Bot vs Market

An AI agent trades a real EUR 1,000 against a buy-and-hold index fund, in public. Internal project name: finance.

Disclaimer. This is an experiment. An autonomous AI agent (Claude) manages a small real-money Scalable Capital Broker account (about EUR 1,000) with an explicit, versioned trading strategy written as code, and publishes every decision, trade, mistake and portfolio snapshot to a public site. The owner accepts the risk of losing the entire amount. Nothing here is investment advice or a recommendation to anyone. Do not copy these trades.

Site: https://fconforti.github.io/bot-vs-market/ · Glossary: docs/glossary.md · Mistakes: MISTAKES.md · Strategy: STRATEGY.md · Rules: MANDATE.md · First-run checklist: docs/handover.md

Status

This section is kept current by whoever last worked on the repo (the build agent so far).

Date	State
2026-11-03	2026-11-02 aborted: the weekly Scalable login hit its 7-day limit on 10-30 and no alert reached the owner. The 10-31 research commit had been refused (an overwritten backtest file) and is now pushed. The keep-alive now emails a reminder a day before a login runs out; it needs ALERT_EMAIL and SMTP_PASSWORD (a Google app password).
2026-10-23	Aborted again: the login made the evening before was deleted by the CLI itself. Upstream bug (ScalableCapital/scalable-cli#38): if the check after a token refresh hits one dropped network request, sc 1.1.0 throws away a valid login. sc upgraded to 1.2.0, which keeps the login and returns a retryable error. Scalable's documented limits: 20-minute access token, 24 hours idle, 7 days maximum from sc login. A keep-alive timer (finance-keepalive.timer, every 8 hours) covers the idle limit over weekends; the owner still has to log in once a week.
2026-10-22	Two days without a session. 10-21: the broker login had expired after at most about 50 hours. 10-22: the login was valid, but Scalable's sign-in server was intermittently unreachable over Wi-Fi and the runner reported it as an expired login. Read-only broker calls now retry on network errors (confirms never do), the pre-session check (bin/broker-check.py) retries for about five minutes and tells "not logged in" apart from "unreachable". 65 tests pass.
2026-10-20	Site: journal entries are split into a reader-facing part and a folded "Technical log" (the daily prompt now writes them that way; older entries are split by section). Mistakes page shows wrong decisions only; failures of the machinery go to OPERATIONS.md and a new Operations page (footer link). Two technical entries already in MISTAKES.md are reclassified there, append-only. Fixed page overflow from code blocks and bare URLs. Universe trimmed from 44 to 42 instruments: the two cash-like ETFs no version ever bought are gone, also from the CLI allow-list.
2026-10-20	Live, no trade yet. Timers on since 2026-10-19 (daily 18:15, research Saturday 10:00). The first scheduled run (10-19 18:16) exited at once on a live-only bash bug; the retry died on the machine's Wi-Fi (about 20 drops a day), as had the 16:03 dry run. A manual 09:01 run on 10-20 declined both buys: wrong time, 1.5% morning spreads, and a cached 10-19 'close' that was an intraday price. Fixed since: runners wait for the network and retry pushes; live sessions only run 17:45 to 21:30; the price refresh never stores an unfinished bar and fills a day Yahoo's daily series lacks from its last hourly bar; the site builds from live snapshots; a second session on the same day can append to the journal (only that day, only by adding lines); snapshots are redacted like trade files. Email alerts built, not yet configured. 58 tests pass. (2 VanEck Semiconductor ETF, 3 Xtrackers MSCI World Value ETF, EUR 1.98 in fees). The session's commit was refused by two pre-commit hooks (the confirm response's idempotency key looked like a secret; the benchmark series file changed for the first time), so the site was published by hand that evening; both hooks fixed.

What this is

An experiment in running a trading strategy with three properties that are usually missing from "AI trades stocks" projects:

1. The strategy is code, not vibes. strategy/current.py is a pure function from (portfolio, prices, parameters) to a list of proposed orders. The daily agent may execute those orders or refuse to trade. It may not invent orders, resize them, or change the algorithm. Only the weekly research session can change the strategy, and only with a written argument and a backtest.
2. The money is protected from the software, not from the market. Every order passes through bin/guardrails.py, the only code path that can confirm a trade with the broker. It re-reads the limits in mandate.toml, checks the proposal against today's account snapshot, asks the broker for a preview, compares the preview with the proposal, and confirms exactly once or not at all. Losing the money to a bad strategy is accepted; losing it to a bug, a misread number or a drifting prompt is not.
3. Everything is public and append-only. The journal, the trade files, the snapshots and the research notes are committed to this repository and rendered on the site. They are never edited after the fact; a pre-commit hook rejects such commits. Mistakes get their own page.

How a day works

Every weekday at 18:15 Europe/Rome (after the Xetra close, while EIX/gettex are still open), bin/run-session.sh starts a headless Claude Code session with prompts/daily.md and a short allow-list of tools. The session:

1. takes a snapshot of the account (bin/snapshot.py),
2. re-reads the last ten journal entries, the strategy, the research inbox and the mistakes page,
3. refreshes prices and reads the news, recording its sources,
4. writes a market view before any decision,
5. reviews each position against the thesis recorded when it was bought,
6. runs the algorithm and pastes its output verbatim,
7. accepts the orders or refuses to trade with a written reason,
8. writes one JSON file per order and hands each to the guardrails,
9. re-snapshots, records fills,
10. appends to the mistakes page or the research inbox if warranted,
11. rebuilds the site. The runner commits and pushes.

Every Saturday, bin/run-research.sh runs prompts/research.md: review the week, address the inbox, form at most one testable hypothesis, backtest it against the current version, and adopt it only with an argument that covers overfitting and regime risk. The default outcome is "no change".

The strategy in one paragraph

v002 is a slow, trend-filtered momentum rotation over 42 liquid EUR-listed instruments (mostly ETFs, a few large caps, gold, one bitcoin ETP). New positions are opened only while the MSCI World ETF trades above its 200-day average; everything is sold if it falls 3% below it. Up to four positions, equal weight, chosen by 12-1 month momentum; exits only in the first three trading days of a month, with a wide hysteresis band so the book does not churn. Two orders a day at most, market orders only. The backtest (2013 to 2026, with fees and spreads) and, more importantly, the reasons not to trust it are in STRATEGY.md.

Safety layers, from the outside in

Layer	What it does
Broker	Order size cap and ISIN allow-list in the Scalable CLI's own [trade_controls] config (set by the owner, outside this repo).
mandate.toml	25% of NAV per order, 40% per position, 2 orders a day, EUR 25 minimum, market orders only, one venue. Re-read on every call.
bin/guardrails.py	Validates, previews, compares, confirms once. Any mismatch, warning, questionnaire or error: no order. Never retries. Idempotent.
Claude Code permissions and hook	The agent cannot call the broker's trade commands, cannot write outside the journal/trades/inbox, cannot edit history, cannot push in a dry run.
Append-only history	Pre-commit hook rejects edits to past journal, trade, snapshot and research files.
Kill switch	bin/kill-switch.sh disables the timers and writes HALTED; every session checks for it first.

Layout

MANDATE.md / mandate.toml   rules and limits (human / machine)
STRATEGY.md                 current strategy, backtest, caveats, changelog
strategy/                   current.py, params.toml, universe.csv, data.py, backtest.py, versions/
bin/                        run-session.sh, run-research.sh, guardrails.py, snapshot.py,
                            session-info.py, build-site.py, check-append-only.py, kill-switch.sh
prompts/                    daily.md, research.md
journal/ trades/ portfolio/ research/    append-only history (dry runs suffixed -dryrun, not published)
web/                        Astro site; src/data/site.json is exported by bin/build-site.py
MISTAKES.md                 append-only log of decisions that turned out wrong (published on the site)
OPERATIONS.md               append-only log of failures of the machinery (site: Operations log)
systemd/                    user units (not enabled)
tests/                      pytest, with a fake `sc` that serves fixtures
docs/                       cli-notes.md (what the Scalable CLI really does), handover.md, glossary.md
finance/                    shared library: sc wrapper, mandate, models, trade files, snapshots

Running it yourself (owner)

cd ~/personal/bot-vs-market
uv sync && uvx pre-commit install && uv run pytest -q

Scalable CLI: enable "Agentic Investing" in the Scalable profile, then

mkdir -p ~/.config/scalable-cli
printf '[auth]\nsession_backend = "file"\n\n[trade_controls]\nmax_order_notional = "250"\n' >> ~/.config/scalable-cli/config.toml
sc login                          # device-code flow, done by a human
sc whoami --json                  # must print ok:true

The keyring backend does not work on this machine (no Secret Service on D-Bus), hence session_backend = "file".

uv run python bin/snapshot.py                  # portfolio/YYYY-MM-DD.json from the account
uv run python strategy/data.py --refresh       # price cache (yfinance)
uv run python strategy/data.py --check-live    # cached closes vs broker quotes
uv run python strategy/current.py              # signals and proposed orders, read-only
uv run python strategy/backtest.py --version current
uv run python bin/build-site.py               # exports web/src/data/site.json
(cd web && npm ci && npm run dev)              # http://localhost:4321/bot-vs-market/

./bin/run-session.sh --dry-run                 # full session, preview only, never confirms
./bin/run-session.sh                           # live
./bin/run-research.sh --dry-run

--dry-run uses the real account when sc is logged in and falls back to a fake sc with fixture data (clearly labelled in the journal) when it is not. Dry-run journals, snapshots and trade files stay in the repo for the record but are not published on the site and never feed the equity curve. FINANCE_TODAY=YYYY-MM-DD replays a date for a rehearsal on a weekend; never set it live.

Automation (not enabled)

mkdir -p ~/.config/systemd/user
cp systemd/finance*.service systemd/finance*.timer ~/.config/systemd/user/
systemctl --user daemon-reload
systemctl --user enable --now finance.timer finance-research.timer finance-keepalive.timer
systemctl --user list-timers
journalctl --user -u finance.service -f

User timers only fire while the owner is logged in, unless lingering is on:

loginctl enable-linger filippo     # currently off; needed for unattended runs

The services load ~/.secrets/finance.env (GH_TOKEN for pushing). Logs go to ~/.local/state/finance/. claude -p uses this machine's Claude Code login. GitHub Pages must be set to "GitHub Actions" once in the repo settings; the workflow in .github/workflows/pages.yml builds web/ with Astro and deploys it on every push that touches web/. The runner commits the exported site.json after each session, so every session redeploys the site.

Kill switch

./bin/kill-switch.sh      # disables both timers and writes HALTED to the repo root

Remove the file and re-enable the timers to resume.