Agentic AI in Trading

Materials for the opening workshop of QuantInsti's Algorithmic Trading Conference 2026, Thursday 24 September.

A coding agent was given a research brief and a panel of daily ETF prices, and it wrote a complete predictive pipeline: data loading, features, a label, a model, an evaluation. This repository holds what it wrote, the transcript of the run that wrote it, and a notebook that runs the code and takes it apart.

Start here

notebooks/audit.ipynb explains what the strategy is, reproduces the agent's headline result, then changes one line at a time and watches the result move.

It is committed with the outputs of a real run, so you can read every number on GitHub without executing anything. To run it yourself, open it in Colab: it takes about three minutes, and needs nothing installed, no account beyond a Google login, no API key, and no agent on your side. During the workshop the presenter runs it; you are not asked to execute anything live.

CHECKLIST.md is the takeaway: six checks under three questions - what are you predicting, what was knowable when, does the evidence say what you think it says - each one demonstrated in the notebook rather than asserted.

What is here

Path	What it holds
notebooks/audit.ipynb	The notebook, with stored outputs. Paired with audit.py via jupytext
src/	The five modules the agent wrote, copied unedited from runs/instructed-precise/
CHECKLIST.md	Six checks, with what each one caught
briefs/	The two research briefs handed to the agent, byte for byte
runs/	Four captured runs: environment, transcript, diff, findings
scripts/fetch_prices.py	Downloads the price panel from Yahoo Finance

The four runs

The same model, the same container, the same data, the same task. Two research briefs across two repository setups:

Run	Brief	Repository	Wall clock	Headline
bare-loose	one sentence	no instructions	4m 02s	long-short Sharpe 0.06 gross, -0.02 net
instructed-loose	one sentence	research standards in AGENTS.md	6m 28s	mean IC 0.0436, naive t 6.90, corrected t 1.99
bare-precise	five numbered steps	no instructions	10m 16s	mean IC 0.0285, naive t 6.10, corrected t 2.01
instructed-precise	five numbered steps	research standards in AGENTS.md	4m 02s	mean IC 0.0282, naive t 6.104, corrected t 2.013

All four answered the research question in the negative, which is the correct answer. They differ in how much evidence each one handed back. Each run directory carries its own ENVIRONMENT.md stating the model, the harness version, the container image, the home directory and the data checksums, so any two can be compared knowing exactly what differed.

src/ in this repository is instructed-precise's output, unedited. Its docstrings and comments are the agent's own and are part of what the notebook audits.

Running it locally

The price data is not shipped with this repository: the panel it came from may not be redistributed, so the fetch script downloads its own from Yahoo Finance.

uv sync
uv run python scripts/fetch_prices.py     # writes data/, gitignored

Then open notebooks/audit.ipynb in Jupyter or VS Code. Python 3.12 and uv. To run the agent's pipeline on its own, exactly as the agent ran it:

uv run python -m src.evaluate

License

MIT, including the code the agent wrote. The price data is not covered: it is not shipped here and is downloaded from Yahoo Finance at run time, under whatever terms that source imposes.

Where this continues

The workshop covers one iteration of one stage of a research workflow. The Machine Learning for Trading repository and book cover the whole of it, including the chapters on autonomous agents and on monitoring a model after it is deployed.