Forecast methodology: how we build and score every forecast

The forecasts themselves are in All Coins, coin by coin and year by year.

Methodology last revised: 1 October 2026 · latest model run: 7 October 2026

Models in use, with the number of coins each one forecasts: alt-ratio 1.9.0 (495) · alt-ratio 1.7.0 (1) · bch-ratio 1.1.0 (1) · bnb-ratio 1.1.0 (1) · btc-cycle 4.0.0 (1) · doge-ratio 1.1.0 (1) · eth-ratio 1.3.0 (1) · hbar-ratio 1.1.0 (1) · leo-ratio 1.1.0 (1) · link-ratio 1.1.0 (1) · ltc-ratio 2.0.0 (1) · sol-ratio 1.1.0 (1) · trx-ratio 1.1.0 (1) · uni-ratio 1.1.0 (1) · xlm-ratio 2.0.0 (1) · xrp-ratio 1.1.0 (1)

Where the data comes from

Prices come from daily closing candles on four exchanges — Bitstamp, Gate.io, Binance and OKX — read in that order of priority. For a coin none of them lists, we fall back to CoinGecko’s aggregate price for the last 365 days.

Bitcoin’s history starts on 18 August 2011, so every cycle low and peak since then is in our own database. Today’s unfinished candle is never stored.

A coin is identified by its CoinGecko id, never by its ticker: thousands of tickers are shared by unrelated tokens. CoinGecko also gives us market capitalisation, supply, rank and each exchange’s trust rating.

We also collect the Fear & Greed index, Bitcoin dominance and US market and interest-rate series. No published model reads them today.

Every return and volatility figure we show is computed from our stored history, and every published forecast can be recomputed from our database and the parameters saved with its run.

Three layers: Bitcoin sets the clock

Bitcoin is modelled on its own. The model combines a long-term support line that rises ever more slowly — a power law in time — with a halving-cycle pattern whose swings shrink from one cycle to the next. It is refitted from scratch on every run, and the fitted parameters are stored with the run.

Ethereum has no clock of its own. Its forecast is Bitcoin’s forecast multiplied by the ETH/BTC ratio, which the model expects to drift back toward its long-run average — an average our data pin down only weakly.

A few other large coins — today BNB, XRP, Solana, TRON, Dogecoin, LEO Token, Chainlink — hold today’s ratio to Bitcoin constant. Each was admitted only after a test on past starting points showed that holding the ratio erred less than the shared altcoin drift; coins that failed it in every form we tried, such as Zcash and Cardano, stay on the drift.

Others — today Stellar, Bitcoin Cash, Litecoin, Hedera, Uniswap — carry today’s ratio to Bitcoin forward with a declared downward drift. The coins that were largest at the 2017 and 2021 market peaks went on to lose about 35% a year against Bitcoin on average; the forecast applies that pace for up to about 7 years — the longest span it was measured over — and then holds the ratio. The drift is not fitted to any single coin, and each of these coins was admitted only after a test on past starting points showed that this form erred less than the shared altcoin drift.

Every other coin is Bitcoin’s forecast multiplied by its current ratio to Bitcoin and a downward drift measured across the whole altcoin universe. Nothing is fitted to a single coin: every per-coin parameter we tested did worse.

That drift is why most forecasts sit below today’s price. Between the 2021 and 2025 Bitcoin peaks, only 3.5% of 368 altcoins gained ground on Bitcoin, and the median one lost 95% against it — and our sample only holds coins that are still trading. Why Most Altcoins Lose to Bitcoin, Cycle After Cycle

The band, and what its level means

Each forecast is a central path with a band around it, in 90-day steps up to about ten years out. We do not extrapolate beyond that.

The band is sized so the price should land inside it a stated share of the time: 85% up to three years, easing to 65% at ten years. That level is a target we chose; the width is what our out-of-sample tests say is needed to meet it. Bitcoin gets its width from tests on its own history; a coin with a layer of its own adds to Bitcoin’s band the spread its own ratio to Bitcoin has shown; the other altcoins get it from a test that pools many coins at once, and the confidence check on the larger coins’ model forms pools coins too. Pooled coins tend to miss together, in the same market phase — so a pooled test is worth as many market episodes as it spans, not as many coins.

Only the nearest step, 90 days out, has enough independent past windows in today’s market conditions to test that level: at least eight, for every model. Out to about three years there are only a handful; from three to about seven years, a single market episode; beyond eight years, none. Past the nearest steps the level is therefore declared, not measured — a claim we choose to make, which is why it is lower further out, and why we say so.

An altcoin’s band is never narrower than Bitcoin’s at the same point. The central path is not a price target: read the band first.

The confidence score

Confidence is how much we are willing to claim about our own forecast — not the odds that a coin reaches a price.

It starts from the project’s research score: a project scored 600 out of 1000 cannot carry more than 6.0 out of 10. Every later check can only lower it — how well documented that research is, the band’s level, Bitcoin’s confidence at the same point, a backtest, in today’s market, of bands built the same way and, once forecasts come due, the accuracy tracker. The lowest wins.

Only the nearest horizons rest on a backtest. Further out, the figure is a declared level — the band’s design level, capped as above — which no backtest can confirm yet; where a backtest found the band holding the price less often, we publish that lower, measured figure instead. A dash means we publish no figure there, most often because the coin has no research yet. It is not a zero, and not a cautious default.

The research score, and what it is not

A language model with web search reads public sources about the projects it covers and returns a score from 0 to 1000, strengths and risks each with its source, and any red flags. We give it the figures from our own data; its answer format has no field for a price.

For a project the research has covered, a score below 150 means we treat it as dead or abandoned and publish no forecast. Coins it has not reached yet are forecast without that check, and show no confidence figure.

The score does not predict prices. In a blind test on coins whose outcome we already knew, a project that lost 95% against Bitcoin between the last two peaks and one that gained ground on it received the same score, 710. It is a gate and a cap on our confidence, never a reason to buy. How We Score Trust in a Crypto Project, From 0 to 1000

The research text on coin pages is written by that model from the sources it cites, not by an editor; each page names the model and the date. It is not the exchange score either: the 0–10 trust rating next to each exchange is CoinGecko’s measure of reported volume and order-book depth, and says nothing about solvency or the safety of your funds.

Where language models are used, and where they are not

We use them for the project research above. A separate search can also read two dates from public sources: the next Fed meeting and the estimated date of the next Bitcoin halving. Only the halving date would reach a model, to place Bitcoin in its cycle; without it, the model uses the theoretical four-year cycle.

They also judge which small coins enter the High Potential selection, and draft some of the site’s translations. No price, return or band on StakeBible is written by a language model. Can AI Predict Crypto Prices? An Honest Answer

Crypto prediction accuracy: how we grade our own forecasts

When a forecast’s date arrives, we compare it with the real closing price: the signed error, and whether the price landed inside the band. Forecasts we cannot grade — a coin delisted, a gap in the data — are counted, not silently dropped.

Grades are kept per model version, so a new model never inherits an old one’s record. The tracker can lower a published confidence once at least eight non-overlapping windows are graded at a horizon; it can never raise it. At 90 days, eight windows take about two years for an unchanged model version, and a new version starts its record from zero — until then, the measurement that can lower a figure is the backtest described above.

The latest model run was on 7 October 2026; its first quarterly points come due on 5 January 2027. We publish no accuracy figure until graded forecasts exist.

What we do not forecast

Stablecoins, fiat currencies and tokenized real-world assets — gold, stocks, treasuries — are tracked for price but never forecast: a cycle model has nothing true to say about a dollar peg or an ounce of gold.

We also publish no forecast for projects our research finds dead, tokens that simply track Bitcoin or Ether, broken price series, or coins with too little history.

Why long-term only

What the model reads is a cycle, and a cycle says nothing useful about tomorrow. A short-term forecast would be a coin flip dressed up with decimals. Why We Forecast Crypto in Years, Not Weeks

This is not financial advice

Forecasts are statistical estimates, not a recommendation to buy, sell or hold. We place no orders and know nothing about your situation. Crypto assets are volatile — never invest more than you can afford to lose.