Vulcan Markets
Published 14 September 2026
How a surface is built

What does a model that draws the next candles actually know?

Where you are. A model drew the next few candles on your chart and they look confident. You want to know what it actually had to go on before you trust the shape.

A candlestick forecast is not a tip. It is not a percentage attached to a direction. It is a model reading a sequence of bars and drawing what it expects the next bars to look like - projected open, high, low and close, at three explicit horizons. What the model knows, and what it cannot know, is exactly what this piece is about.

What Goes In

The input is not a price. It is a sequence.

The Candlestick Forecast on Vulcan Trading reads the last 300 bars before it draws anything - open, high, low, close and volume, in order, bar by bar. That sequence is the signal. The model is looking at structure: how each bar relates to the one before it, how volume moves with price, how ranges expand and contract across the window.

This is the first distinction that separates it from an indicator. An indicator takes a price or a set of prices and applies a formula. A sequence model reads the shape of a market over time. The 300-bar context window is not a parameter chosen for convenience - it is the span the architecture needs to read the rhythm of a market before it projects forward.


What Comes Out

The output is a full candle, not a label.

At each of the three horizons - 12 bars, 24 bars and 48 bars - the model draws a projected candle body: open, high, low and close. The shape of that candle carries information that a direction label alone does not. A projected candle with a wide range and a close near the high tells you something different from a narrow-range candle closing in the middle, even if both point nominally upward. The geometry is the forecast.

Three horizons mean three explicit forward windows, not one ambiguous prediction. You can see whether the model's projected structure changes between the near term and the extended term. That difference is itself information. You set the instrument - from 27 available, spanning FX pairs, equity indices, US large-cap equities, gold and crypto - and you read the candles at each horizon before you read anything else.


Where the Model Comes From

The architecture is a decoder-only Transformer with a BSQ Tokenizer, purpose-built for OHLCV sequence modelling. It was not adapted from a language model or a general-purpose classifier. The design is specific: the model's job is to take a bar sequence and output the next bars in the same format. That specificity matters because it means the architecture was chosen for the problem, not fitted to it afterward.

The training corpus spans data from 45+ global exchanges. The model was not trained on one market or one asset class. The breadth of the corpus is what allows it to operate across FX, equities, gold and crypto from the same surface.

This model was published at AAAI 2026 - a peer-reviewed venue with a public record. You can read the methodology. Most retail tools that describe themselves as AI-powered do not offer this. The publication is not a performance claim; it is a transparency fact. The methodology is checkable, which is a different thing from the results being guaranteed.

In-product, the surface describes itself by architecture type and context window - not by an accuracy score. That framing is deliberate. It is also the honest framing.


What It Is Not

A forecast drawn against actuals is not a win rate.

The surface shows what the model projected. It does not attach a percentage to those projections. It does not show a P&L figure or imply a return. If you look at a forecast candle and then look at the bar that followed, you are seeing two data points - not a performance record. A methodology display is not a claim about what the model earns.

No accuracy figure, no win rate, no implied return appears on the surface or in this piece. The model explains the structure it reads. What you do with that explanation is your decision.


Where It Sits

The Candlestick Forecast is one surface of fifteen on Vulcan Trading. The model behind it runs that surface alone. Fourteen other engines sit behind the rest of the platform - order-book intelligence, tail-risk regime detection, Treasury curve decomposition, prediction-market flow, sentiment, and more. Breadth is the headline. The Candlestick Forecast is the entry point to a larger analytical stack, not the whole of it.

The surface is open on the Vulcan Trading free tier. Open it, set an instrument, and read the method note beside the forecast candles before you read the candles themselves. The method note is there for a reason.

Start free at vulcan-trading.ai

#TradingIntelligence #CandlestickForecast #VulcanTrading

The Candlestick Forecast surface in Vulcan Trading: a chart with history and forecast candles, a prediction output row naming the horizon, bars used, instrument and model, and a panel describing the model. Direction, forecast change and the price axis are blanked.
Where this lands in Vulcan Trading. A candlestick forecast as the surface presents it: a window of bars in, forecast candles out, and the method note beside them. What the forecast said is blanked.
  1. The chart: the history in one colour, the forecast candles in the other, with the boundary marked. The price axis is blanked.
  2. The prediction output row: horizon, bars used, instrument, the model's name and the inference time. Direction and forecast change are blanked.
  3. About the model, in the surface's words: the type, the context window, the input, the output, and that inference runs on demand.
Blanked in this still: the direction and forecast-change cells; the price axis. The method cells are shown; no result figure is.

The objection

You might say
If it can draw the next candles, why can it not tell me the accuracy?
The answer
A forecast drawn on live bars and an accuracy figure are two different claims, and this page makes only the first: a window of recent bars in, forecast candles out across three horizons, with the method note beside them saying what the model is and what it was trained on. No accuracy figure appears here.

How Vulcan computes this

Surface
Candlestick Forecast
What it does
Full OHLC forecast candles at 12, 24 and 48 bars ahead, across 27 selectable instruments.
Instruments
27 selectable instruments
Method
Forecasts are framed impersonally, as what the model projects. The surface states what it is projecting and over what horizon, and it never tells you what to do about it.

Described as capability. This page reproduces no figure from the product and nothing on it is a live read; the method is the point, not a result.

First action

Open Forecast in the Retail view, pick an instrument you follow, and read the method note under the chart before the candles.

See it on the surface, free. One free account opens the Retail view: the desk, the watchlist, Stocks with the pairs lab inside it, Gold Vault, Order Flow, Forecast and Replay. No card. You are leaving a page about method for a product that applies it. The page stays here.

Create a free account

Next in this path

Door You read order flow and want the book, not the last print · page 4 of 4

That is the end of this path. Pick another door, or read on below.