Trading-agent study separates strong portfolio returns from evidence of AI value

What the system actually does

The November 8, 2026 preprint describes a hybrid system: language models assess financial information, while fixed software rules select stocks, allocate capital and translate targets into orders. Its results are preprint findings, not established evidence of deployable investment performance.

Two specialist roles use Claude Haiku 4.5. One assesses company headlines alongside price and sector context; the other examines market indicators and macroeconomic news. Their structured reports contain numerical scores, confidence levels and, for the macro analyst, an action label such as “hold.” Unusable responses trigger deterministic fallbacks—fixed procedures rather than another model judgment.

The portfolio’s main engine is momentum: ranking stocks by their recent relative price performance, here over one and three months. It targets ten stocks. Momentum rank dominates the weighting rule, while a composite score incorporating other inputs makes a smaller adjustment. Individual target positions are capped at 25%, but the strategy has neither a sector-neutrality constraint nor a volatility target.

The model analysts therefore do not have unrestricted trading authority. Their influence depends on which fields the coordinator reads and what those fields are allowed to change.

Strong returns, substantial risk—and no historical AI attribution

The broker-recorded paper account started with $100,000 and returned 34.5% between March 11 and August 28, 2026. Over identical dates, SPY, an exchange-traded fund tracking the S&P 500, returned 14.4%; QQQ, tracking the Nasdaq-100, returned 18.2%.

But this was simulated capital, not live money. More importantly, the account used deterministic fallbacks throughout the recorded period. Its gains cannot be credited to language-model analysis.

The higher return also came with greater risk. Maximum drawdown—the largest fall from a previous portfolio-value peak—was 26.8%, compared with 6.3% for SPY. The account’s Sharpe ratio, which summarizes return relative to variability, was lower than SPY’s under the paper’s calculation. Beating the index at the endpoint did not mean delivering a better risk-adjusted path.

Testing what the models add

To examine the analyst integrations, the researchers reconstructed 119 historical decision dates and compared three configurations. A baseline used only fallbacks. An advisory version let model scores influence weights and cash reserves. A sizing version additionally translated the macro analyst’s action label directly into a cash target.

In the original reconstruction, returns were 56.4%, 55.5% and 58.7%, respectively, including the initial trading interval. These are simulated outcomes over a different window from the account record, not additional realized account returns.

The sizing configuration mostly recommended an action producing 5% target cash. A simple constant-5%-cash control returned 59.0%, close to the sizing version without model calls. That comparison leaves a straightforward exposure explanation plausible: holding less cash, rather than extracting superior information, could explain the difference. The controls do not perfectly match every policy feature, however, and wide uncertainty intervals prevent a precise ranking.

Company sentiment alone had little allocation influence at the chosen settings. It changed no selected stocks relative to the baseline and produced small weight adjustments. That reflects its limited position in the allocation rules, not proof that financial news is universally uninformative.

Better formatting can change the advice

The study also tests the interface between model responses and software. Shorter company-report formats eliminated observed parsing failures in two matched rich-headline test groups at the same response-token limit. Parsing means successfully reading the reply into the required structured fields.

Yet easier parsing did not guarantee stable recommendations. Shortening the macro report changed its action label on 57 of 117 dates where both versions parsed, despite strongly correlated numerical scores. A formatting change could therefore alter cash decisions, not merely improve reliability.

A separate offline experiment selected model-call dates near cash-rule thresholds. A cutoff calibrated on the first 60 dates captured about 98% of later allocation differences while selecting about 68% of the remaining dates. This measures preserved allocation changes—not preserved profits—and still needs prospective testing.

Why the investment conclusion remains limited

The original replay sometimes combined later intraday inputs with simulated opening-price fills. Deferring every assessment to the next session removed that particular timing mismatch, and all three configurations still finished above the benchmarks. But it changed signal age and holdings paths rather than recreating historical trading.

Other limitations remain: substituted historical headlines, model responses collected after most decision dates, a stale event calendar in primary prompts, incomplete strategy history and unavailable price archives. The reconstruction also exceeded the recorded account substantially, and the proprietary strategy prevents full public reproduction.

The clearest contribution is an inspectable account of how analyst reports reach portfolio decisions. Demonstrating persistent investment value will require prospective comparisons with fixed information and execution clocks, explicit fallback records and matched exposure controls.