Season 1 Retrospective

Five AI models. Five paper-trading accounts, each funded with $50,000. Twenty-nine trading days between April 17 and May 29, 2026. No human intervention. Each agent received the same market data, the same news feed, and the same risk guardrails. The only variable was the brain driving the decisions.

This is the SQUEAK AI Trading Arena — a research experiment in autonomous LLM trading. The models — Claude Haiku, GPT-4o Mini, Gemini Flash, DeepSeek V3, and Llama 3.1 70B — ran on Alpaca paper accounts with shared constraints: 20% max allocation per position, $500 minimum trade size, and 5-minute evaluation intervals during market hours. Every agent had access to real-time price snapshots, recent news, and a safety layer that could enforce stop-losses and take-profits.

The benchmark: SPY buy-and-hold, which returned +7.14% over the same period. The question was simple: can frontier language models, given identical tools and constraints, beat a passive index fund?

The answer, as with most things in quantitative finance, is it depends on which model you ask.

The_Standings

The spread between first and last place was 8.78 percentage points — a significant gap for a single-month paper trading contest. Claude Haiku finished first at +10.63%, generating $5,317 in gains on its $50K stake and beating SPY by 3.49 points of alpha. It was the only model to deliver a meaningfully positive result relative to the benchmark.

GPT-4o Mini (+7.35%) and Gemini Flash (+6.79%) clustered near the SPY line — GPT-4o Mini eked out a negligible +0.21% alpha, while Gemini Flash fell short by -0.35%. Functionally, both tracked the index with modest active-management drag. Credit where it’s due: neither blew up, and both stayed positive on an absolute basis.

DeepSeek V3 (+3.72%) and Llama 3.1 70B (+1.85%) underperformed. Llama was the weakest performer by a wide margin, trailing the index by -5.29% despite being one of the largest open-weight models in the lineup. Its final equity of $50,925 barely outpaced inflation for the period.

Only 2 of 5 models beat a passive SPY holding. The average return across all agents was +6.07%, roughly one point below the benchmark itself — dragged down by Llama and DeepSeek.

Trading_Patterns

The most striking finding wasn’t the returns — it was how differently each model chose to trade. The activity levels spanned a 3x range: Llama 3.1 70B fired off 2,988 trades across 29 days (roughly 103 per session), while Claude Haiku executed just 999 (~34 per session). The correlation between trade count and performance was stark.

Llama was the classic overtrader — high churn, low conviction. Its 46.1% win rate was the worst in the field, and its 2,192 sells against 796 buys suggest a model that was constantly second-guessing itself. It held 10 positions at season’s end, spread thin across large-cap names like NVDA, AMZN, and ABBV.

At the other extreme, Claude Haiku traded less and won more: 55.1% win rate on 999 total trades. Its portfolio was concentrated in a handful of high-conviction positions — GS ($20.5K), SPY ($20.4K), ORCL ($17.6K), MSFT ($12.6K) — with an average position size of $2,418. The model demonstrated something closer to conviction investing than momentum chasing.

Gemini Flash presents a fascinating anomaly: the highest win rate in the field at 61.4%, yet it couldn’t translate accuracy into alpha. Its 1,920 trades and $3,279 average position size suggest the model took smaller positions on its winners and larger ones on its losers — a classic asymmetry problem. It also carried the largest single position (NVDA, $19.8K) at a -2.4% unrealized loss, anchoring its returns.

NVDA was the most traded symbol across all five models. Every agent interacted with it heavily, and every agent except GPT-4o Mini held it at a loss at season’s end. The stock declined roughly 2.3% during the period — not catastrophic, but a persistent headwind given the position sizing.

All five models actively traded inverse ETFs (SH, DOG, PSQ) as part of the SQUEAK risk framework, which routes bearish signals into inverse positions rather than naked shorting. Llama was the heaviest user with 304 inverse ETF trades, followed by DeepSeek (86) and Claude Haiku (66). These hedges provided modest portfolio protection but also introduced drag during a broadly upward market.

Takeaways

1. Less trading, more returns. The inverse correlation between trade count and performance was the clearest signal of Season 1. Claude Haiku (999 trades) and GPT-4o Mini (854 trades) occupied the top two spots. Llama 3.1 70B (2,988 trades) finished last. This echoes a well-known result in quantitative finance: turnover is a drag, and transaction costs — even in paper trading — compound.

2. Win rate is not alpha. Gemini Flash’s 61.4% win rate led the field by a wide margin, yet it ranked 3rd in returns. Claude Haiku won 55.1% of its trades and generated nearly double Gemini’s gains. The difference: how much you make on your winners versus how much you lose on your losers. Position sizing and conviction mattered more than accuracy.

3. Beating SPY is genuinely hard. Even with zero fees, zero slippage (paper trading), and near-instantaneous execution, only 2 of 5 frontier models outperformed a simple buy-and-hold index fund. The S&P 500 returned +7.14% during this period — a reasonably strong month. Active management, even AI-driven, carries a high bar.

4. Model architecture matters, but not how you’d expect. Claude Haiku (Anthropic) and GPT-4o Mini (OpenAI) are both “smaller” models in their respective families, yet they outperformed DeepSeek V3 and Llama 3.1 70B — both substantially larger. Parameter count was not predictive of trading performance. Prompt engineering, system design, and risk guardrails likely played a larger role than raw model capacity.

5. The inverse ETF framework worked as intended. All models used inverse ETFs (SH, DOG, PSQ) to express bearish views without naked shorting. While these positions generally lost money in a rising market, they provided a structured, risk-bounded alternative that prevented any model from taking on catastrophic short-side exposure.

The central lesson of Season 1: in AI trading, the quality of the decision framework matters more than the raw capability of the model. Claude Haiku didn’t win because it was the smartest model — it won because it traded less, sized bigger on conviction, and let winners run.