Case study · Always-on data systems
Two trading bots that have to prove themselves on paper first
The long-term goal is a bot that trades real money. The short-term rule is that it has to earn that first. So I built two research systems with Claude Code. They watch live markets around the clock and trade them with pretend money, so every idea gets tested against the real market before any real dollars are involved.
Where it started
It started as a bare-bones starter prompt I found on X. The first bot Claude Code built from it ran, but it didn't really work. Over 200+ commits I turned it into two full systems: real market data, multiple strategies, testing, monitoring, reporting and backups, running around the clock.
Paper first, real money later
For now the bots only read public market data and trade pretend money. No brokerage account or keys are connected, and an automated check on every change makes sure it stays that way until a strategy has proven itself and I decide it's time. Paper trading is the proving ground. Real money comes only after the results hold up.
What they do
- Record live markets around the clock and keep a full audit trail of every decision the system makes.
- Run several strategies in parallel on pretend money, so ideas can be compared on identical data.
- Protect themselves. They have built-in loss limits, and a one-file kill switch pauses all simulated trading while data recording keeps going.
Honest numbers, not exciting ones
The hardest part of this kind of research is not fooling yourself. The systems are designed to be skeptical of their own results: every strategy is measured against a simple benchmark, the reports flag when a result could be luck, and ideas that fail stay labeled as failed instead of being quietly deleted.
Built to run without me watching
- Live dashboards that update the moment something happens. A red banner means something is broken and yellow means it needs a look.
- A one-page check-in brief: is it healthy, is anything working, are the backups current.
- Daily and weekly reports, plus phone notifications when something needs attention.
- Nightly verified backups stored off the main drive, with alerts if they go stale and a written restore procedure.
- Plain-English operating docs so a non-programmer (me) can run, pause, deploy and recover it.
My role
I'm not a traditional programmer. Claude Code wrote the code. My part is the direction and the iteration: I took a starter prompt that didn't work and kept prompting, running it, seeing what came back, and asking for the next improvement, round after round, until it turned into a real system.
Now the bottleneck isn't building. It's data. The next improvements on the roadmap depend on months of live results, so for now the bots run and collect, and the next round of work picks up when there's enough real data to learn from.
Why the code isn't public: the strategies, settings and findings are my own research, so the repositories are private. This page covers how the systems are built and run, not what they trade or how well it's going. I'm happy to walk through the architecture in an interview.
Nothing here is financial advice. These are research tools that use pretend money only.
Stack: Python · Docker · GitHub Actions CI · automated testing · live web dashboards · Claude Code