The foundation model
AlphaMind Prediction Engine
A decoder-only Transformer trained on price itself. It reads a chart bar by bar and returns a distribution of where price can go next rather than a single number.
How it works
One pass through the engine
The figure follows a single request at the timeframe you are charting. Recent history goes in as one sequence, the model writes the continuation a step at a time, and the sampled paths collapse into a band.
Sequence ingestion
Discrete tokenization
Each candle and price movement is compressed into tokens the AI model can read.
Decoder stack
A transformer engine reads them like a sentence.
Probabilistic paths
Future paths are sampled and organized into probability quantiles.
p(x̂t+1:t+H | xt−499:t) = ∏i p(zi | z<i; θ)
The algorithm that turns candles into probable future paths.
500
Bars of history read as one sequence
20
Bars written forward, a step at a time
q.05 / q.95
The band the sampled paths collapse into
~10s
From request to a complete written read
The output
What the engine hands back
A forecast is only worth having if it can be acted on and disproved. Four things come back, and the fourth is the one most systems leave out.
A direction, with a confidence
Not a verdict. A leaning, carrying a number for how strongly the sampled paths agree with one another.
A band, not a line
The paths collapse into quantiles. The distance between the outer bands is the honest width of the forecast, and it widens when the market gives the model less to stand on.
The levels the path implies
Where the median path meets the structure already on the chart. Those crossings become the target and stop candidates the rest of the platform works from.
The price that ends the idea
Every forecast carries the level at which it is finished. A read you cannot invalidate is not a read.
What it does not read
Price, and nothing else. No headlines, no positioning, no sentiment. Context of that kind enters one layer later, where the forecast is turned into words, and the separation is deliberate: it keeps what the chart says apart from what is being said about the chart.
In the product
Where the engine runs
Three surfaces sit on the same model. What separates them is what each one does with the distribution.
The most literal use of the engine. You are on a chart, you ask for a read, and the bars in front of you become the input sequence. The forecast comes back, the platform measures it against the structure already on that chart, the average range, the recent swing highs and lows, the round numbers within reach, and the analysis layer writes the result as something you can act on. Nothing in that write-up is improvised: every level it names was computed before the text was written.
Signals are published, timed and tracked, so the bar is higher than it is for a read. A trend read can be early and still be useful; a signal cannot. The engine is applied at the trigger, where its forecast is weighed against the conditions the rules describe, and what reaches you has already been checked against what the model expects the next bars to do. Each signal carries its entry, its invalidation and a history that stays published whether it worked or not.
MindX GPT was trained on the engine's output, which is what separates it from a general model handed a chart. It knows the shape of the forecasts it is discussing. With a read on your screen you can ask what the band means, why the median path turns where it does, or what would have to happen for a level to fail, and the answer comes from the evidence the read was built on rather than from a paraphrase of it.
