The AI stack behind a trading system

A model is the smallest part of a production AI system. Around it sits a stack of infrastructure that gets data in, turns it into something the model can use, runs the model fast enough to matter, and watches the whole thing for failure. In trading, where decisions are tied to money and time, every layer of that stack carries risk.

The layers

1. Ingestion: collecting market data, news and reference data from exchanges, vendors and brokers.
2. Storage: keeping raw and cleaned data in a form that can be queried by time, instrument and session.
3. Feature engineering: turning raw data into the inputs a model uses, such as returns, volatility, session context or pattern flags.
4. Training: fitting models on historical data, usually on separate, heavier compute.
5. Serving and inference: running the trained model on live data and returning a decision or score.
6. Orchestration: scheduling jobs, moving data between stages and recovering from failures.
7. Monitoring: measuring the health of the data, the model and the decisions it produces.

Research versus production

Most AI work in trading begins in a research environment: notebooks, ad-hoc scripts and one-off datasets. That is the right place to explore. It is the wrong place to run money. Production systems need the same logic rebuilt with versioned code, tested data pipelines, reproducible results and clear ownership of every component.

The gap between a model that works in a notebook and a system that works in a live market is where most of the engineering effort in trading AI actually goes.

Educational content only. Not financial advice. This course describes engineering practice and does not recommend any trading system.