Crypto prices and prediction methodology
Where our prices come from, how forecasts are built, and how to read their limits.
- Market pricesTimestamped provider data
- 30D · 90D · 1Y forecastsModelled outcomes at a target date
- Annual scenarios through 2030Separate milestone-adjusted projections
Prices and sources
Coin profiles report stored provider quotes, market capitalization, trading volume and supply. Pricing data is provided by CoinMarketCap; embedded charts are provided by TradingView. Other profile information is collected and maintained by CryptoSlate.
Each quote has its own timestamp. Missing data is withheld. An older quote is not a claim of current trading or an executable exchange price, and a provider outage alone does not establish that an asset is inactive.
How market signals differ from forecasts
Market Signal describes observed price conditions. Trading Activity describes market participation and available liquidity evidence. Peg Health describes observed price stability around a reference peg; it does not assess reserves or redemption safety. These indicators describe observed conditions and are separate from future-price forecasts.
Reading a forecast
A forecast describes possible prices at a fixed future date, measured from a frozen reference close. Published horizons are 30, 90 and 365 days where a valid snapshot exists. The median is the middle of the modelled outcomes, not a guaranteed future price.
P10 to P90 contains the middle 80% of modelled outcomes. The chart can also show 50% and 95% model intervals. These percentages describe the model distribution; they are not guaranteed real-world coverage rates.
An interval at the target date does not predict whether a price will be touched before that date. Percentage changes compare each target price with its snapshot’s reference close, which can differ from the latest market quote.
Annual scenarios & milestones
Illustrative outlooks target exactly 365 days ahead and December 31 closes through 2030. They use a separate simulation with frozen editorial milestone assumptions, rather than extending a backtested one-year forecast. One simulation supplies those targets using up to three years of price history and reduced historical drift.
- Sourced events
- Primary sources establish dates, protocol changes and regulatory milestones. Estimated windows remain estimates; related news provides context rather than proof of a price effect.
- Editorial assumptions
- Editors set separate upside, central and downside effects. These are editable scenario inputs, not source-backed price estimates. A neutral central effect leaves the central baseline unchanged.
Illustrative fitting requires at least 366 consecutive valid daily closes, providing one full year of returns; up to 1096 closes are used. This engineering floor permits fitting but does not establish predictive accuracy. Assets with less history remain unavailable. Short-term backtests do not validate illustrative horizons, and scenario bounds are not calibrated probability intervals.
How milestone effects enter annual scenarios
The annual baseline uses P10, P50 and P90 prices. Eligible dated milestones adjust those baselines with separate scenario effects, phased in from the event or window start to the end of its phase-in period. Each effect is applied once and then held constant; effects combine multiplicatively.
Undated events and events beyond the forecast window remain context only, with no numerical adjustment. Cancelled or completed events do not add a new effect. Overlapping assumptions, or effects already reflected in price, can distort results and require editorial review.
The published snapshot preserves the milestones and numerical assumptions used for that run. Later edits do not rewrite that forecast’s evidence.
Validation & limitations
We test short-term forecast distributions against a simple baseline on held-out historical data. Each prediction page reports its run’s measured improvement and uncertainty, including results that are worse than the baseline.
Past performance does not ensure future accuracy. Regime shifts, exchange failures, thin liquidity and limited samples can make forecasts unreliable. Pending targets do not establish predictive accuracy, and these scenarios are not financial advice or guarantees.
How we test against a baseline
CRPS, the Continuous Ranked Probability Score, evaluates the entire forecast distribution. Lower scores are better. Calibration chooses model weights; separate held-out forecast origins evaluate them after an embargo. Non-overlapping validation outcomes reduce dependence between samples.
The reported comparison uses the strongest declared baseline and a one-sided 95% lower confidence bound. Coverage and hit-rate claims require resolved outcomes, a stated sample size, the interval width and the forecast horizon.
Freshness & publishing
A coin’s prediction page must be enabled and have a published backtested forecast or an eligible illustrative outlook. Private evaluation runs stay private. Paused forecasts withhold targets; older published snapshots carry a freshness notice rather than a claim of a new daily run.
The reference close anchors the calculation, the target date specifies the forecast endpoint, and the snapshot timestamp records the run. These dates serve different purposes. A newly loaded page does not mean the underlying quote or forecast has just been updated.
Technical reference
Each prediction page’s model details identify the model version, inputs and validation evidence behind its published snapshot. The short-term model and annual simulation have separate methods and limitations.
Model details and reproducibility
The short-term model combines Student-t GARCH, filtered historical simulation and regularized quantile regression. Component weights and input provenance are stored in each immutable snapshot, along with the training window, simulation path count and reproducible random seed.
Annual scenarios use a separate horizon-adjusted GARCH simulation and a frozen set of milestone assumptions. Their reproducible seed, baseline prices and applied effects are retained with the published annual outlook.
Why prediction markets are shown separately
External prediction-market probabilities are displayed as context and have zero weight in the published model. Their contracts, settlement rules and observation times may differ from our target-date forecasts. Any future model contribution would require separate held-out validation.