How to use
What the system does with a bar, what it refuses, and how to read a signal without over-trusting it.
How the AI thinks
Collect
Closed one-minute bars are pulled from the exchange's public feed. A bar still forming is never stored — its high, low and close are all still moving, and training on one means learning a candle that never existed.
Describe
Each bar becomes ~40 scale-free features: ratios, z-scores and bounded positions. Nothing is in price units, so the same feature means the same thing on BTC at $63,000 as on ADA at $0.18.
Predict
The model returns three probabilities — higher, lower, and no-edge. The third class is the important one: it lets the model decline, instead of being forced to pick a side on every bar.
Price the bet
A binary paying 85% must win 54.05% of the time to break even. Confidence below that line loses money on average no matter how convincing it looks, so it is refused before anyone sees it.
Commit
A surviving call is written to an append-only chain, each row hashed against the one before it. Nothing can be edited or deleted afterwards, so a losing call cannot quietly disappear.
Settle
At expiry the outcome is read from the stored bar at that exact timestamp. If that bar is missing the signal is voided rather than guessed, and the result — win or loss — goes onto the record.
The decision
Refusals are recorded as carefully as signals. Without them there is no way to tell a model that never fires from one that fires and is always wrong — and no way to ask later what a different threshold would have done.
Reading a signal
Confidence is not a promise
The percentage is the model's own estimate that its call is right. It is only as good as its calibration, which is exactly what the live record tests. A model can be confident and wrong, and the only way to find out is to let the calls settle.
The bar moves with the payout
At an 85% payout you need 54.05% to break even; at 70% you need 58.82%. A signal that clears at one payout can be a losing trade at another, so the required rate is shown on every card.
Empty is a decision
Most of the time there is nothing on the signals page. That is the system working. A page that always has a trade on it is a page optimising for engagement rather than for expectancy.
Never average down
Doubling after a loss — martingale — converts a small edge-free strategy into a fast one. Volta will never emit a recovery ladder, and a sequence of losses is information about the model, not an invitation to raise the stake.
Where this stands today
655 won, 568 lost — 53.6% against a 54.05% bar.
The model serving now has not demonstrated an edge on live outcomes. Treat every signal as an experiment being run in public, not as advice. Trading carries a real risk of loss, and nothing here is financial advice.