What a trading model actually sees

When people imagine an AI trading system, they often picture it "looking at the chart" the way a person does, recognising shapes, sensing mood. That is not what happens. A model does not see a candlestick chart; it sees numbers. Understanding which numbers, and how they are chosen, demystifies most of what an AI-assisted system is really doing, and reveals where the real craft lies.

[01]Features: the model’s senses

The inputs a model reads are called features, and each is a specific measurement derived from raw market data: recent momentum, current volatility, the spread, the session or hour, the distance from a moving average, the relationship between timeframes. Every feature is one sense, one way of quantifying a property of the market. The model does not experience price; it experiences this bundle of measurements, refreshed each time it is asked to decide.

[02]Garbage in, garbage out

The uncomfortable reality of machine learning is that the choice of features usually matters more than the choice of model. Feed a system noisy, redundant or irrelevant inputs and even a sophisticated algorithm will learn little of value. Feed it a small set of genuinely informative measurements and a fairly simple model can do well. This is why which indicators you choose is not a trivial preliminary, it defines the entire universe the model can reason within.

[03]Weighing, not triggering

Where a hand-coded rule treats an input as an on/off switch, a model treats each feature as evidence with a weight. It learns, from a great deal of history, how much a given signal should count in a given context, leaning on momentum when momentum has been informative, discounting it when it has not. The output is less a verdict than a balance of many weighted considerations, which is what lets it adapt as conditions shift rather than fire the same rule forever.

[04]The limits of seeing numbers

This framing also makes the limits clear. A model can only see what it is given; a feature nobody thought to compute is invisible to it, and a genuinely novel event outside its historical experience is something it has never learned to weigh. It is powerful within the world of its features and blind beyond it, which is exactly why these systems are bounded by risk limits and monitored, not handed the keys.

This is educational and illustrative, not financial advice. No model, however well-fed, can guarantee results; understand what a system can and cannot see, and manage risk deliberately.