Methodology

Every number on DUKOIN has a source, a timestamp and a method you can read. This page is the map: how raw market data becomes scores, forecasts and AI answers, and the checks in between.

Our principles

• Show the data age. Every value carries a freshness state, so you never mistake an old number for a live one.
• Rank, do not promise. Scores express research priority. Forecasts express uncertainty. Neither is a promise of profit.
• Publish the record. Forecasts are logged before outcomes are known and scored in public, including the misses.
• Say what is missing. When data is not connected, we show it as missing rather than estimating it quietly.
• Keep AI on a short leash. The assistant explains verified data. It does not invent numbers, generate probabilities or move money.

1. Data

Market data (prices, market caps, volume, supply, percentage changes and daily history) comes from the CoinGecko API and is cached on our servers. News headlines come from publisher RSS feeds; we store the headline, link and timestamp, not article text. The event calendar is editorial and manually verified, with a source link for every event. Derivatives (Bybit), DeFi (DefiLlama) and Bitcoin network data (mempool.space) are connected; per-asset on-chain flows are not yet.

Every value shows one of five freshness states:

Full details: Data sources.

2. Research Score

The Research Score v1 ranks roughly 250 large assets from 0 to 100 on eight weighted components: Trend, Momentum, Liquidity, Fundamentals/Tokenomics, On-chain, Derivatives, Catalysts and Risk resilience. Each raw metric is converted to a percentile across the universe. Components without data are left out and the remaining weights are re-normalized, and every score shows its coverage and a data-quality grade. Assets with very low volume, missing market cap or stale data are excluded. Full formula: Research Score methodology.

3. Forecasts

Our forecasting model, DK-EWMA-LN v1, is a volatility model. It estimates how widely price is likely to move over 24 hours, 7 days and 30 days, and outputs a median, 50/80/90% intervals and threshold probabilities. It does not claim to know direction. A directional probability is shown only for models that pass a walk-forward test against a 50/50 baseline; otherwise you see “No demonstrated directional edge”. Full method: Forecast methodology.

4. The ledger and evaluation

Every forecast is written to an append-only ledger at the moment it is created. Entries are never edited or deleted. When a forecast’s horizon matures, it is evaluated automatically against the realized price. Results, including hit rate, Brier score, log loss, calibration error and interval coverage, are published on Model performance. Until at least 30 forecasts have matured for a horizon, that page shows “Insufficient evidence for a reliable calibrated forecast”.

5. AI assistant and research agents

DUKOIN’s assistant and research agents are powered by Claude, Anthropic’s AI model. They work through tools that read DUKOIN’s own verified data: market data, scores, dip flags, forecasts, and, for signed-in users, only that user’s own watchlists and portfolios. Answers include sources and timestamps.

Claude explains forecasts; it never generates probabilities or prices. Content fetched from third parties, such as headlines, is treated as untrusted and kept behind prompt-injection boundaries, so it cannot instruct the AI. The AI has no ability to trade, place orders or move funds.

6. Quality assurance

Before an AI answer is shown, a validation step checks it against the data the AI was given. Figures in an answer are checked against that data: price-target-style numbers that are not in it are blocked, and any other unmatched figure is flagged as unverified. The morning brief is rejected outright if any figure fails the check. This is designed to stop the most damaging failure of language models in finance: confident, specific numbers that are simply made up.