AI Trading Ledger: eleven bots, one pre-registered public track record
The first trading day is September 8, 2026. Until then the chart is empty on purpose — nothing is backfilled.
| Arm | Status | Return | Max drawdown | vs SPY | Trades |
|---|---|---|---|---|---|
| Loading /paper-ledger.json… | |||||
The rules, fixed before the first trade
| Arm | Rule | Why it is here |
|---|---|---|
| SPY hold | Buy SPY at the first close, hold. | The benchmark every "AI trader" has to beat. |
| QQQ hold | Buy QQQ at the first close, hold. | The benchmark our 13F copy-homework page uses. |
| 60/40 | 60% SPY, 40% AGG, rebalanced on the first trading day of each month. Added 2026-09-05, before the start. | The static bar every timing rule has to clear — GEM's post-publication record (2014–2026) did not clear it. |
| AGI basket | Ten names from the exposure tool's two presets (NVDA, AMD, TSM, AVGO, MU, MSFT, GOOGL, AMZN, AAPL, META), equal weight, rebalanced on the first trading day of each month. | The naïve "buy the AI thesis" portfolio, with no stock picking after the start. |
| Tracker mix | Weight in the AGI basket = Thesis Tracker score ÷ 100 (62.5% at start), remainder in SPY, rebalanced monthly. | The only arm that reads this site's own verdicts. If the scorecard has information, it should show here. |
| 200-day rule | Faber (2007): at each month-end close, hold SPY if it is above its 200-day average, otherwise cash. Executed at the next close. | The best-known published timing rule — a decades-old baseline, not a discovery. |
| LLM agent | Once a week, a language model is shown the last 60 days of prices, the Tracker score and this site's llms.txt, and returns long-only weights across the ten names, SPY and cash. Executed at the next close; its stated rationale is logged. Not started until an API key exists — the arm shows "not started", never a backfill. | The actual experiment: the thing the funding wave is selling. |
| GEM dual momentum | Antonacci (2014) Global Equities Momentum, monthly: if SPY's 12-month return beats T-bills (BIL), hold whichever of SPY or ex-US equity (VEU) has the higher 12-month return; otherwise hold US aggregate bonds (AGG). Added 2026-09-05, before the start. | The most-followed published dual-momentum rule; its live record since publication is part of what the read date will test. |
| GTAA-5 | Faber (2007/2013) five-asset version: SPY, VEU, IEF, VNQ, DBC at 20% each, held only while above the 200-day average, otherwise that slice sits in cash; monthly. Added 2026-09-05. | The multi-asset form of the 200-day rule; historically the drawdown-cutting arm, not the return-maximising one. |
| SPY vol target | Hold SPY at a weight of 10% ÷ realised 21-day volatility, capped at 100% (never levered), rest in cash; monthly. Added 2026-09-05. | Moreira-Muir (2017) style volatility management, in the only form a cash account can run. |
| Basket momentum 5 | Inside the same ten AI names: rank by 12-1 month momentum (Jegadeesh-Titman), hold the top five at 20% each; monthly. Added 2026-09-05. | The one arm that picks stocks — with a published rule, not an opinion. |
Shared assumptions: $10,000 paper per arm · adjusted closes (dividends included) · 5 basis points per side on every trade · signals computed at a close are executed at the next close · data from a public price API on a GitHub runner, cached ticker-by-ticker; a ticker that cannot be fetched keeps its last value and is flagged in the JSON, never invented. Machine-readable: /paper-ledger.json (CC BY 4.0). Code: tools/paper_ledger.py in the public repository.
LLM arm decisions
Pre-registered judgement
- Read date: March 8, 2027 (six months). Every arm's return, maximum drawdown and excess versus SPY are published as they stand. No arm is promoted, renamed, or quietly restarted.
- What would change the parent verdict: nothing on this page alone. Six months in one regime is a demonstration of method, not evidence of skill; the parent page's flip condition is net user-level outperformance after costs, in a vendor's own audited ledger.
- What this ledger will never become: a signal service, a subscription, a "connect your broker" button, or a page that tells you what to buy. Positions appear only after their paper execution date.
- Pipeline kill: more than five consecutive weekday runs without fresh prices turns the workflow red until fixed; the ledger is never advanced on stale data.
FAQ
Is the AI trading ledger investment advice or a signal service?
No. It is a paper-trading experiment with rules fixed before the start date. No security is recommended, nothing is executed for anyone, no real money is involved, and positions are shown only after the paper execution date. Nothing on this page is a recommendation to buy or sell anything.
Why publish a paper ledger at all?
Because /do-ai-trading-agents-work holds vendors to a standard — positions timestamped before outcomes, every result graded, misses never deleted, methodology fixed in advance — and the only honest way to demand that standard is to meet it ourselves. The ledger is the demonstration, not a product.
What happens if one arm beats the S&P 500?
Nothing automatic. The pre-registered read date is 2027-03-08 (six months). On that date every arm's return, maximum drawdown and excess versus SPY are published as they are. Six months of paper results in one market regime is not evidence of skill, and the page will say so.
Related: Do AI trading agents make money? · Does copying 13F filings work? · AGI investing, mapped · Browse all AGI questions. Educational information only; nothing here is investment advice or a recommendation to buy or sell any security.