Ask AI about this page: Open in Claude Open in ChatGPT View as Markdown

How stocks are chosen¶

Claude chooses. Code bounds the choice. No code ranks the 50 stocks. The harness decides what Claude can see and what can go through; inside those bounds the pick is Claude's judgement.

flowchart TD
    U["Universe: 50 allowlisted tickers<br/>config/universe/*.csv"] --> L["list_universe<br/>symbol, price, my_pick flag<br/>(default: only if 1 share fits the position cap)"]
    L --> H["get_price_history<br/>per ticker, max 15 per run"]
    H --> J["Claude's judgement<br/>system prompt: objective, costs, diversification<br/>+ recent_activity"]
    J --> C["check_orders (optional)<br/>dry-run through the gate"]
    C --> S["submit_orders<br/>each order with a written reason"]
    S --> G["Risk gate (code)<br/>allowlist, 25% cap, cash, limit price, 10/month"]
    G --> E[Executor → IBKR]

What Claude sees¶

Tool	Gives Claude	Limit	Code
`get_portfolio`	cash, holdings with weights, average cost and unrealized P/L, drawdown, orders used, max value per holding, `recent_activity` (last 10 orders and outcomes)	—	`agent.py` → `_portfolio()`
`list_universe`	symbol, currency, price (local and AUD), `my_pick` flag. No company names or sectors.	filters: affordable, my picks, symbols	→ `_universe()`
`get_price_history`	6-month summary: returns (1w/1m/3m/6m), SMA20/50, 6m high/low, drawdown from high, annualised volatility, last 10 closes	15 calls per run	`market.py` → `summarize_history()`
`check_orders`	the gate's verdict and reasons for proposed orders	5 calls per run	→ `_check()`

What it does not see: news (web search is off by default), fundamentals, earnings, analyst views, or any model score.

What guides the judgement¶

The system prompt (agent.py → SYSTEM_PROMPT) sets the objective and trade-offs, not a formula:

maximise long-term total return in AUD within the limits; holding is a valid decision
each US order costs about US$1, so prefer fewer, deliberate trades
diversify, because a concentrated portfolio can hit the kill switch
prices may be ~15 minutes delayed
my_pick stocks are candidates, not obligations
don't blindly repeat orders that were just blocked or didn't fill

A real example (paper account, first day)¶

From the journal and the run transcripts:

Run	Stocks it looked at in detail	Bought	Claude's stated reasoning (abridged)
Open	NVDA, GOOGL, JNJ, MA, AMZN	all five	one per sector (semis, internet, healthcare, payments, consumer/cloud); positive 6-month momentum; JNJ as a lower-volatility anchor
Close	CSCO, FTNT, NFLX	CSCO	put idle cash (~8%) to work in a sixth holding; CSCO had lower volatility than FTNT

In the open run Claude fetched history only for the five it bought: it shortlisted from the ticker list using its own knowledge of the companies, then used the data to confirm. It did not compare all 50 on data. The close run did compare three candidates.

What this is and isn't

A language model's judgement on a small amount of price data, not a tested strategy.
Non-deterministic: the same inputs can produce different picks.
Breadth is bounded by the 15 history calls per run.
The paper-trading phase is how we find out whether the choices are any good.

Extension point: signal providers (planned, not built)¶

Selection is modular: Claude only ever sees tool results, and the gate never depends on how a proposal was reached. A new source of evidence, such as an ML model, plugs in as another read-only tool, with no change to the gate, executor or journal.

flowchart TD
    subgraph Offline["Offline, e.g. weekly"]
        D[price history + other features] --> T[train + walk-forward validation]
        T --> M[(versioned model file)]
    end
    M --> P["SignalProvider<br/>scores(universe, as_of)"]
    P --> Tool["get_signals tool<br/>score, horizon, model version, as_of"]
    Tool --> J[Claude's judgement]
    J --> G["Risk gate (unchanged)"]

Contract for any signal provider:

Rule	Why
Read-only; returns scores, never orders	the gate stays the only path to the broker
Every score carries its model version and as-of time, and is logged in the journal	audit and later evaluation ("did high scores outperform?")
Fails closed: missing, stale or invalid output → a tool error	Claude falls back to today's behaviour
Shadow mode first: scores are logged but not shown to Claude	measure out-of-sample value before letting it influence trades
Covered by scenario evals (outage, stale model, drift)	same bar as the rest of the harness

Where it would plug in: a provider passed into run_once() like the market data (pipeline.py), exposed as a tool in agent.py (TOOLS, _dispatch()), and simulated in evals/sim.py.