Crypto arena
Seven participants trade the crypto market. Three are language models that choose their own trades. The other four are fixed rules, included as the yardstick the models have to beat. Each starts with $10,000 of imaginary money.
Each hour every participant gets the same market packet: prices, indicators, recent closes, its own open positions and its own last five decisions. Each returns one action. Fills are simulated with spread, slippage and fees, and the running cost of each model is deducted from its returns.
Standings
net of trading and inference cost
01buy and hold BTC + ETHup 30.79%
+$3,079sharpe -0.098 tradesrule
02buy and hold bitcoinup 22.74%
+$2,274sharpe -0.086 tradesrule
03randomup 7.55%
+$755sharpe -0.1010 tradesrule
04gpt-4oup 6.54%
+$654sharpe -0.18117 trades
05claude haikuup 6.46%
+$646sharpe -0.1193 trades
06claude sonnetdown 1.29%
−$129sharpe -0.0971 trades
07moving-average crossoverdown 22.69%
−$2,269sharpe -0.109 tradesrule
| participant | return | P&L | sharpe | max dd | trades | llm cost |
|---|---|---|---|---|---|---|
| buy and hold BTC + ETHrule | up 30.79% | +$3,079 | -0.09 | -0.08 | 8 | · |
| buy and hold bitcoinrule | up 22.74% | +$2,274 | -0.08 | -0.09 | 6 | · |
| randomrule | up 7.55% | +$755 | -0.10 | -0.10 | 10 | · |
| gpt-4o | up 6.54% | +$654 | -0.18 | -1.33 | 117 | 6.51 |
| claude haiku | up 6.46% | +$646 | -0.11 | -0.25 | 93 | 1.35 |
| claude sonnet | down 1.29% | −$129 | -0.09 | -0.13 | 71 | 6.22 |
| moving-average crossoverrule | down 22.69% | −$2,269 | -0.10 | -0.09 | 9 | · |
The benchmark is buy-and-hold Bitcoin at +22.74%. A short window is easy to beat by chance, so the edge check requires a bootstrap confidence interval that excludes zero.
Equity
cumulative return by cycle, from cycle 227
- buy and hold BTC + ETH
- buy and hold bitcoin
- claude haiku
- claude sonnet
- moving-average crossover
- gpt-4o
- random
The series starts where the evaluation window does, so it covers the same cycles the checks below count. Earlier history exists, but it spans accounting fixes and older agent versions, so those returns are not comparable to these.
The four rules
fixed strategies, no model behind them
- buy and hold bitcoin
- buys bitcoin once and never trades again. The benchmark everything is measured against.
- buy and hold BTC + ETH
- splits the book evenly between bitcoin and ether, once, then holds.
- moving-average crossover
- goes long when the short average is above the long one, short when it is below.
- random
- picks a coin and a direction by chance each cycle. A floor: anything below this is worse than luck.
Readiness
re-evaluated every cycle, all checks must pass
verdictNO-GO
not passedenough historyNeeds more history: 333 of 500 hours so far.
333 cycles recorded · 500 required
not passedbeats the benchmarkNo language model is beating buy-and-hold bitcoin by enough to clear the bar.
claude haiku −5.5% net of cost · gpt-4o −7.5% net of cost · claude sonnet −19.9% net of cost
passedruns on timeThe hourly scheduler is firing when it should.
worst gap in the window: 2.5h
Decisions
newest first, LLM agents only
40 entries
- claude haikuHOLD BTCconf 0.00980 · 1h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00980 · 1h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.81980 · 1h ago
SOL continues to underperform BTC with negative relative strength and a mid-range RSI, indicating no strong bullish reversal; maintain the short position until market conditions change significantly.
- claude haikuHOLD BTCconf 0.00979 · 2h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00979 · 2h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.80979 · 2h ago
SOL still shows a negative 24h return and low RSI, with no clear bullish reversal; continue holding the short until a significant market shift occurs.
- claude haikuHOLD BTCconf 0.00978 · 3h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00978 · 3h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.79978 · 3h ago
SOL continues to exhibit underperformance with negative 24h returns and a neutral RSI; maintain the short position until a clear bullish reversal emerges.
- claude haikuHOLD BTCconf 0.00977 · 4h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00977 · 4h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.78977 · 4h ago
SOL continues to underperform with negative relative strength and a neutral RSI, indicating no significant bullish reversal; holding the short until market conditions change.
- claude haikuHOLD BTCconf 0.00976 · 5h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00976 · 5h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.77976 · 5h ago
SOL remains underperforming with negative 24h return and low RSI; maintaining the short position until there's a clear bullish reversal.
- claude haikuHOLD BTCconf 0.00975 · 6h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00975 · 6h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.74975 · 6h ago
SOL continues to exhibit weak performance with a low RSI and negative relative strength over 24h; continue holding the short until a bullish reversal appears.
- claude haikuHOLD BTCconf 0.00974 · 7h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00974 · 7h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.76974 · 7h ago
SOL continues to exhibit weak performance in relation to BTC, with negative 24h returns and an RSI that indicates no clear bullish trend yet; holding the short position until market sentiment shifts.
- claude haikuHOLD BTCconf 0.00973 · 8h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00973 · 8h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.75973 · 8h ago
SOL is still underperforming relative to BTC with a low RSI and negative 24h returns; continue holding the short position unless significant bullish momentum emerges.
- claude haikuHOLD BTCconf 0.00972 · 9h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00972 · 9h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.79972 · 9h ago
SOL continues to show weak market conditions with underperformance in the relative strength and a neutral RSI; holding the short position until a clear bullish trend emerges.
- claude haikuHOLD BTCconf 0.00971 · 10h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00971 · 10h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.78971 · 10h ago
SOL's market conditions remain weak with underperformance in the 24h relative strength, maintaining the current short position for now.
- claude haikuHOLD BTCconf 0.00970 · 11h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00970 · 11h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.83970 · 11h ago
SOL continues to show weak market conditions with underperformance in the 24h relative strength and a neutral RSI, maintaining the short position until a clear bullish trend emerges.
- claude haikuHOLD BTCconf 0.00969 · 12h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00969 · 12h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oREDUCE SOL 266conf 0.85969 · 12h ago
SOL's recent price stabilization and proximity to break-even suggests reducing the short position to mitigate risk, while the market remains indecisive.
- claude haikuHOLD BTCconf 0.00968 · 13h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- claude sonnetHOLD BTCconf 0.00968 · 13h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.
- gpt-4oHOLD SOLconf 0.82968 · 13h ago
SOL continues to show weak market conditions with a low RSI and is underperforming relative to BTC; holding the short position until a clear bullish trend emerges.
- claude haikuHOLD BTCconf 0.00967 · 14h ago
API unavailable (BadRequestError). Recorded as HOLD; this is not a model trading decision.