Skip to main content

Strategy Arena · Public research

LLM Trading Benchmark

Strategy results, their baselines and their limits. Not a universal ranking of AI models.

Published snapshot · 2026-09-13

Simulation · archive

This selection preserves the reported leader of each of the nine markets in the World Arena file. It is neither a new run nor a controlled LLM experiment. Strategies and periods differ between markets.

Markets in source
9
Strategy × market rows
196
Leaders displayed here
9

Period boundaries, detailed costs, engine version and candle hash are not supplied by this source. Update dates do not replace backtest dates. These figures alone do not reproduce a strategy or authorize its distribution.

Reported values, not annualized. Alpha is the gap to buy-and-hold in percentage points, not risk-adjusted alpha.
Market / strategyReturnBuy & HoldGap (pp)DDTrades
BitcoinDarvasFewer than 20 reported trades 17.01%-12.47%29.481.08% 10
CAC 40GA CAC 40 MomentumFewer than 20 reported trades 35.40%8.19%27.217.14% 13
DAXGA DAX ConquestFewer than 20 reported trades 22.38%5.25%17.135.67% 18
EUR/USDGA EUR/USD MomentumFewer than 20 reported trades 2.02%-0.13%2.151.55% 8
GoldGA Gold Mean Reversion 55.26%26.46%28.806.10% 28
Nasdaq 100Nasdaq 100 Buy & HoldPassive reference, not an AI win 38.88%38.88%0.0012.05% 0
OilOil Trend Conquest v2Fewer than 20 reported trades 74.28%53.70%20.5813.89% 7
S&P 500S&P 500 Buy & HoldPassive reference, not an AI win 24.88%24.88%0.009.18% 0
SilverGA Silver RobustFewer than 20 reported trades 142.10%91.55%50.5512.16% 16

SA-WA-20260913-e493a6a25189

What the leaders do not show

Showing only the leaders hides the spread within each arena. The same source also reports the following rows as the worst in their respective markets.

Promotion status is a registry declaration, not an independent verdict obtained in this publication. An AI name in a strategy does not establish its model version or call history.

Three different things to measure

Strategy designer
The model proposes rules and an engine executes them. The exact code and provenance of its creation matter.
Decision-making agent
The model receives information and decides during the experiment. Timestamped responses, tools, costs and actions are needed.
Simulated strategy
The result depends on rules, data and execution. This snapshot describes that third level without certifying the other two.

Dates and reproduction

The dates below are preserved as reported. Several markets show June 2026: publication in September does not make those observations recent.

Inspect market identities and dates
  • Bitcoin · BTC · BTCUSDT
    native_arena_state
    Reported update :
  • CAC 40 · CAC40 · ^FCHI
    native_world_arena_generator
    Reported update :
  • DAX · DAX · ^GDAXI
    native_world_arena_generator
    Reported update :
  • EUR/USD · EURUSD · EURUSD=X
    native_world_arena_generator
    Reported update :
  • Gold · GOLD · GC=F / GLD
    native_world_arena_generator
    Reported update :
  • Nasdaq 100 · QQQ · QQQ ETF
    native_world_arena_generator
    Reported update :
  • Oil · USO · USO ETF
    native_world_arena_generator
    Reported update :
  • S&P 500 · SPY · SPY ETF
    native_world_arena_generator
    Reported update :
  • Silver · SILVER · SI=F / SLV
    native_world_arena_generator
    Reported update :

For a future controlled comparison: fix model identities, rules, dates, costs and baselines before the trials, then publish trades and the engine receipt. No new backtest was run for this page.

Archived at :
Source-reported dateModified :