1. The inputs: candles on four timeframes
The model sees open, high, low and close prices (and volume) on 15-minute, 1-hour, 4-hour and daily candles for all 37 markets at once, plus the calendar (session and weekend structure). Price history comes from public market-data feeds: Dukascopy historical prices and, since May 2026, delayed quotes from Yahoo Finance, with futures-based series rescaled onto the same price level so the history is continuous. Every input is checked to be information that was available at the moment of the forecast — nothing from the future can leak in. Prices can differ slightly from your broker's quotes, and UTC candle boundaries can differ from your broker's chart.
2. The engine: 256 simulated paths
Instead of drawing one line, the model generates 256 plausible 48-hour paths, candle by candle. A candle network reads the multi-timeframe history; a statistical range model scales the spread of the paths to the market's current volatility. All four timeframes are cut from the same paths, so the 15-minute candles add up exactly to their hour, their 4-hour block and their day.
3. The outputs
- Central expected path and close — the middle of the 256 paths at each candle. It is a midpoint, not a target.
- 80% range — at every candle, the band that holds the middle 80% of simulated closes. If the model were perfectly calibrated, about one close in five would still land outside it.
- Direction reading — the share of paths that end above the starting price. With 256 paths, readings within about ±6 points of 50% are statistical noise and are labelled “no clear edge”; an expected move smaller than a tenth of the 80% half-width is labelled “about flat”.
- Reach odds — the chance, counted over the paths, that price touches one typical move up or down within 24 or 48 hours.
What the 80% range means
The 80% range is the main output. At every candle it holds the middle 80% of the 256 simulated closes, so it shows how far price could plausibly travel: narrow when a market is calm, wide when it is not. It is a statement about probabilities, not a promise — if the model were perfectly calibrated, about one close in five would still land outside it, and in practice coverage varies by market and period; each market page shows the coverage actually achieved.
4. Issued every 15 minutes, fixed at issue
A new forecast is issued every 15 minutes while a market is open (Bitcoin around the clock, every day of the week). Once issued, it is never redrawn or “repainted”: the next hour brings a new forecast, not an edited old one. When a market closes for the weekend, the next forecast is issued at the reopen.
5. How it was validated
The model currently published is PRISM-2B: Volatility-aware forecast ranges, updated every 15 minutes and fixed at issue.
Selected on Oct 2025-Mar 2026 and compared on Apr-Sep 2026; historical results do not guarantee future accuracy. The evaluation is chronological and out of sample: a model is compared on months it was not tuned on, against a statistical range model scaled to live volatility, and only out-of-sample results count.
6. The honest limits
- Direction is close to a coin flip. The model is built without a directional lean. Treat direction readings as information, not as signals.
- The 80% range is a target, not a guarantee. Coverage varies by market and period: calm weeks can be over-covered, sudden news can push many closes outside. Each market page shows the coverage actually achieved.
- It does not read the news. The price model knows only prices. Central-bank surprises, data releases, geopolitical events and weekend gaps are not foreseen. (The separate AI briefing reads headlines, but it can be wrong too.)
- Probabilities are not calibrated per market. A “60% up” reading does not mean the market rises 60% of the time.
- Past results do not guarantee future ones. A model that held up over one year can do worse in the next.
7. The AI news briefing
On paid plans each market also carries a short briefing. A large language model, reached through OpenRouter, is given recent public headlines and PRISM's own forecast numbers and writes what the two imply, with levels to watch. It is told never to give buy or sell instructions and never to invent a headline or number, but language models can still misread news. The briefing is scored after 48 hours, and its own hit rate is shown on each card.
Before using any forecast, please read the risk disclaimer.
Frequently asked questions
How often is a PRISM forecast issued?
Every 15 minutes while a market is open, and for Bitcoin around the clock, every day of the week. Once issued, a forecast is never redrawn; when a market closes for the weekend, the next forecast is issued at the reopen.
What does the 80% range mean?
At every candle, it is the band that holds the middle 80% of the 256 simulated closes. If the model were perfectly calibrated, about one close in five would still land outside it; in practice coverage varies by market and period.
How often does PRISM get the direction right?
There is not yet enough tracked data to give a figure. The model is built without a directional lean, so direction readings are close to a coin flip and are information, not signals.
Does the PRISM model read the news?
No. The price model knows only prices, so central-bank surprises, data releases, geopolitical events and weekend gaps are not foreseen. The separate AI briefing on paid plans reads headlines, but it can be wrong too.
Is a PRISM forecast financial advice?
No. It is model output about probabilities, published for information, and past results do not guarantee future ones. Read the risk disclaimer before using any forecast in a trading decision.
Questions? Contact us