The experiment
Can an AI actually manage a portfolio — and which kind of AI does it better?
Every market day, Claude agents review live market data and trade a $100,000
paper account through Alpaca. No real money moves. Each account runs a different
mandate — a written statement of what it is trying to do and what
it is forbidden from doing — and every order passes an independent risk desk
enforced in code before it reaches the broker.
The point isn't a leaderboard. It's a controlled comparison: the same market,
the same capital, the same guardrails, and genuinely different philosophies of how
an AI should decide. One book reasons freshly every morning. Another executes a
strategy that was decided once and is now enforced mechanically. Both publish every
trade and every rationale here, including the bad ones.
mainDiscretionary
Eric (mandate) + Claude (daily decisions) · live since 2026-06-28
Claude runs a fully discretionary review each market day: it reads the tape, researches positions and candidates, and decides what to buy, sell, or hold — a diversified book of quality large-caps weighed on momentum, valuation, and macro together, with capital preservation ahead of home runs. An independent risk desk in code checks every order against ratified limits: the agent proposes, the desk disposes.
- Max 20% of the book in one name; max 10% in any single order
- New buys stop for the day after a 2% loss — no revenge trading
- A 10% fall from the high-water mark halts buying until a human re-arms it
- 10% of equity must stay in cash after any purchase
- At most 3 buy orders per day; order sizes halve when VIX is above 25
- No new buys at all when VIX is above 35, or if the VIX feed is unavailable
fableSystematic
Claude Fable 5 · first session 2026-08-24
A deterministic engine ranks a fixed 43-name universe (mega-caps, sector ETFs, gold) by blended 12-1/6/3-month momentum behind a 200-day trend filter, holds the top 8 at ~12% each with weekly rebalances and enter-8/exit-14 hysteresis, and parks everything uninvested in T-bills. The daily LLM session executes the precomputed plan with veto-only discretion — it may refuse a buy on catastrophic news or broken data, never invent or resize one. The design bet: LLM churn destroys alpha, so judgment was spent once, in the design, and discipline is enforced daily.
- Top 8 names at ~11.75% each, ranked on blended 12-1 / 6m / 3m momentum
- Must be above its 200-day average AND positive on 12-1 momentum to be eligible
- Enter in the top 8, only exit below rank 14 — hysteresis, not a fresh cut
- Rebalance weekly; trend-break exits fire any day after 2-day confirmation
- Everything uninvested sits in T-bills (BIL), not idle cash
- The LLM may veto a planned buy on catastrophic news — never add or resize one
slot 3Open
unassigned
Provisioned and funded, waiting on a mandate. Reserved for a strategy from a friend of Eric's. Adding it is two environment lines and one registry entry; the account package may even bring its own session architecture rather than reusing the shared daily loop.
Total return
The race itself — every account starts at the same $100,000.
Sharpe ratio
Whether the return was earned or just borrowed from risk.
Max drawdown
The worst peak-to-trough fall — what holding it would have felt like.
vs S&P 500
The honest null hypothesis: could you have just bought the index?
Beta vs alpha
How much came from market exposure versus the agent's own decisions.
Every decision
Each session's trades and full written reasoning, kept verbatim.
FRTB capital
What a bank would have to hold against this book under the Basel market-risk rules.
Risk-adjusted metrics stay hidden
until an account has 20 sessions. Below that they are noise dressed
up as insight, and the verdict on the comparison itself was pre-committed at 60+
sessions before either book placed a trade.
Move off the desktop
The daily job runs on a machine that sleeps. Target is an always-on host; the prerequisite is moving per-account state into D1 so any runner can pick it up.
Fill the third slot
A third mandate turns a duel into a real comparison — and a third data point is what separates 'momentum beat discretion' from 'this one book had a good month'.
Publish risk metrics
Sharpe, max drawdown and S&P-relative performance stay hidden until an account clears 20 sessions. Below that they are noise dressed up as insight.
A third of sessions never ran
Only 23 of 39 trading days actually fired since June — the host machine was asleep, and a job that never starts says nothing. Both books miss together, so the comparison stays fair, but neither strategy is being executed as designed.
The sample is far too small to conclude anything
Momentum's edge is a few percent a year against much larger swings. Over weeks that is a coin-flip with a slight lean. The pre-committed bar is 60+ sessions, and anything read before then is narration, not evidence.
Paper fills are not real fills
Orders execute against simulated liquidity with no slippage or market impact. A concentrated book rebalancing weekly would meet real costs that never appear here.
The two books are not symmetrically falsifiable
Fable's strategy was frozen in writing with a stated evaluation protocol and predicted failure modes before its first trade. The discretionary book has a mandate but no pre-registered protocol, so it is harder to say it was 'wrong'.
Momentum's known crash case has not happened yet
Sharp V-shaped reversals are exactly where this strategy class hurts most: slow out of the old leaders, slow into the new ones. No such reversal has occurred during the test window, so the downside is untested rather than absent.
Survivorship in the universe
Fable's 43 names are today's liquid large-caps and sector ETFs — partly there because they already had momentum. That bias poisons backtests; it matters less for a forward-only run, which is one reason no backtest was used.