with only 3 of 6 factors computed still gets an honest score from what is known, and the screener discloses exactly how many factors fed each score.

Correlation to BTC and ETH is shown for context but deliberately excluded from the score itself: moving independently of the majors isn't inherently more or less risky on its own.

The overall 0–100 score sorts into four bands: Low (0–24), Medium (25–49), High (50–74), Critical (75–100). Scores refresh automatically every 6 hours from market data pulled from CoinGecko and Kraken.

A Low Risk score is not a recommendation, an endorsement, or a safety guarantee — a Low Risk coin can still lose value. A High or Critical score is not an allegation of fraud, mismanagement, or wrongdoing — it reflects volatility, drawdown, and liquidity metrics only, computed the same way for every coin.

This is v1 of the methodology and will be revised as more factors ship. The score behaves as designed — coins scored higher risk have, on average, drawn down more historically — see the public validation page for that check and its limitations, updated weekly. This is a check on the score's internal consistency, not a claim of predictive edge.

If you represent a project and believe its score is wrong, the issue is almost always in the underlying market data (price, volume, market cap), not in a human judgment call — there is none. Contact [email protected] with the coin and the specific input you believe is inaccurate, and we'll check it against the source data.