Every AI trading system, from a hobby script to an institutional desk, runs the same five-stage pipeline. Knowing the stages lets you cut through any pitch and ask the only question that matters: where is the edge supposed to come from?
1. Ingest the data
It starts with a live feed — prices, volume, the order book, sometimes options data or news. Garbage in, garbage out applies brutally here: a laggy or dirty feed poisons everything downstream. See polling vs streaming data for why the feed design matters.
2. Turn data into features
Raw prices are useless to a model. The system computes features — things like relative volume, trend slope, volatility, or order-flow imbalance — that summarize what is happening. This step is where most of the real skill lives; a model is only as good as the features it is fed.
3. Run the model
The model maps features to an output: a direction, a probability, a score. This can be a simple rule, a classic ML model, or a neural network. Bigger is not better — a model that overfits its training data looks brilliant on history and fails live (see look-ahead bias).
A model that dazzles on a backtest and dies live isn't unlucky — it memorized the past instead of learning the market.
4. Apply rules and risk limits
A raw signal is not a trade. Before anything fires, it passes through hard rules: position sizing, stop-loss, maximum exposure, and vetoes for bad conditions. This is the guardrail layer, and it is what separates a tool from a liability — more in why AI needs guardrails.
5. Execute and manage
Finally the order goes to a broker via API, and the system manages the exit. Fast, consistent, unemotional execution is the payoff — the part humans do worst. NoVo runs this stage in your own broker account, non-custodially (your money, your keys).