AI Crypto Signals Explained: What the Model Really Does (and Doesn't)
Every signal service now says it uses AI. Here is exactly what the ASI engine's model does, what it is trained on, and what no model can do.
Two different jobs: finding setups and scoring them
The ASI engine separates two tasks that marketing usually blurs together.
Finding setups is rule-based. Every four hours the engine scans 100 markets on the 1H and 4H timeframes for Smart Money structure: breaks of structure, fresh order blocks, liquidity sweeps. The rules are explicit and the same every time; this part is deterministic.
Scoring setups is where the model comes in. Each candidate setup is described by measurable features: the size of the displacement that created the zone, the zone's height relative to volatility, the volume profile, the higher-timeframe trend, how far price currently is from the zone, and more. A statistical model turns those features into a probability that the setup will reach its first target.
How the model learns
The model is trained on the engine's own history: thousands of past setups, each with a measured outcome. It is a calibrated classifier, which means a score of 82 is meant to correspond to roughly 82 percent of similar setups reaching target 1. Calibration is checked against live results, and the grade thresholds were last adjusted when live data showed that setups scoring 84 and above won at a higher rate than those at 82 to 83.
Every Sunday the model is retrained on the newest outcomes. The new version replaces the old one only if it scores better on a held-out set under a strict error measure; otherwise the old model stays. That gate prevents a bad week from quietly degrading the engine.
What the model is not
- It is not a chatbot and does not read news or social media.
- It does not predict prices. It estimates the probability of one specific event: a defined setup reaching a defined target before a defined stop.
- It does not place trades or know your account. It is analysis.
Why a quality filter matters more than the model
The biggest driver of the published win rate is the threshold, not the cleverness of the model. Publishing everything the engine finds would produce many more signals with a lower win rate. ASI publishes only setups scoring 82 or above, which currently means a handful per day. The trade-off is fewer signals; the benefit is a record of 78.7% over 704 closed trades that you can inspect on the results page.
Honest limits
- A calibrated 82 percent still loses roughly one trade in five, and losses cluster. Expect streaks.
- The model knows the past. A regime it has never seen, such as an exchange failure or a flash crash, will not be in its training data.
- Fill rates matter. Many setups never reach their entry zone and expire as no-trades; the engine filters for zones close enough to price that about half get filled.
- The model is retrained on the engine's own signals, so it improves the engine's own style of setup rather than discovering new ones.
How to use an AI score well
Treat the score as a ranking tool, not a guarantee. Two signals at 82 and 85 are both good; the 85 deserves the larger of your two position sizes, not a bigger stop or more leverage. And verify the calibration yourself: over any month of the journal, trades scoring 82 to 83 should win less often than trades scoring 84 and above. If they do not, the score has lost its meaning, and the retraining gate exists to catch exactly that.
Frequently asked questions
Which AI does ASI Signals use?
A calibrated statistical classifier trained on the engine's own measured outcomes, retrained weekly with a quality gate. It is deliberately simple so that its scores can be checked against live results. It is not a large language model.
Can AI predict the crypto market?
No model predicts prices reliably. What a model can do is estimate the probability of a well-defined event, such as a specific setup reaching its first target, and be honest about that probability. Calibration, not prediction, is the useful claim.
Why only a few signals per day?
Because the quality bar is high. Lowering it would produce more signals and a lower win rate. The current threshold was chosen from live data to balance the number of filled trades per day against the win rate.
Educational market analysis, not financial advice. Trading cryptocurrencies involves substantial risk of loss. Past results do not guarantee future results. Figures quoted are from the public ASI Signals journal as of 6 October 2026; a win means the first target was reached.