Argentarius · a rules-based crypto trading bot

Project Argentum

A personal project built almost entirely with Claude Code and run like a small research lab: specs first, pre-registered tests, every decision logged, and an AI reviewer checking the AI builder.

Not financial advice. No real money is trading yet.

Now Paper forward test Strategy D1 on BTC and ETH, Kraken spot, running unattended.
Next The QC loop Connecting the automated AI reviewer, with a triage protocol backed by evidence.
Then Simulate, go live A simulated-clock harness, 14 clean days, then small live trading.

Roadmap

Each phase starts only when the one before it is clean.

  1. 1

    Foundation and research

    Done

    Data importer, backtest engine and validation suite. The first four short-horizon strategies didn't qualify for paper trading; entry-level exchange fees set a high bar.

  2. 2

    D1, a pre-registered daily trend strategy

    Done

    Rules and gates were written before testing. It didn't pass all of them, so instead of tuning it, it runs unchanged as a forward test against frozen expectation bands.

  3. 3

    Paper operations

    Done

    Scheduled runs, Telegram briefings with strict privacy routing, a static read-only dashboard, security hardening, daily shadow checks against the backtest.

  4. 4

    Phase 3b · engine readiness

    Done

    Golden files that pin D1's outputs, change rules until live, and a tidy-up of the builder's rules. It replaced the multi-strategy platform, cancelled so the forward test isn't disturbed.

  5. 6

    Phase 3d · simulated clock

    Planned

    Months of scheduled runs in minutes: month and year ends, daylight-saving switches, missed runs, restarts and exchange faults.

  6. 7

    14 clean days, then live

    Planned

    Two weeks of clean paper on the final code, then a Linux server and small live trading with explicit approval at each step.

  7. 8

    Crypto ETF track

    Future

    A second strategy track trading crypto ETFs through a stock broker's API, researched under a written statistics standard built from 26 sources.

How it's built

Advisor
Claude in a Project: plans, reviews, writes phase prompts
Builder
Claude Code: proposes in plan mode, builds after approval
Reviewer
The Censor: headless, read-only, in its own repository
Owner
Me: approves every plan and every behavior change

Read the workflow

Guardrails

  • Plan mode first; nothing changes before approval.
  • More than 70 numbered decisions with their reasoning.
  • Golden files must stay byte-identical through refactors.
  • Deploys stay outside run windows, with the journal backed up first.
  • Secrets scanned at every commit and push; exchange keys can't withdraw.
  • Account data only ever goes to a private chat.

Research discipline

  • Pre-registered rules and gates; holdout used once.
  • Every variant logged, so results are judged against the number of tries.
  • Estimates with intervals, never "statistically significant".
  • Deflated Sharpe Ratio and data-snooping tests for any new signal.
  • Paper trading tests the plumbing; proving an edge takes years.

Read the methods

Stack

Core
Python 3.12 · uv · pytest · ruff
Data
CCXT · pandas · Parquet · DuckDB
Journal
SQLite · SQLAlchemy · Alembic
Ops
Task Scheduler → systemd · Telegram · static HTML
AI
Claude Code (desktop and headless) · Claude Projects