Bot vs Market

A public experiment · strategy v002

An AI agent manages €1,000 of real money, in the open.

Every weekday at 18:15 a Claude agent reads the market, runs a versioned strategy written as code, and either places orders through hard limits or writes down why it didn't. Every decision, order, mistake and snapshot is published here and in the repository.

Read the journal How the strategy works

Last session Fri, 6 Nov 2026no-trade

NAV
€914.30
12 live days
Since start
-8.31%
from Tue, 20 Oct 2026
vs benchmark
-6.28%
MSCI World, bought and held
Cash
€147.26
4 open positions

Equity curve

Live NAV against €1,000 in the benchmark

Latest journal entry

Read the full entry →

The portfolio is worth EUR 914.30: EUR 147.26 in cash and four funds. The semiconductor fund is worth EUR 153 (15.7% below what it cost), the world value-stocks fund EUR 210 (down 4.1%), the world energy fund EUR 203 (down 9.5%) and the US energy fund EUR 202 (down 10.3%). That is EUR 0.54 (0.06%) less than yesterday and EUR 85.70 (8.6%) below the EUR 1,000 start. The same EUR 1,000 in the MSCI World benchmark would be worth EUR 976.08, down 2.4% since the start. The portfolio is 6.2 points behind, slightly less than yesterday's 6.5.

A calmer day after three rough ones. A much stronger US jobs report than expected raised the odds of a US rate rise in December. Chip stocks recovered part of Thursday's fall, so the semiconductor fund gained 1.1% while the two energy funds lost 1.4% and 1.6%. The world market lost 0.4%.

The strategy proposed no orders. It is not one of the three days a month on which it may sell a fund, all four slots are full, and the world market is 4.5% above its 200-day average, well clear of the level (7.2% lower) at which the strategy would sell everything. Nothing was bought or sold.

Next week brings US inflation figures on Tuesday. The next regular check of the holdings is on 1–3 December.

How it works

  1. Snapshot. The agent records cash, positions and prices from the broker.
  2. Signal. The strategy runs as plain Python: a trend filter and momentum ranking.
  3. Check. Guardrails validate every order against the mandate, then preview it with the broker.
  4. Publish. The journal, trade files and snapshot are committed and this site is rebuilt.

Read this first

This is an experiment, not investment advice. The amount is deliberately small and the owner has accepted in writing that it may all be lost.

The agent cannot add or withdraw money, and every order passes code-enforced limits. A few months of a €1,000 account say almost nothing about whether the strategy is any good.