The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Not reliably more net returns, at least in the setup tested. In their 2026 study of LLMs trading U.S. equities, Jiayi Chen and Guiling Wang report that increasing test-time reasoning did not reliably improve portfolio returns after trading costs. That is a bounded result—not proof that reasoning effort never helps in other markets, periods, or trading systems.
What did the study find?
Chen and Wang’s paper, The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading?, reports no reliable improvement in net portfolio returns from additional reasoning across the tested DeepSeek, GPT, and Gemini model families. The authors’ abstract puts it this way: “Across all three model families, additional reasoning does not produce a reliable improvement in net portfolio returns.”
As an Amazon Associate I earn from qualifying purchases.
The result is about portfolio performance after the experiment’s trading costs—not whether a model produced a longer rationale, changed a forecast, or earned a better prediction-level score. For an investor, the relevant question is whether those changes lead to a better portfolio once costs are counted.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchHow was reasoning effort tested?
The authors varied reasoning effort while holding information available at each portfolio formation date, prompts, output formats, and portfolio construction fixed. The evaluation covered one year of U.S. equities and compared three input conditions: numerical information, identifiable news, and masked news. The paper reports more than 800,000 asset predictions and repeated model generations.
#1 Best Overall
That design aims to isolate the economic effect of changing reasoning effort within the study’s process. It does not make the result a general test of every way an LLM might be used in finance: a model that answers questions about company fundamentals, generates trading signals, designs a strategy, or executes trades is doing a different job.
Why isn’t more reasoning automatically better?
Additional inference can change a model’s output, but a changed output only matters economically if it improves portfolio decisions enough to outweigh the costs of trading and the extra computation. A polished explanation is not itself evidence of better returns. Nor does a stronger score on a financial question necessarily establish that a strategy is profitable.
The study’s reported pattern also cautions against assuming a simple “more effort, better result” curve. For DeepSeek, outcomes across reasoning levels were nonmonotonic: increasing effort did not produce a steadily improving sequence. Across repeated generations, treatment effects and selected portfolios were unstable even when overall scores were similar.
What can—and can’t—the result establish?
The conclusion applies to the models, one-year U.S. equity setting, information conditions, and portfolio process that Chen and Wang evaluated. It should not be extended automatically to other markets, time periods, model versions, trading frequencies, or agent designs.
The accessible abstract-level account does not state exact return magnitudes, model version identifiers, the transaction-cost schedule, confidence intervals, or detailed statistical significance results. Those details are therefore not established here; the reported conclusion is about the absence of a reliable net-return improvement in the tested setup, not a quantified estimate of how much reasoning helped or hurt.
How do other financial LLM evaluations differ?
Financial benchmarks test different capabilities, so their results are useful context rather than a substitute for a portfolio-return experiment.
| Evaluation | What it tests | What the reported evidence says | What it does not establish |
|---|---|---|---|
| FinTradeBench: A Financial Reasoning Benchmark for LLMs (2026) | 1,400 questions grounded in NASDAQ-100 companies over a ten-year historical window, grouped into fundamentals-focused, trading-signal-focused, and hybrid questions. | The authors evaluate 14 LLMs and report a clear performance gap. Retrieval substantially improves reasoning over textual fundamentals but provides limited benefit for trading-signal reasoning. | Benchmark question performance is not evidence of live investment results or net portfolio returns. |
| QuantEval: A Benchmark for Financial Quantitative Tasks in Large Language Models (2026) | Financial knowledge questions, quantitative mathematical reasoning, and quantitative strategy coding. Generated strategies are evaluated in a CTA-style backtesting framework. | The paper says it releases a deterministic configuration covering the asset universe, cost model, and metric definitions. | A backtest under a specified benchmark configuration is not, by itself, proof of live trading performance. |
| Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens (PMLR, 2026) | An inference-time sample-selection method called Think@$n$. | The paper’s page reports matching or exceeding standard self-consistency performance while lowering inference costs through early rejection of unpromising generations. | This general reasoning-efficiency result does not show that more reasoning improves trading returns. |
| AlphaForgeBench: Benchmarking End-to-End Trading Strategy Design with Large Language Models (arXiv / KDD 2026 paper page) | Trading-strategy design, including generating executable factors and strategies. | The benchmark describes run-to-run variance, inconsistent action sequences, and action flipping as evaluation concerns. | Strategy design is distinct from testing reasoning effort in a fixed portfolio process or evaluating a direct execution agent. |
What should a claim that reasoning “pays” include?
A meaningful claim needs to connect the added inference to both an economic outcome and the conditions under which it was measured. At minimum, check:
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →- Net outcome: portfolio returns after the experiment’s trading costs, not forecast or question-answering scores alone.
- Compute cost: how much additional inference the reasoning uses and whether its cost is justified by any portfolio change.
- Repeatability: whether portfolio choices and effects remain stable across repeated generations.
- Information timing: what data was available at each decision date and how the evaluation controls for information leakage.
- Task definition: whether the system is being assessed on financial knowledge, signal generation, quantitative reasoning, strategy design, or trade execution.
Chen and Wang provide a useful controlled test of one part of that question. Their reported result argues against treating extra reasoning as a dependable shortcut to better net returns; it does not settle whether a different reasoning method, model, or trading system could succeed under a separately measured evaluation.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
