Can Claude out-trade a bot?
About this video
A shared market on tradefloor has one order book, one clock and $1,000,000 for each seat, trading 17 Nasdaq-100 tech stocks. Seat 1 is Claude, which trades through tradefloor's MCP tools and leaves a note on each order. Seat 2 is the momentum bot from the crash video at 1.5 times leverage, and seat 3 buys and holds all 17 stocks in equal weights.
The bots run from the terminal, and the market moves one day each time all three seats are ready. The page shows each seat as a lane on the branch graph and every seat's net worth on one chart, with the index dashed. A liquidity crisis lands on day 8. After day 25, Claude is asked how it did against the bots and answers from tradefloor's data.
Chapters
- The market
- Claude trades
- Result
- Debrief
Figures on screen
Every figure here is on screen in the video and was checked against the product's own data.
- On day 8 the VIX went from 14 to 47 and order books thinned to 40% of their depth.
- Claude held a long book at about 1.6 to 1.8 times on days 1 to 7. After the crisis hit it sold every long and went short, up to 1.29 times, then held a 0.5 times long book from day 16 to the close.
- Net worth on day 25: Claude $920,840 (−7.9%), buy and hold $816,321 (−18.4%), momentum bot $758,169 (−24.2%).
The guide
Claude against two bots makes a good show. Testing an AI agent properly takes a little more ceremony, and Testing an LLM trading agent without look-ahead bias covers it: markets no model has seen, seeds sealed before the run and every model answer recorded for replay.
More demos
2:15Would my trading bot survive a crash?Branch a market into a rate shock and a liquidity crisis, run the same bot on every branch, ask Claude why it lost, and replay the crisis with Claude's one-line fix.
2:20Does my trading bot pass CI?A pull request raises a bot's leverage and tradefloor's stress suite fails it in CI. Claude fixes the bot from the CI log without loosening the check, and CI passes.
3:02Skill or luck?Claude tunes a trading bot on 8 public markets and none of its changes beats the original. On fresh hidden markets the unchanged bot scores anywhere from 33 to 51.
3:18A trading classA teacher sets a credit crisis from a template. One student's AI agent trades it, another trades by hand and hands in, and the teacher watches live and marks.
3:47No installA new account writes a trading bot in the browser, tests it against three crises in the same tab and asks Claude which one hurt, with nothing installed.