Research report
AI Hedge Fund
Claude proposes, code disposes. Every decision is journaled, scored and published.
Headline numbers
- Champion equity $9,999.53
- Champion return +0.00%
- S&P 500 return +0.00%
- Brier skill n/a
- Model cost $0.32
Equity
Each portfolio is indexed to 100 at its first point. The champion line is thicker. Baseline lines are thin and grey.
Leaderboard
Sorted by return after model cost.
| Portfolio | Kind | Return | Net of LLM cost | Max drawdown | Sharpe | Fees | Funding | Cycles |
|---|---|---|---|---|---|---|---|---|
| baseline:equal_weight | baseline | +0.00% | +0.00% | 0.00% | n/a, needs 30 cycles | $1.55 | $0.00 | 1 |
| baseline:random | baseline | +0.00% | +0.00% | 0.00% | n/a, needs 30 cycles | $0.76 | $0.00 | 1 |
| baseline:sp500 | baseline | +0.00% | +0.00% | 0.00% | n/a, needs 30 cycles | $0.90 | $0.00 | 1 |
| baseline:trend | baseline | +0.00% | +0.00% | 0.00% | n/a, needs 30 cycles | $1.31 | $0.00 | 1 |
| config:haiku-5.5/single | config | +0.00% | -0.00% | 0.00% | n/a, needs 30 cycles | $0.36 | $0.00 | 1 |
| config:sonnet-5.5/single | config | +0.00% | -0.00% | 0.00% | n/a, needs 30 cycles | $0.00 | $0.00 | 1 |
| config:opus-5.5/single | config | +0.00% | -0.00% | 0.00% | n/a, needs 30 cycles | $0.22 | $0.00 | 1 |
| config:opus-5.5/team | config | +0.00% | -0.00% | 0.00% | n/a, needs 30 cycles | $0.00 | $0.00 | 1 |
Calibration
No calibration bins for the champion yet.
| Config | Resolved | Brier | Brier skill |
|---|---|---|---|
| opus-5.5/single | 0 | n/a | n/a |
| opus-5.5/team | 0 | n/a | n/a |
| sonnet-5.5/single | 0 | n/a | n/a |
| haiku-5.5/single | 0 | n/a | n/a |
Current positions
Champion paper account
| Market | Asset class | Weight | Entry | Unrealised | Stop | Thesis |
|---|---|---|---|---|---|---|
| Ether | Crypto | -0.05 | 2,377.31 | -0.05% | 2,543.70 | ETH is the weakest crypto asset, 7.5% below its SMA20 and 14% off its 20-day high, with lower closes over the past several days and a risk-off tech backdrop. A small short is warranted given 43% volatility. |
Latest decision
Weekend session with all US-linked markets closed; only crypto can take new exposure. The macro backdrop is risk-off for tech: hot CPI and PPI driven by energy, an oil spike above $100 on Middle East attacks, and a 5% weekly Nasdaq-100 drop led by chips and AI. Crypto is drifting lower with ETH weaker than BTC, but the weekend makes conviction low, so positioning stays minimal.
| Market | Requested | Approved | Status | Reasons | Probability | Thesis |
|---|---|---|---|---|---|---|
| S&P 500 | +0.00 | +0.00 | approved | none | 0.49 | The market is closed, so no exposure can be added. The index is in a mild pullback, 1.6% below its SMA20 after hot inflation data. No view on weekend drift. |
| XYZ US 100 | +0.00 | +0.00 | approved | none | 0.49 | The market is closed. The index fell 6.3% on the week on worries about AI spending and China chips. Deeply negative funding signals crowded shorts, which offsets the weak momentum. |
| NVIDIA | +0.00 | +0.00 | approved | none | 0.49 | The market is closed. NVDA is 12.7% off its 20-day high in the chip selloff, while very negative funding suggests crowded shorts. Net, no edge over the weekend. |
| Apple | +0.00 | +0.00 | approved | none | 0.50 | The market is closed. Apple has shown relative strength above both its SMAs, but I have no weekend catalyst. |
| Microsoft | +0.00 | +0.00 | approved | none | 0.50 | The market is closed. MSFT is range-bound near its SMA20, and I have no view. |
| Tesla | +0.00 | +0.00 | approved | none | 0.48 | The market is closed. TSLA is in a steady downtrend, down 8.8% on the week, but very negative funding limits further downside conviction. |
| Gold | +0.00 | +0.00 | approved | none | 0.50 | The market is closed. Gold is flat near its SMA20 with no supply news, so I have no view. |
| WTI Crude Oil | +0.00 | +0.00 | approved | none | 0.53 | The market is closed. Oil is above $99, up 9% on the week, after Houthi attacks on Aramco and Iranian attacks on shipping in Hormuz. Geopolitical risk skews slightly up, but the market is extended 7.8% above its SMA20. |
| USD/JPY | +0.00 | +0.00 | approved | none | 0.50 | The market is closed. USD/JPY is in a tight range around 158 with low volatility, and I have no view. |
| Bitcoin | +0.00 | +0.00 | approved | none | 0.48 | BTC is drifting lower, 2.8% below its SMA20 and 7% off its 20-day high, alongside the tech selloff. Funding is neutral and there is no structural news, so the edge is too small to trade. |
| Ether | -0.05 | -0.05 | approved | none | 0.46 | ETH is the weakest crypto asset, 7.5% below its SMA20 and 14% off its 20-day high, with lower closes over the past several days and a risk-off tech backdrop. A small short is warranted given 43% volatility. |
Reasoning summary
It's Saturday evening with US markets closed, so I can only add exposure to crypto, meaning everything else stays flat at zero target. Looking at BTC and ETH, both are in a slight downtrend with low weekend conviction, and ETH looks weaker, trading well below its SMA20 and SMA50, which makes me consider a small short position there. I'll set ETH to a modest short of -0.05 with probability 0.46, and keep BTC flat near 0.49. For the closed markets, I'm just assigning probabilities close to 0.5 across the board, nudging CL slightly up to 0.52 given elevated geopolitical risk, while leaving NVDA and XYZ100 near 0.5 given their negative funding dynamics.
Recent cycles
| Time | Mode | Summary | Orders | Cost |
|---|---|---|---|---|
| 17 Oct 2026, 21:44 UTC | PAPER | Weekend session with all US-linked markets closed; only crypto can take new exposure. The macro backdrop is risk-off for tech: hot CPI an... | 1 | $0.32 |
Method and caveats
Evaluation is forward only. Large language models have memorised prices up to their training cutoff, so a backtest over that history would be contaminated. Removing tickers or dates does not repair it. Figures on this page start at the first published cycle.
Every configuration sees the same snapshot and the same news digest. Differences come from the configuration.
Fees, funding, and model costs are included. The leaderboard sorts by return after model cost. Sharpe is reported after 30 cycles.
Paper accounts share one fill model. When a live account is enabled, only the champion sends orders to the execution venue (Hyperliquid). Testnet uses that same path with test funds. Paper and live can differ because fills, funding, and timing differ.
Autonomy for the real venue has three modes. Paper never sends a real order. Approve prints the gated plan and waits for a person to confirm. Auto sends the gated plan without a confirmation. The default is approve.
This is an experiment, not investment advice.