Key takeaways
- The model compares normalized price, range, volume, and multi-candle context rather than raw price levels.
- Confidence communicates relative model strength and target difficulty, not certainty.
- Automated generation, administrative data review, and objective outcome tracking serve different purposes.
The model begins with market data, not a story
Every daily candle summarizes four prices: open, high, low, and close. Volume adds a measure of activity. A sequence of candles contains information about trend, volatility, failed moves, acceptance above or below levels, and changes in participation.
An AI chart model tries to recognize useful relationships inside that sequence. It does not read the chart as an image in the same way a person looks at a screen. The data is transformed into numerical features that preserve price structure while allowing different securities to be compared.
Why raw prices must be normalized
A $2 crypto asset and a $2,000 stock can form visually similar patterns. Comparing their raw prices would be meaningless. The model instead works with relationships such as percentage change, range relative to price, distance from recent highs, and volume relative to a rolling baseline.
Normalization also helps with corporate actions and different quote conventions. The data pipeline must keep candle semantics consistent so a split, missing value, or provider adjustment does not masquerade as a breakout.
Context matters more than one candle
One large green candle does not define a high-quality breakout. The model evaluates what came before it:
- Was price compressed or already extended?
- How many times was resistance tested?
- Were lows rising toward the boundary?
- Was volume expanding or fading?
- Did previous attempts fail quickly?
- Was volatility stable enough to define a meaningful stop?
The same closing price can therefore produce different scores depending on the path that led to it.
Learned patterns and engineered measurements
A robust system can combine two types of input.
Engineered measurements encode well-understood chart concepts such as rolling returns, ranges, moving relationships, volume expansion, and distance from support or resistance.
Learned representations allow the model to combine features in ways that are not expressed by a single traditional indicator. These combinations can capture subtle sequences that are difficult to name or draw manually.
The value of machine learning is not that it creates magic. It is that it can compare many normalized relationships across a large history with consistent rules.
Separate ranking for breakouts and breakdowns
Upside and downside structures are related but not identical. Downside moves can accelerate as liquidity disappears. Upside breakouts can depend more heavily on sustained demand. The model therefore ranks breakout and breakdown candidates in separate groups before publication thresholds are applied.
Each candidate receives a score relative to other setups in its direction. Only the strongest candidates that also pass minimum quality rules proceed.
From score to confidence
The displayed confidence value is a compact way to communicate relative model strength. It may incorporate rank, pattern quality, target distance, and data-quality checks.
It should not be interpreted as a calibrated promise that the target will be reached. Markets are open systems affected by information and behavior that no historical chart model can fully observe.
Confidence is most useful for comparison. A user can understand that one setup ranked more strongly than another under the same model and data process.
Defining target and stop
A signal is incomplete without an objective and an invalidation point.
For a breakout, the target is above entry and the stop below. For a breakdown, the target is below and the stop above. The distances can reflect expected movement, volatility, support or resistance, and the model's maximum evaluation window.
Targets that require larger moves are harder to reach. The platform therefore reduces displayed confidence as target distance grows rather than presenting every ambitious target with the same apparent certainty.
Outcome tracking
Once approved, the signal is frozen as a historical record. Future daily candles determine whether target, stop, or time limit ends the lifecycle.
The platform separates:
- Current profit or loss for active signals.
- Targeted profit required for success.
- Final profit or loss after a lifecycle outcome.
- Days active or days to success or failure.
That distinction prevents temporary favorable movement from being reported as a completed success.
Why the model does not rely on news
News and narratives can matter to markets, but they introduce challenges. The same headline can produce opposite reactions in different regimes. Text arrives from many sources with different latency, credibility, and licensing constraints.
A chart-only model makes a narrower claim. It asks whether price and volume behavior contain a setup that resembles historically useful breakout or breakdown structures. It does not claim to know why market participants are acting.
This focus makes the process repeatable but also limits it. A sudden fundamental event can invalidate a chart pattern instantly.
Human review and automation
The signal itself is generated automatically. Administrative review checks operational quality: valid symbol, coherent candle history, plausible levels, and complete data. Review is not meant to substitute a new discretionary thesis for the model.
The separation matters. The model creates the candidate; the platform verifies that the candidate is safe to publish as data.
Known limitations
An AI chart model can fail because:
- Historical relationships change.
- Several markets become correlated during stress.
- Candle data arrives late or incorrectly.
- Thin liquidity produces misleading prices.
- Gaps bypass modeled stop levels.
- A rare event has no useful historical analogue.
No amount of training removes these risks. Responsible use requires transparent outcomes, risk limits, monitoring, and the willingness to publish failures.
What users should inspect
When reviewing a signal, look beyond confidence. Inspect the direction, entry, target, stop, maximum window, current status, and the chart before and after the signal date. Compare similar signals over time.
The model is most useful as a disciplined filter over a large market universe. It narrows attention to structured candidates. It does not replace judgment, risk management, or personal responsibility.
Continue exploring
Review the signal framework in context.
Create a free account to inspect delayed signal history and portfolio performance, or review the related public example where one is available.