← Back to K3vl4r

How K3vl4r forecasts

We'd rather show you how the forecasts are made than make promises about them. Last reviewed June 2026.

The model

Every forecast on K3vl4r is generated by Kronos, an open foundation model for financial time series. It's an autoregressive transformer trained on historical OHLC (open/high/low/close) candles across many assets — it reads a window of recent price action and generates a probable continuation of it, one candle at a time. We run it continuously on our own GPU; forecasts refresh on a fixed schedule, not on demand per visitor.

It is a general market model, not a bot tuned to any single coin or a strategy fit to past winners. It has no access to news, order books, on-chain flows, or insider information — only price and volume history.

What we forecast

We forecast a live watchlist of cryptocurrencies and equities across multiple horizons. The headline per-symbol pages show four: 1-hour, 4-hour, daily, and weekly. Shorter, intraday horizons are inherently noisier than longer ones — we say so on the forecast itself rather than hiding it.

Each horizon uses sampling settings appropriate to its length, so a multi-step weekly path doesn't behave like a single-step intraday one.

How a single forecast is produced

  1. We pull the most recent candles for the symbol and timeframe.
  2. The model generates a forward path of candles from that history.
  3. For daily and weekly horizons we run the model many times (a Monte Carlo pass) to produce an upper/lower confidence band and a bullish probability — a distribution of outcomes, not a single line.
  4. The result is written to our database and rendered on the chart, with the time it was produced.

What the confidence bands mean

The shaded band around a forecast is a range of modelled outcomes, not a promise that price stays inside it. A wider band means the model sees more uncertainty for that asset and horizon.

We are candid about calibration: getting a band to hold its stated confidence (e.g. price landing inside an "80%" band 80% of the time) is hard, and we are actively measuring and improving ours. Until that measurement is solid, treat the bands as a relative sense of uncertainty, not a precise probability.

What we do not claim

This is the important part, and most of the industry buries it:

  • We do not claim a proven accuracy rate or track record. We are early, we are collecting resolved outcomes over time, and we will not publish a performance number until it clears a bar we set in advance (see below).
  • Forecasts are scenarios, not signals. They are one input for your own research — not a recommendation to buy or sell, and not financial advice.
  • Short-term price direction is genuinely hard.For many assets and horizons, no model reliably beats simple baselines, and we'd rather tell you that than pretend otherwise.
  • Past patterns do not guarantee future prices. Markets change regime; models fail in ways their history didn't show.

See our full Risk Disclosure for the complete picture.

How we're validating it — before we claim anything

Rather than eyeball results and cherry-pick good calls, we run a standing, read-only evaluation on a fixed held-out set of symbols. Every week it records honest metrics — did the model beat a naive "assume-the-trend-continues" baseline, and did prices actually land inside the confidence bands — and appends them to a growing log we never edit after the fact.

We pre-registered the thresholdsthat would let us act: we'll only recalibrate the live bands, promote a new model, or publish a public track record once the measurement clears those bars over a sustained window. That discipline is what keeps us from marketing on noise.

See it yourself

Every forecast is public and free to view. Browse today's top picks, the market heatmap, or read more about K3vl4r and our FAQ.