Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents

Evaluating Investment Logic in Large Language Models: A Real-World Benchmark Towards Personalzied Financial Agents

Yuanhong Jiang * † 1 2 Jingjie Zou * 3 Zhenghong Lin 1 Xusheng Yu 4 Qiqi Huang 4 Shuai Jia 2 Shijie Dai † 5 2

arXiv:2608.06108v1 [cs.AI] 6 Aug 2026

Abstract

1. Introduction

The dominant paradigm for evaluating Large Language Models remains anchored in chat and question-answering (QA) tasks, creating a fundamental mismatch with the competencies required for reliable agent execution in real-world domains. This disconnect is acutely evident in financial decision-making, where proficiency is defined not by knowledge recall, but by the capacity to align Person—the investor’s unique profile, constraints, and risk preferences—with the interpretation of unstructured market Events (E), the synthesis of coherent investment Reasoning (R), the formulation of corresponding Decisions (D), and the pursuit of profitable Outcomes (O). Without evaluation grounded in this integrated P→E→R→D→O chain, progress toward capable financial agents remains stalled.

The rapid integration of Large Language Models (LLMs) into finance has transformed tasks ranging from sentiment analysis and earnings call summarization to macroeconomic forecasting, yet their deployment in live trading environments reveals significant performance limitations. This progression has culminated in the emergence of live-trading arenas where LLM-based agents compete with real capital, relying solely on real-time news and market data (Fan et al., 2025). These platforms provide an unforgiving, forwardlooking test of practical capability, revealing a striking dissonance: models that excel on established financial reasoning benchmarks often generate unprofitable or unstable strategies in live markets. The observed “Capabilities-Performance Paradox”(Ding et al., 2024) points to a fundamental distinction between financial knowledge and investment intelligence. Current LLMs demonstrate impressive proficiency at retrieving facts, solving textbook problems, and performing quantitative calculations—skills meticulously measured by benchmarks such as FinBen (Xie et al., 2024) and BizBench (KoncelKedziorski et al., 2024). Yet they lack the higher-order logic required for successful market participation: the ability to distill salient signals from unstructured news, connect disparate geopolitical, corporate, and macroeconomic events into a coherent narrative, adjudicate between conflicting data points under uncertainty, and formulate—then steadfastly maintain—a strategic thesis that aligns with a consistent investment style.

To bridge this gap, we introduce I NVEST L OG IC B ENCH 2026, a benchmark derived from the documented decision-making of real-world investors. It provides structured, validated traces that explicitly instantiate such investment logic chain for each investment thesis. Our evaluation reveals a pronounced investment logic gap: state-of-the-art LLMs encounter significant bottlenecks in both emulating the reasoning of expert investors and translating such reasoning into profitable outcomes. Our benchmark establishes the first reproducible and market-aligned standard for quantifying this deficiency. This work shifts the evaluation paradigm from passive knowledge recall to active reasoning, paving the way for building robust, end-to-end LLM investment agents that generate market-aligned and personalized decisions adaptable to diverse strategies. The project is publicly accessible in https://anonymous.4open.science/r/InvestLogicBenchC2D7.

We argue that this failure is not incidental but systemic, stemming from three structural mismatches between the prevailing LLM paradigm and the cognitive requirements of successful investing: Objective Mismatch: LLMs are trained via next-token prediction on vast corpora, optimizing for linguistic coherence over causal reasoning. Raw financial text, while abundant, rarely contains explicit annotations of the investment decision cycle: how an investor’s unique Person (P)—comprising their objectives, risk constraints, and behavioral biases—interacts with market Events (E), which are then interpreted and synthesized into investment Rea1

I NVEST L OGIC B ENCH 2026

2. Can Language Models Make Profitable Investment in Real-World Markets?

soning (R), leading to corresponding Decisions (D), and ultimately aiming at profitable Outcomes (O). Without exposure to structured P→E→R→D→O investment logic chains, models struggle to internalize the instrumental logic of investment decision-making. (Chen et al., 2024).

Despite demonstrating outstanding competence in information retrieval and accessing specialist domain knowledge, current state-of-the-art models frequently struggle to capture critical financial events and conduct rigorous logic analysis when navigating the stochastic, adversarial, and non-stationary dynamics of live trading scenarios, which we termed the “Capabilities-Performance Paradox”. This disconnect highlights a fundamental reliability gap: the proficiency to explain financial concepts does not guarantee the agency required to analyze and execute profitable strategies under uncertainty.

Evaluation Gap: Existing financial benchmarks overwhelmingly assess static QA skills—fact retrieval, calculation, and well-defined problem-solving. They evaluate a model’s prowess as a financial analyst but fail to assess its competency as a portfolio manager: its ability to identify consequential signals in real-time noise, synthesize narratives from unfolding events, and maintain strategic discipline. This creates an evaluation opacity: we can measure whether an agent succeeds (via P&L) but cannot diagnose the underlying reasoning quality, thus obscuring whether success stems from sound logic or stochastic luck (Ding et al., 2024).

2.1. The U. S. Equity Trading Arena: A Live-Fire Market Competition To rigorously evaluate the investment logic of LLMs in market-aligned conditions, we instantiated a live U.S. equity trading arena. From November 19, 2025, multiple LLMbased agents, each initialized with $10,000 in simulated capital, competed autonomously to maximize risk-adjusted returns through discretionary hourly trading of U.S. stocks and ETFs.

To bridge this chasm, we shift the paradigm from evaluating what models know to assessing how they think. We introduce InvestLogicBench2026, the first benchmark constructed from the dynamic, articulated reasoning processes of successful market participants. It provides structured, validated traces that explicitly instantiate the P→E→R→D→O chain for each investment thesis. Our framework extracts and structures investment theses from experienced fund managers and key opinion leaders (KOLs) into explicit P→E→R→D→O chains, capturing market events, logical rationale, concrete actions, and market-verified outcomes.

The arena was designed to test the complete P→E→R→D→O chain under uncertainty. Agents maintained strategic continuity through access to their full Portfolio Report and Trading History. To simulate professional practice, they actively foraged for information via a dual-tool Perception-Action Loop—querying structured data with FinQuery and unstructured intelligence with Search—then synthesized these signals into executable trades with explicit reasoning. This forward-testing format eliminated retrospective bias, providing a pure assessment of an LLM’s ability to form and act on market logic in real time.

Our contributions are threefold: 1. Through a seven-week experimental study comparing live agent profitability with benchmark results, we empirically identify a pronounced capability-performance paradox in leading LLMs when applied to real-world market investment. 2. We present I NVEST L OGIC B ENCH 2026, the first benchmark for evaluating practical investment logic, featuring 201,247 structured decisions annotated with events, reasoning , decisions, actions, and marketverified outcomes from 151 real-world investors.

2.2. The Capabilities-Performance Paradox

3. We quantify a significant investment logic gap: leading LLMs achieve only 1.8/5 reasoning score against expert traces, revealing systematic failures in eventnoise discrimination and long-horizon consistency that explain their live trading underperformance.

As visualized in Figure 1 (Right), we contrasted the models’ Foundation Aptitude Scores—an average score derived from MMLU-Pro, SWE-bench, AIME 2025, and SimpleQA—against their realized financial returns. While scaling laws typically postulate a positive correlation between general reasoning capabilities and downstream task performance, our empirical data exhibits a distinct counterintuitive misalignment.

By establishing a diagnostic benchmark for market-aligned reasoning under uncertainty, I NVEST L OGIC B ENCH 2026 shifts the evaluation focus from financial knowledge recall to the critical ability to formulate and execute coherent investment theses.

Strikingly, models possessing superior foundation scores, such as Claude-Sonnet-4.5 (Score: 77.2) and GPT-5 (Score: 73.5), failed to translate their general reasoning dominance into effective market execution. Throughout the 7-week live trading window (Nov. 19, 2025 – Jan. 9, 2026), GPT-5 ended with a portfolio value of $9,901.1 (-1.0%), 2

I NVEST L OGIC B ENCH 2026

Total Account Value Comparative Analysis of LLMs: Foundation vs. Real-World Trading Performance

$20,000

$16,000

S&P 500 benchmark Cathie Wood Dave Mazza

80.0% DeepSeek-V3: $16362.1 (Score: 64.6)

$16,362.11 $15,173.11 $14,765.98

$14,000

$13,216.42 $12,582.84 $11,419.70 $11,252.74 $10,268.81

$12,000 $10,000

Real-World Cumulative Return (%)

$18,000

DeepSeek-V3 Qwen3-Max Gemini 2.5 Pro Claude Sonnet 4.5 GPT-5

60.0% Qwen3-Max: $14766.0 (Score: 76.1)

40.0%

Claude Sonnet 4.5: $12582.8 (Score: 77.2)

20.0% Gemini 2.5 Pro: $11419.7 (Score: 70.7) GPT-5: $10268.8 (Score: 73.5)

0.0%

$8,000 Nov 19

60.0

Dec 12

Jan 07

Jan 30

Feb 24

Mar 17

Apr 06

Apr 23

May 13

May 24

62.5

65.0

67.5

70.0

72.5

75.0

Foundation Aptitude Score

77.5

80.0

Figure 1. The “Capabilities-Performance Paradox” in financial domain. (Left) Cumulative equity curves between autonomous LLM agents and human expert proxies in a 7-week live U.S. stock trading competition. DeepSeek-V3 achieved the highest cumulative return, significantly outperforming the S&P 500 benchmark, while models like GPT-5 and Claude-Sonnet-4.5 demonstrated lackluster performance in seizing opportunities. Key market inflection points highlight periods where expert logic diverged from model behavior. (Right) Comparative analysis between Foundation Aptitude Scores and Real-World Cumulative Returns. It suggests that superior performance on general benchmarks fails to reliably transfer into financial analytical acumen or effective real-world trading execution.

while Claude-Sonnet-4.5 yielded negligible gains, closing at $10,072.8 (+0.7%). Both models failed to surpass the passive S&P 500 (SPY) benchmark, which returned approximately +4.3% over the same period, indicating that most models were paradoxically less effective in the live financial domain. In sharp contrast, as detailed in the cumulative equity curves in Figure 1 (Left), models with comparatively lower general evaluation scores demonstrated significantly stronger market adaptability. DeepSeek-V3 achieved the highest cumulative return of +14.2% ($11,420.3), significantly outperforming both the market baseline and its higherscoring counterparts.

superposition of fading liquidity support and reignited AI valuation fears triggered a sharp sell-off. • Fed Delivers Hawkish Cut (Dec. 9, 2025): The Fed delivered a widely expected 25 basis point cut with only one cut for 2026. The lack of a fresh positive catalyst triggered a “sell-the-news” dynamic with subsequent market adjustment. • CPI Surprise Cools Anxiety (Dec. 18, 2025): November CPI cooled to 2.7% (3% forecast), reigniting expectations for further rate cuts. This data thoroughly eliminated market tension, acting as the catalyst for a sustained subsequent rally.

This divergence suggests a critical transfer barrier, where current generalist metrics fail to capture the nuances of financial intelligence-specifically the ability to synthesize noisy, non-stationary signals into profitable decisions. This finding exposes a fundamental gap in current evaluation paradigms: the proficiency to solve static problems is decoupled from the agency required to execute rigorous financial logic under uncertainty.

• Post-Correction Rebound (Jan. 2, 2026): Year-end profit-taking and selling pressure were fully digested, forming a consensus around a soft landing. This renewed optimism launched a fresh rally to start the year. By decomposing the aforementioned turning points and the timeline, we observed that while the broader market (SPY) reacted dynamically to these pivotal events, models like GPT-5 and Claude-Sonnet-4.5 failed to correctly interpret the signal, resulting in flatlining performance.

2.3. Deconstructing the Divergence: Consensus Bias and Reasoning Fragmentation To diagnose the underlying mechanisms of the “CapabilitiesPerformance Paradox”, we conduct a retrospective analysis of the trading period and identify four pivotal market events that exerted the key influence on U.S. market.

Initially, the homogenous drawdowns observed during the “Hawkish Minutes & Nvidia Volatility” phase suggest a shared baseline risk sensitivity across all agents. However, a significant bifurcation emerged during the subsequent market rally. Claude-Sonnet-4.5 and GPT-5 executed a premature risk-off rotation—effectively liquidating positions during the transient shock—thereby failing to capture the ensuing liquidity-driven upside.

• Hawkish Minutes & Nvidia Volatility (Nov. 19–20, 2025): Hawkish FOMC minutes dampened rate cut hopes, coinciding with a reversal in Nvidia stock. The 3

I NVEST L OGIC B ENCH 2026

Specific inefficiencies were observed in GPT-5, which incurred a maximum drawdown exceeding the SPY benchmark following the “Fed Hawkish Cut”. Crucially, it failed to re-enter the market effectively during the “CPI Surprise” and “Post-Correction Rebound” phases. In contrast, despite Qwen3-Max suffering a comparable drawdown, it exhibited a rapid recovery dynamic, repairing losses immediately as market sentiment shifted. Notably, DeepSeek-V3 displayed advanced strategic timing effectively accumulating assets during the market cooling phase and realizing outsized returns during the recovery.

The deconstruction of experts’ unique decision-making chains provides critical evidence for understanding the performance deficiencies of LLMs in dynamic gaming scenarios. Deep analysis of this divergence reveals that the expert advantage stems from variant perception. When the Key Events occurred, the experts leveraged an independent reasoning framework to gather information and identify the mispricing, whereas LLMs lack this depth of analysis. Integrating human experts’ analysis logic is essential not just for comparison, but as a diagnostic reference system.

By comparing the decision traces of LLMs, two fundamental cognitive failures—Consensus Bias and Reasoning Fragmentation—are revealed as the drivers of the capabilitiesperformance paradox. Models like GPT-5 and ClaudeSonnet-4.5 tend to converge toward the statistical mean of their training data to maximize token likelihood; this results in merely reproducing priced-in market consensus rather than generating the variant perception necessary for alpha. Simultaneously, when confronted with a complex, non-stationary environment, these agents exhibited Reasoning Fragmentation. Overwhelmed by the volume of exogenous data, they reverted to rigid, myopic thinking patterns that failed to account for forward-looking dynamics, ultimately missing the pivotal rebound that drove the benchmark’s outperformance.

3. The I NVEST L OGIC B ENCH 2026 Dataset 3.1. Design and Construction I NVEST L OGIC B ENCH 2026 is constructed from the dynamic decision-making processes of influential real-world investment agents (e.g., major KOLs and fund managers), whose viewpoints are widely followed and accepted by substantial investor communities. Its design is built upon the E→R→D→O investment logic chain, where market Events (E) inform Reasoning (R) leading to a Decision (D) aimed at an profitable Outcome (O). This framework shifts evaluation from static question-answering to the assessment of active reasoning required for real-world investing. The benchmark is built through a four-stage pipeline (Fig. 2): 1) Data Preparation ingests multi-source data (e.g., video commentary, fund reports). 2) Agent Logic Processing employs a dual-channel analysis: (i) extracting explicit investment logic from the agent’s own verbalized theses and (ii) using our proprietary Stock Investment Logic (SIL) Engine to infer implicit logic patterns from an agent’s trading history, constructing a holistic behavioral profile. 3) RAG-Augmented Verification ensures fidelity by checking logic based on real-market data including general search engine, financial database, and event prediction markets where event contracts are traded (e.g., Polymarket, Kalshi). 4) Benchmark Assembly integrates verified decisions with agent profiles into a structured, queryable knowledge graph.

2.4. The Human Benchmark: Establishing Expert Logic Anchors The divergent performance of LLMs raises a critical question: How would human experts perform during this trading period? To incorporate expert logic, we introduced two professional investors: • Bill Miller: The founder of Miller Value Partners, renowned for his legendary 15-year streak of beating the S&P 500 (Emulate through the Miller Value Partners Appreciation ETF). • Cathie Wood: The founder and CEO of ARK Invest, who commits to find technologies where declining costs trigger exponential demand growth (Emulate through the ARK Innovation ETF).

3.2. Structured Schema Each data instance is a JSON object containing an agentProfile and a list of decisions, instantiating the E→R→D→O chain.

As illustrated in Figure 1 (Left), human expert portfolios demonstrated superior resilience and excess return generation. During critical shocks like the “Hawkish Cut” and the “Nvidia Volatility”, the experts exhibited counter-cyclical aggression. Unlike most investors and consensus-following LLMs, human experts can employ forward-looking reasoning to avoid drawdowns and accelerate decisively when opportunities arise. This allowed them to capture the full magnitude of the “Post-Correction Rebound” with pleasing cumulative returns (e.g., Miller at +10% and Wood at +8%).

Profile Fields characterize the agent: agentProfile (personalized investment style and preference), capital (scale), security (asset classes), and metadata (name, link). Decision Fields capture each logic-action triplet: • date, events: Decision timestamp and catalyst events.

4

I NVEST L OGIC B ENCH 2026

Figure 2. The two-stage construction pipeline of I NVEST L OGIC B ENCH 2026. Stage 1 creates a static investor profile by aggregating multiple data points from a single agent. Stage 2 performs per-decision annotation, which includes a quality control loop where logic is verified and potentially fed back to refine the profile.

3.5. Automated Evaluation via Judge LLM

• finLogic: The complete E→R→D chain (e.g., ’[Event] → [Thesis] → Action: [Action]’). • investDecision: The executable action (e.g., ’Long NVDA’). • investPeriod, endDate: Evaluation horizon and validation date. • outcome: The market-validated outcome (e.g., ’+5.2%’). • finLogicReview: System-generated post-mortem analysis.

A high-capacity LLM (e.g., GPT-5) serves as an automated judge, scoring outputs based on: - Fidelity: Similarity to ground-truth agent logic (Task 1). - Coherence & Plausibility: Logical soundness and market relevance (Tasks 2, 3). - Consistency: Adherence to initial style and logical through-line (Task 3).

4. Empirical Analysis: The Investment Logic Gap

3.3. Statistics and Tasks

We evaluate some leading LLMs on I NVEST L OG IC B ENCH 2026 using Task 2 (Logic & Thesis Generation), where models must synthesize an investment thesis from retrieved market events and an agent profile. Performance is measured along two axes: (1) Reasoning Quality—comprising Event Coherence (fidelity to input events) and Logical Plausibility (internal validity of the E→R→D chain), and (2) Decision Outcome—comprising Win Rate and simulated Total Profit over the investment horizon.

The inaugural version of I NVEST L OGIC B ENCH 2026 comprises 201,247 annotated investment theses sourced from 151 distinct financial KOLs (data details in Appendix A.1 and B). The statistical summary of the investor profiles, which underscores the dataset’s focus on real-world, eventdriven strategies, is presented in Table 1. 3.4. Benchmark Tasks We define three progressive tasks: 1. Logic Comprehension: Given agent text, extract and summarize the finLogic. (Evaluates understanding) 2. Logic & Thesis Generation: Given current events and an agentProfile, generate a plausible finLogic and decision. (Evaluates synthetic reasoning) 3. End-to-End Investment Simulation: According to the investment decisions, the outcome is evaluated with real-world market data. Both static and dynamic investment actions are evaluated depending on whether the account’s position (holdings) status is disclosed.

4.0.1. C ONDITIONAL S CORING F RAMEWORK The overall score synthesizes reasoning quality and decision outcome. The Decision Outcome Score is a weighted composite of two critical performance dimensions: the agent’s absolute profitability and its strategy win rate, each evaluated relative to the human expert benchmark. First, the Profit Score (Sprofit ) is calculated conditionally based on the agent’s total profit (Pa ) against the human expert benchmark (Ph ) and a neutral midpoint (Pm =

5

I NVEST L OGIC B ENCH 2026

Figure 3. A working example from the Stock Investment Logic (SIL) Engine. It demonstrates how a specific market event (here, an antitrust probe announcement by China’s SAMR targeting Nvidia on 2025-09-15) is temporally linked to a security (NVDA) and its subsequent price series. The engine structures and stores such Event-Security-Time tuples to form a queryable knowledge base for inferring recurring investment logic patterns and enriching agent profiles.

max(0, Ph )): Table 1. Statistical Summary of Agent Profiles in I NVEST L OG IC B ENCH 2026 Category Statistics Total Investment Experts Total Annotated Decisions Avg. Decisions per KOL Avg. Decisions per fund agent

  if Pa ≥ Ph 5.0,   P − P  a m  5.0 × , if Pm ≤ Pa < Ph Ph − P m Sprofit = P  5.0 × a , if 0 ≤ Pa < Pm   Pm    0.0, if Pa < 0

151 201,247 1332.8 2.5

Primary Style (agentProfile) - Event-Driven - Long-Term Value - Technical - Macro

102 (67.5%) 16 (10.6%) 8 (5.3%) 30 (19.9%)

Capital Scale (capital) - Below $1M - $1M - $10M - $10M - $100M - Above $100M

19 (12.6%) 41 (27.2%) 74 (49.0%) 17 (11.3%)

Primary Security Type (security) - Equity - Options - Futures/Commodities - ETF - Others/Combined

42 (27.8%) 16 (10.6%) 6 (4.0%) 15 (9.9%) 72 (47.7%)

Primary Sample Type - Event-driven decisions - Forward-looking event predictions Web-verifiable Prediction-market verified - Technical/quantitative analysis

94.0% 3.0% 2.9% 0.1% 3.0%

Second, the Win Rate Score (Swin ) measures the consistency of positive returns. The agent’s win rate (Wa ) is compared to the human’s (Wh ):  5.0, if Wa ≥ Wh Swin = Wa 5.0 × , if Wa < Wh Wh The final Decision Outcome Score is the weighted sum of these components, emphasizing absolute profit: Soutcome = 0.7 × Sprofit + 0.3 × Swin The Reasoning Quality Score remains the average of its Event Coherence (EC) and Logical Plausibility (LP), each on a 5-point scale: Sreasoning = (SEC + SLP )/2. The final Overall Score is the sum of the reasoning and outcome scores, providing a balanced assessment on a 0-10 6

I NVEST L OGIC B ENCH 2026

scale:

KOL/Models

Event

Reasoning

Decision

Outcome

U.S. government shutdown, ADP payroll report shows job losses, Fed cuts interest rates

Cooling labor market → Safe-haven demand Shutdown and weak ADP indicate uncertainty, supporting safe-haven assets

Buy: GLD 50%, TLT 50%(Short-term 20251231)

+3.53%

DeepSeek V3

LDP elects Sanae Takaichi as leader, Japan stocks rally on policy hopes, Fed cuts rates

Pro-stimulus fiscal → Equity boost + Yen depreciation Takaichi’s policies boost Japanese equities and depreciate Yen

Buy: TOPIX (513800¥ )

+2.44%

Claude-4.5

Japan’s LDP elected Takaichi, Apple increased iPhone 17 production, US new tariffs

Pro-stimulus shift → Bullish but tariff risk Loose monetary stance vs tariff fears lead to defensive positioning

Buy: TOPIX (Short-term 20251018)

+0.70%

GPT-5

Japan’s LDP elects Takaichi, Apple increases production, U.S. funding market stress

Stimulus hope → Conflated signals Ignored funding stress, mixed with unrelated Apple noise

Buy: TOPIX (Short-term 20251011)

-15.67%

Qwen3 Max

China-Bangladesh 50th anniversary, Jiaolong Arctic dive, Pinglu Canal progress

Infrastructure milestone → Logistics benefit Pinglu Canal 65% completion benefits Western Land-Sea Passage

Buy: 000582.SZ¥

+11.23%

Soverall = Sreasoning + Soutcome | {z } | {z } (0–5)

(0–5)

This composite metric ensures a robust and multidimensional evaluation of an LLM’s investment logic, equally valuing the quality of its reasoning process and the practical efficacy of its decisions.

Cathie Wood

4.0.2. E VALUATION R ESULTS As shown in Table 2, four of the five evaluated models exceeded the expert profitability benchmark, each earning a perfect outcome score. This indicates that LLMs can generate financially effective decisions. However, their reasoning quality scores show significant variance, with notable weaknesses in event grounding. The results confirm that achieving profitability does not guarantee logically coherent or well-explained investment theses, highlighting a key area for future improvement.

Figure 4. Decomposing the Investment Logic Gap: A comparative case study of investment reasoning across the P→E→R→D→O chain between a human expert (Cathie Wood) and LLMs instantiated with her agent profile. The analysis focuses on the market event cluster of October 4, 2025 with empty starting position, and the performance divergence is substantial: the expert returned +3.53%, the best LLM (Qwen) +11.23%, and the worst LLM (GPT) -15.67%, resulting in a 26.9% LLM performance gap. According to the table, LLMs exhibit significantly divergent investment logic and outcome. We use ¥ to present security code from Chinese financial market. Notably, the distinct geographical bias in asset recommendations across different LLMs reflects inherent market preferences shaped by their training data.

4.1. Case Study: Event-Driven Reasoning Failure We present a simulated case where an agent correctly predicted a rally in semiconductor stocks. The agent’s logic wove together nuanced observations from earnings call transcripts, industry capacity reports, and geopolitical news. When leading LLMs were provided the same raw event data, they either fixated on the most headline-grabbing news piece, produced a generic ”strong demand” summary, or generated internally contradictory statements. This case exemplifies the event-noise discrimination and integrative reasoning deficits of current models.

identified by I NVEST L OGIC B ENCH 2026: LLMs struggle to distinguish signal from noise and fail to maintain logical consistency across multiple data sources, explaining their underwhelming performance in live trading environments despite strong performance on static financial benchmarks.

Figure 4 illustrates a simulated event-driven investment scenario where Cathie Wood correctly predicted a rally in semiconductor stocks by integrating nuanced observations from multiple data sources. In contrast, LLMs exhibit systematic reasoning failures when provided the same raw event data.

5. Related Work LLM evaluation in finance has followed two main tracks: static benchmarks testing analytical knowledge, and dynamic agent frameworks measuring trading performance. However, a critical gap remains in evaluating the investment reasoning process itself—the logical synthesis of unstructured events into actionable theses. Recent studies expose temporal awareness failures and other reasoning limitations (Sehgal et al., 2026), underscoring the need for processcentric evaluation. I NVEST L OGIC B ENCH 2026 addresses this gap by introducing the first benchmark grounded in explicit P→E→R→D→O decision chains, enabling processdiagnostic assessment of investment logic.

Key Observations: (1) Human experts excel at integrative reasoning, weaving together earnings call transcripts, industry capacity reports, and geopolitical news into a coherent investment thesis. (2) LLMs display poor event-noise discrimination, either fixating on headline-grabbing news (e.g., government shutdown) or producing generic summaries lacking specificity. (3) The quantitative evaluation reveals a 25.7% gap in overall logic quality and a 2.2-point deficit in noise discrimination between human experts and the bestperforming LLM (Claude-4.5-Sonnet).

5.1. Static Benchmarks: Knowledge and Analysis Early work evaluates LLMs as analysts through static QA and numerical reasoning. FinQA (Chen et al., 2022) and TAT-QA (Zhu et al., 2021) require multi-step reasoning over

Implications: This case exemplifies the core limitations 7

I NVEST L OGIC B ENCH 2026 Table 2. Comprehensive Evaluation of LLMs on I NVEST L OGIC B ENCH 2026

Reasoning Quality

Model DeepSeek-V3 Qwen3-Max Claude-4.5 GPT-5

Decision Outcome

Event Coherence (/5)

Logical Plausibility (/5)

Win Rate

Total Profit

2.8 0.8 2.2 1.8

4.1 4.2 3.9 4.1

72% 51% 46% 40%

+16.7% +7.7% +5.4% +3.8%

Overall Score /10 7.8 5.7 5.3 3.1

Note: Logical Plausibility evaluates the coherence of the E→R→D chain. Total Profit is the average return of each outcome.

financial reports and hybrid tabular-textual content. Subsequent benchmarks expand scope: FinanceBench (Islam et al., 2023) tests open-domain QA on SEC filings, while BizBench (Koncel-Kedziorski et al., 2024) reframes business problems as quantitative tasks. Comprehensive suites like FinBen (Xie et al., 2024) aggregate these tasks.

soning under pressure (Sehgal et al., 2026) and advanced market analysis using LLM agents (Fatouros et al., 2025). Most pertinent is the Chain-of-Decision framework by Chen et al. (2024), which argues that modeling professional decisions requires distilling explicit intermediate reasoning steps between information intake and actions. This directly informs our P→E→R→D→O schema. However, their work focuses on report generation rather than the full investment cycle encompassing decision, action, and marketvalidated outcome.

The key limitation is retrospective analysis of given information. As noted by Ding et al. (2024), strong performance reflects calculation proficiency but not prospective decision-making under uncertainty—the core investor function. This limitation persists even in specialized models like BloombergGPT (Wu et al., 2023) and FinGPT (Yang et al., 2023), which remain predominantly evaluated on static benchmarks.

Unlike prior work measuring input knowledge or output performance, we evaluate the reasoning chain connecting them by (1) grounding evaluation in real expert decision trails, (2) assessing logical coherence independent of P&L, and (3) enabling style-conditioned diagnosis. This provides the substrate to quantify the investment logic gap and align models toward robust, interpretable, expert-aligned financial reasoning.

5.2. Dynamic Agents: Real-World Performance Recent work shifts from evaluating models as tools to assessing them as autonomous agents in simulated or live markets. This paradigm prioritizes outcome-based metrics (P&L). Frameworks like FinDABench (Liu et al., 2024) and XFinBench (Zhang et al., 2025) place LLM agents in complex financial problem-solving environments. Fully autonomous arenas such as AI-Trader (Fan et al., 2025) and DeepFund (Li et al., 2025) use real-time data with actual capital, providing high ecological validity. Architecturally, this space features multi-agent systems (Yu et al., 2024) and diverse agent-based modeling approaches (Gao et al., 2023).

6. Conclusion This paper identifies the absence of benchmarks for practical investment logic—the reasoning chain—as a fundamental obstacle to the development of investment-oriented agents. We introduce I NVEST L OGIC B ENCH 2026, the first benchmark constructed from the event-driven reasoning traces of real-world investors. It provides structured, validated data to evaluate an LLM’s ability to formulate and maintain coherent investment theses. Our analysis confirms a significant investment logic gap: state-of-the-art LLMs fail to match the nuanced, contextual reasoning of human experts, explaining their poor performance in live trading. This reproducible benchmark shifts the evaluation paradigm from passive knowledge recall to active reasoning. It lays the essential foundation for future work in training logic-aligned models and architecting personalized, market-aligned investment agents.

While crucial for end-to-end stress-testing, these approaches are fundamentally outcome-opaque. They indicate whether an agent profits but not why. Profitable trades may stem from sound logic or noise; losses may reflect poor strategy or black-swan events. This black-box nature hinders targeted improvement and obscures the root cause of the “Capabilities-Performance Paradox” (Ding et al., 2024). 5.3. Bridging the Gap: Process-Diagnostic Evaluation To address these limitations, recent work diagnoses specific reasoning failures. Look-ahead bias remains a critical challenge, causing performance decay under point-in-time constraints. Complementary studies examine temporal rea8

I NVEST L OGIC B ENCH 2026

7. Impact Statement

A new benchmark for financial question answering. arXiv preprint arXiv:2311.11944, 2023. URL https://ar xiv.org/abs/2311.11944.

This work introduces I NVEST L OGIC B ENCH 2026, a benchmark designed to advance machine learning by evaluating the core capability of dynamic, sequential reasoning under uncertainty—essential for developing reliable autonomous agents in finance.

Rik Koncel-Kedziorski, Michael Krumdick, Viet Lai, Varshini Reddy, Charles Lovering, and Chris Tanner. BizBench: A quantitative reasoning benchmark for business and finance. arXiv preprint arXiv:2311.06602, 2024. URL https://arxiv.org/abs/2311.06602.

By shifting evaluation from opaque performance metrics to the diagnosis of reasoning quality, the benchmark facilitates the development of trustworthy, interpretable, and logically sound AI systems. It serves as a foundational step toward transparent and accountable decision-making AI, with broader implications for improving financial decisionsupport tools.

Changlun Li, Yao Shi, Yuyu Luo, and Nan Tang. Will LLMs be professional at fund investment? DeepFund: A live arena perspective. arXiv preprint arXiv:2503.18313, 2025. URL https://arxiv.org/abs/2503.1 8313. Shu Liu, Shangqing Zhao, Chenghao Jia, Xinlin Zhuang, Zhaoguang Long, Jie Zhou, Aimin Zhou, Man Lan, and Yang Chong. FinDABench: Benchmarking financial data analysis ability of large language models. In Proceedings of the International Conference on Computational Linguistics, 2024. URL https://api.semanticsc holar.org/CorpusID:266844370.

References Chung-Chi Chen, Hiroya Takamura, Ichiro Kobayashi, and Yusuke Miyao. Distilling analysis from generative models for investment decisions. arXiv preprint arXiv:2410.07225, 2024. URL https://arxiv.or g/abs/2410.07225. Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, and William Yang Wang. FinQA: A dataset of numerical reasoning over financial data. arXiv preprint arXiv:2109.00122, 2022. URL https://arxiv.or g/abs/2109.00122.

Neil K. R. Sehgal, Sharath Chandra Guntuku, and Lyle Ungar. Real-time deadlines reveal temporal awareness failures in LLM strategic dialogues. arXiv preprint arXiv:2601.13206, 2026. URL https://arxiv.or g/abs/2601.13206. Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. BloombergGPT: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023. URL https://arxiv.or g/abs/2303.17564.

Han Ding, Yinheng Li, Junhao Wang, and Hang Chen. Large language model agent in financial trading: A survey. arXiv preprint arXiv:2408.06361, 2024. URL https://arxiv.org/abs/2408.06361. Tianyu Fan, Yuhao Yang, Yangqin Jiang, Yifei Zhang, Yuxuan Chen, and Chao Huang. AI-trader: Benchmarking autonomous agents in real-time financial markets. arXiv preprint arXiv:2512.10971, 2025. URL https://arxiv.org/abs/2512.10971.

Qianqian Xie, Weiran Wang, Zhenyu Chen, Ruoyu Xiang, Xiao Zhang, Yansong Deng, et al. Finben: A holistic financial benchmark for large language models. In Advances in Neural Information Processing Systems 38 (NeurIPS 2024), 2024. doi: 10.52202/079017-3033. URL https://proceedings.neurips.cc/p aper_files/paper/2024/hash/adb1d9fa8 be4576d28703b396b82ba1b-Abstract-Dat asets_and_Benchmarks_Track.html.

George Fatouros, Kostas Metaxas, John Soldatos, and Manos Karathanassis. MarketSenseAI 2.0: Enhancing stock analysis through LLM agents. arXiv preprint arXiv:2502.00415, 2025. URL https://arxiv.or g/abs/2502.00415.

Hongyang Yang, Xiao-Yang Liu, and Christina Dan Wang. FinGPT: Open-source financial large language models. arXiv preprint arXiv:2306.06031, 2023. URL https: //arxiv.org/abs/2306.06031.

Chen Gao, Xiaochong Lan, Nian Li, Yuan Yuan, Jingtao Ding, Zhilun Zhou, Fengli Xu, and Yong Li. Large language models empowered agent-based modeling and simulation: A survey and perspectives. arXiv preprint arXiv:2312.11970, 2023. URL https://arxiv.or g/abs/2312.11970.

Yangyang Yu, Zhiyuan Yao, Haohang Li, Zhiyang Deng, Yuechen Jiang, Yupeng Cao, Zhi Chen, Jordan W. Suchow, Zhenyu Cui, Rong Liu, Zhaozhuo Xu, Koduvayur Subbalakshmi, Guojun Xiong, Yueru He, Jimin Huang, Dong Li, and Qianqian Xie. FinCon: A synthesized LLM

Pranab Islam, Anand Kannappan, Douwe Kiela, Rebecca Qian, Nino Scherrer, and Bertie Vidgen. FinanceBench: 9

I NVEST L OGIC B ENCH 2026

multi-agent system with conceptual verbal reinforcement for enhanced financial decision making. In Advances in Neural Information Processing Systems, volume 38, 2024. Zhihan Zhang, Yixin Cao, and Lizi Liao. XFinBench: Benchmarking LLMs in complex financial problem solving and reasoning. arXiv preprint arXiv:2508.15861, 2025. URL https://arxiv.org/abs/2508.1 5861. Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance. arXiv preprint arXiv:2105.07624, 2021. URL https://arxiv.or g/abs/2105.07624.

10

I NVEST L OGIC B ENCH 2026

A. Appendix A.1. KOL Sample Construction and Measurement Notes This appendix documents a curated sample of YouTube-based key opinion leaders (KOLs) for descriptive comparison of investment-logic-oriented content. Channels are selected based on whether their content primarily articulates generalizable decision frameworks (e.g., macro–industry–firm reasoning pipelines; valuation, risk management, and portfolio construction principles; cycle and sentiment indicators; and rule-based trading systems), as opposed to predominantly event-driven or news-reactive commentary. The resulting sample covers multiple market segments (equities/funds/macro, options and trading education, and crypto/Web3), enabling cross-domain characterization of audience scale and content supply. As shown in Table 3, Subscribers is used as a coarse proxy for audience size, and Videos reflects cumulative content output and, to a limited extent, channel maturity. Both measures are time-varying; thus, the table should be interpreted as a descriptive snapshot rather than a fixed ground-truth record. For subsequent quantitative analyses, we recommend re-scraping and freezing all statistics within a single, pre-specified time window to ensure comparability.

Table 3: Statistics of Selected KOL Creators Name

Description

Subscribers

Videos

Cathie Wood

Investment strategist, founder of ARK Invest

619K

743

Zip Trader (Charlie)

Day trading and swing trading education

857

819

Stock Moe

Stock and crypto trading strategies

714K

897

Data Dispatch

Market data and analytics insights

28.4K

611

Paul Thomas Investing

Penny stocks and growth investment analysis

57.4K

478

Stas Talks Stocks

Stock market discussion and analysis

61.6K

4283

Lark Davis

Crypto investment strategies

638K

4936

Miles Deutscher

Cryptocurrency analysis and portfolio strategies

224K

2284

Financial Education

Investing education

911K

2800

Everything Money

Investing

376K

3019

Joseph Carlson Show

Investing

499K

449

Learn to Invest

Investing basics

289K

597

Sven Carlin

Value investing

263K

2176

New Money

Finance commentary

1090K

415

Tom Nash

Market analysis

572K

880

Chicken Genius Singapore

Singapore investing

249K

249

Ticker Symbol YOU

Investing

572K

382

Jimmy Dividend Investing

Dividend investing

289K

597

TraderTV Live

Professional live market trading sessions

550K

6315

The Plain Bagel

Animated financial education

1140K

263

The Compound

Casual market commentary

223K

1885

Seeking Alpha

Fundamental stock analysis

32.7K

947

Patrick Boyle

Academic finance perspective

1100K

460

Option Alpha

Systematic options trading education

296K

862

NaNa Talks US Stocks

Chinese-language US stock analysis

301K

1573

Morningstar

Fund analysis and long-term strategies

129K

2956

Money or Life

Financial freedom philosophy

71.8K

217

11

I NVEST L OGIC B ENCH 2026 Name

Description

Subscribers

Videos

Investopedia

Financial dictionary and concepts

298K

545

Investor’s Business Daily

CAN SLIM investing system

169K

4961

Hedgeye

Macroeconomic analysis

78.3K

4131

Coin Bureau

Comprehensive cryptocurrency education

2730K

1773

The Defiant

DeFi and Web3 news and analysis

131K

1346

Unchained

Blockchain and crypto industry insights

6770K

1649

Crypto Casey

Beginner-friendly cryptocurrency education

678K

875

The Maverick of Wall Street

Alternative investment strategies

146K

1686

Larry Jones

Stocks and crypto analysis for entrepreneurs

691K

2154

Meet Kevin

Real estate and financial independence education

2050K

6095

Graham Stephan

Personal finance and real estate investing

5140K

1402

Mark Moss

Economic trends and investment strategies

803K

1365

Real Vision

In-depth financial markets analysis

89.7K

1210

Garys Economics

Economic analysis and market commentary

1520K

369

Let’s Talk Money!

Personal finance education and tips

738K

1199

ClearValue Tax

Tax strategies and financial planning

2820K

151

Kitco NEWS

Precious metals market news and analysis

750K

4818

12

I NVEST L OGIC B ENCH 2026

B. Appendix: KOL Resources B.1. KOL Channel URLs and Reproducibility To support reproducibility and facilitate follow-up analyses, Table 4 provides the canonical channel URLs for the curated KOL sample reported in Table 3. These links enable consistent access to each source and allow downstream data collection (e.g., channel metadata, upload cadence, topical coverage, and representative playlists) under a fixed protocol. In empirical settings where platform statistics and content inventories evolve over time, anchoring the sample with persistent URLs is essential for re-scraping within a pre-specified time window and ensuring that subsequent measurements remain comparable across runs.

Table 4: List of KOL Channels and URLs KOL Name

Channel URL

Cathie Wood

https://www.youtube.com/@ARKInvest2015

Financial Education

https://www.youtube.com/@FinancialEducation

Everything Money

https://www.youtube.com/@EverythingMoney

Joseph Carlson Show

https://www.youtube.com/@JosephCarlsonShow

Learn to Invest

https://www.youtube.com/@LearnToInvest

Sven Carlin

https://www.youtube.com/c/InvestwithSvenCarlinPhD

New Money

https://www.youtube.com/@NewMoneyYouTube

Tom Nash

https://www.youtube.com/@TomNashTV

Chicken Genius Singapore

https://www.youtube.com/@ChickenGeniusSingapore

Ticker Symbol YOU

https://www.youtube.com/@TickerSymbolYOU

Jimmy Dividend Investing

https://www.youtube.com/@JimmyDividend

TraderTV Live

https://www.youtube.com/c/TraderTVLive

The Plain Bagel

https://www.youtube.com/c/ThePlainBagel

The Compound

https://www.youtube.com/@TheCompoundNews

Seeking Alpha

https://www.youtube.com/c/SeekingAlpha

Patrick Boyle

https://www.youtube.com/c/PatrickBoyleOnFinance

Option Alpha

https://www.youtube.com/c/OptionAlpha

NaNa Talks US Stocks

https://www.youtube.com/@NaNaShuoMeiGu

Morningstar

https://www.youtube.com/@morningstar

Money or Life

https://www.youtube.com/@Money_or_Life

Investopedia

https://www.youtube.com/c/Investopedia

Investor’s Business Daily

https://www.youtube.com/c/InvestorsBusinessDaily

Hedgeye

https://www.youtube.com/user/Hedgeye

Zip Trader (Charlie)

https://www.youtube.com/c/ziptrader

Stock Moe

https://www.youtube.com/channel/UCoMzWLaPjDJBbipihD694 pQ

Data Dispatch

https://www.youtube.com/@DataDispatch

Paul Thomas Investing

https://www.youtube.com/c/PaulThomasInvesting

Stas Talks Stocks

https://www.youtube.com/@StasTalksStocks

Lark Davis

https://www.youtube.com/@TheCryptoLark

Miles Deutscher

https://www.youtube.com/@MilesDeutscher

Coin Bureau

https://www.youtube.com/@CoinBureau

The Defiant

https://www.youtube.com/@TheDefiant

13

I NVEST L OGIC B ENCH 2026 KOL Name

Channel URL

Unchained

https://www.youtube.com/@UnchainedOff

Crypto Casey

https://www.youtube.com/@CryptoCasey

The Maverick of Wall Street

https://www.youtube.com/@TheMaverickofWallStreet

Larry Jones

https://www.youtube.com/c/STOCKUPwithLarryJones

Meet Kevin

https://www.youtube.com/@MeetKevin

Graham Stephan

https://www.youtube.com/@GrahamStephan

Mark Moss

https://www.youtube.com/@1MarkMoss

Real Vision

https://www.youtube.com/@RealVisionFinance

Garys Economics

https://www.youtube.com/@garyseconomics

Let’s Talk Money!

https://www.youtube.com/@josephhogue

ClearValue Tax

https://www.youtube.com/@clearvaluetax9382

Kitco NEWS

https://www.youtube.com/@kitco

14

I NVEST L OGIC B ENCH 2026

C. Appendix:KOL Fund Reference Tabless C.1. Institutions, Fund Managers, and Morningstar Ratings Table 5 reports a compact snapshot of the fund universe by listing each record’s institution (fund firm), fund manager(s), fund name, and the Morningstar overall star rating. We include the star rating as a standardized, coarse descriptor of historical risk-adjusted performance within a fund’s peer group, and use it strictly for descriptive comparison rather than causal interpretation. Because manager assignments and ratings may change over time, the table should be interpreted as a snapshot tied to the data extraction window.

Table 5: Institutions, Fund Managers, Fund Names, and Morningstar Overall Ratings Institution (Firm)

Fund Manager(s)

Fund Name

Aegis Alger Alger Allspring Global Investments Allspring Global Investments Gotham American Century Investments BNY Mellon Ave Maria Mutual Funds VALIC

Scott L. Barbee Patrick Kelly, Ankur Crawford Gregory S. Adams Justin P. Carr, Robert M. Wicentowski

Aegis Value I Alger Focus Equity Y Alger Growth & Income Z Allspring Disciplined US Core R6

WesMark Wasatch Hood River Capital Management GMO Leavell SEI T. Rowe Price T. Rowe Price T. Rowe Price T. Rowe Price BlackRock BlackRock BlackRock BlackRock Franklin Templeton Investments Copley Hotchkis & Wiley Independent Franchise Partners Fidelity Investments Fidelity Investments Fidelity Investments

MS Rating 5 5 4 5

John R. Campbell, Vincent Fioramonti Allspring Large Cap Core R6

5

Joel Greenblatt, Robert Goldstein Gotham Index Plus Institutional Keith Lee, Jeffrey R. Bourke, Tong Li American Century U.S. Equity Focus I

5 3

Peter D. Goslin George P. Schwartz, Timothy S. Schwartz Jeffrey D. Parker, Brian Demain, Philip Cody White Douglas C. Robertson Garrett S. Lamb, Sean A. Schaub Michael J. Delf, John H. Beck

BNY Mellon Appreciation Instl Ave Maria Rising Dividend I

5 5

VALIC Company I International Growth

5

WesMark Small Cap I Wasatch Small Cap Value Inst Hood River Small-Cap Growth I

4 4 5

Ben Inker, Jeremy Grantham Thomas A. Rowe, Ryan T. Leavell SEI Investments Management Corporation Larry J. Puglia Justin P. White Brian W. Berghuis Brian C. Rogers Robert R. Shillito, Michael C. Clemente Tony DeSpirito, John M. H. Lawler, Kevin B. Winkler Tony DeSpirito, John M. H. Lawler, Kevin B. Winkler Tony DeSpirito, John M. H. Lawler, Kevin B. Winkler David M. Winters, Todd Brighton

GMO Quality Fund III Leavell Investment Counsel Mid Cap Growth SEI Large Cap Growth Instl (SIMT)

5 5 4

T. Rowe Price Blue Chip Growth I T. Rowe Price Growth Stock I T. Rowe Price Mid-Cap Growth I T. Rowe Price Equity Income I BlackRock Advantage Small Cap Core Instl

4 4 4 4 4

BlackRock Global Equity Instl

4

BlackRock Mid-Cap Growth Equity Instl

4

BlackRock Small Cap Growth Equity Instl

4

Franklin Small Cap Growth Adv

4

Jeffrey M. Fischer, Benjamin W. Hall Copley Focus Equity Institutional Daniel G. Hotchkis, Shannon P. Kress Hotchkis & Wiley Value Opportunities I Andrew L. Moffett, David W. Potter Independent Franchise Partners US Equity I

4 4 4

Will Danoff Joel P. Tillinghast Geode Capital Management

5 4 4

Fidelity Contrafund K6 Fidelity Low-Priced Stock K6 Fidelity Total Market Index

15

I NVEST L OGIC B ENCH 2026 Institution (Firm)

Fund Manager(s)

Fund Name

Dodge & Cox Dodge & Cox ClearBridge Investments ClearBridge Investments Cohen & Steers Columbia Threadneedle Investments Dimensional Fund Advisors Dimensional Fund Advisors Dimensional Fund Advisors Invesco J.P. Morgan J.P. Morgan

Dodge & Cox Dodge & Cox David C. Kupfer, Peter W. Fondren Robert E. Taylor, John D. Freeman Joseph M. Harvey, Jason T. Yablon James P. Tierney

Dodge & Cox Income Dodge & Cox Stock ClearBridge Large Cap Growth IS ClearBridge Large Cap Value IS Cohen & Steers Realty Shares Instl Columbia Seligman Global Technology Instl 2

4 4 4 4 3 5

Dimensional Fund Advisors

DFA Emerging Markets Core Equity I

4

Dimensional Fund Advisors

DFA International Core Equity I

4

Dimensional Fund Advisors

DFA US Core Equity 1 I

4

Amit D. Kumar Clare H. Hart, Michael D. Marini Jonathan D. Simon, Jennifer L. Williams Kevin D. Clark, Joseph Flaherty

Invesco Developing Markets Instl JPMorgan Equity Income R6 JPMorgan Large Cap Growth R6

3 4 4

MFS Growth R6

4

Robert M. Zito, James H. LeBoeuf

MFS Value R6

4

Scott A. Mather, Esteban J. Burbano William R. Bowen, Allen E. Jackson Schwab Equity Ratings David Samra, Kiki P. Pillai Benjamin Herrick, David Samra, Kiki P. Pillai Ronald Baron, Michael J. Lippert Ronald Baron, Michael T. Sheehy, Neal E. Clark Vanguard Equity Index Group Donald M. Kilbride, Michael D. Reckmeyer Vanguard Equity Index Group Vanguard Equity Index Group Vanguard Equity Index Group Vanguard Equity Index Group Vanguard Equity Index Group Vanguard Equity Index Group John T. Keogh, Thomas J. Martin John T. Keogh, Thomas J. Martin John T. Keogh, Thomas J. Martin John T. Keogh, Thomas J. Martin Vanguard Equity Index Group

PIMCO Total Return Instl Principal MidCap R6 Schwab S&P 500 Index Artisan Developing World Instl Artisan International Value Instl

3 4 4 4 4

Baron Partners Instl Baron Small Cap Instl

4 4

Vanguard 500 Index Admiral Vanguard Dividend Growth Investor

4 4

Vanguard Extended Market Index Admiral Vanguard FTSE All-World ex-US Index Admiral Vanguard Growth Index Admiral Vanguard Small-Cap Index Admiral Vanguard Total Stock Market Index Admiral Vanguard Value Index Admiral Vanguard Wellington Investor Vanguard Windsor II Investor Vanguard Windsor Investor Vanguard Wellesley Income Investor Vanguard World Stock Index Investor

4 4 4 4 4 4 4 4 4 4 4

MFS Investment Management MFS Investment Management PIMCO Principal Charles Schwab Artisan Partners Artisan Partners Baron Capital Baron Capital Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard Vanguard

16

MS Rating

I NVEST L OGIC B ENCH 2026

C.2. Fund Managers and Trailing Return Metrics Table 6 reports a compact summary of each fund’s manager(s) and trailing performance metrics. We use TR 1Y to denote the trailing one-year total return, and TR 3Y / TR 5Y to denote the annualized total returns over the trailing three- and five-year horizons, respectively. These values are time-varying and should be interpreted as a descriptive snapshot tied to the data extraction window.

Table 6: Fund Managers and Trailing Returns Fund Name

Manager(s)

Aegis Value I Alger Focus Equity Y Alger Growth & Income Z Allspring Disciplined US Core R6 Allspring Large Cap Core R6 Allspring Special Small Cap Value R6 American Funds Washington Mutual R6 Artisan Developing World Instl Artisan International Value Instl Baron Partners Instl Baron Small Cap Instl BlackRock Advantage Small Cap Core Instl BlackRock Global Equity Instl BlackRock Mid-Cap Growth Equity Instl BlackRock Small Cap Growth Equity Instl ClearBridge Large Cap Growth IS ClearBridge Large Cap Value IS Cohen & Steers Realty Shares Instl Columbia Seligman Global Technology Instl 2 DFA Emerging Markets Core Equity I DFA International Core Equity I DFA US Core Equity 1 I Dodge & Cox Income Dodge & Cox Stock Fidelity Contrafund K6 Fidelity Low-Priced Stock K6 Fidelity Total Market Index Franklin Small Cap Growth Adv Invesco Developing Markets Instl JPMorgan Equity Income R6 JPMorgan Large Cap Growth R6 MFS Growth R6 MFS Value R6 PIMCO Total Return Instl Principal MidCap R6 Schwab S&P 500 Index T. Rowe Price Blue Chip Growth I T. Rowe Price Equity Income I T. Rowe Price Growth Stock I T. Rowe Price Mid-Cap Growth I Vanguard 500 Index Admiral Vanguard Dividend Growth Investor Vanguard Extended Market Index Admiral Vanguard FTSE All-World ex-US Index Admiral Vanguard Growth Index Admiral Vanguard Small-Cap Index Admiral

S. Barbee P. Kelly, A. Crawford G. Adams J. Carr, R. Wicentowski J. Campbell, V. Fioramonti J. Campbell, V. Fioramonti M. Pacheco, K. Martell, B. Reick D. Samra, K. Pillai B. Herrick, D. Samra, K. Pillai R. Baron, M. Lippert R. Baron, M. Sheehy, N. Clark R. Shillito, M. Clemente T. DeSpirito, J. Lawler, K. Winkler T. DeSpirito, J. Lawler, K. Winkler T. DeSpirito, J. Lawler, K. Winkler D. Kupfer, P. Fondren R. Taylor, J. Freeman J. Harvey, J. Yablon J. Tierney Dimensional (DFA) Dimensional (DFA) Dimensional (DFA) Dodge & Cox Dodge & Cox W. Danoff J. Tillinghast Geode D. Winters, T. Brighton A. Kumar C. Hart, M. Marini J. Simon, J. Williams K. Clark, J. Flaherty R. Zito, J. LeBoeuf S. Mather, E. Burbano W. Bowen, A. Jackson Schwab L. Puglia B. Rogers J. White B. Berghuis Vanguard Index D. Kilbride, M. Reckmeyer Vanguard Index Vanguard Index Vanguard Index Vanguard Index

17

1Y

3Y

5Y

29.37% 51.38% 14.82% 18.56% 19.50% 16.06% 10.84% 10.61% 9.68% 37.09% 17.52% 17.89% 24.73% 33.92% 26.72% 35.58% 11.51% 6.58% 46.84% 8.96% 13.62% 18.48% 1.67% 9.53% 27.86% 17.11% 18.25% 28.73% 10.46% 10.24% 36.55% 30.64% 10.53% 1.46% 25.36% 18.34% 37.78% 11.05% 33.65% 33.17% 18.35% 8.65% 19.55% 13.11% 25.75% 15.17%

16.90% 32.67% 16.25% 18.61% 19.34% 12.17% 10.86% 5.73% 6.28% 17.08% 10.71% 10.62% 12.81% 15.86% 10.76% 15.01% 10.48% 1.68% 19.32% 5.62% 9.40% 11.49% 2.57% 9.88% 13.63% 9.69% 11.24% 9.78% 4.55% 10.62% 14.63% 12.54% 9.85% 2.38% 11.83% 10.99% 14.29% 10.36% 12.77% 12.39% 11.00% 9.91% 8.72% 9.43% 13.03% 7.25%

24.01% 19.60% 16.23% 17.29% 17.89% 12.77% 10.71% 7.13% 10.21% 19.25% 12.45% 11.57% 13.42% 15.05% 10.50% 13.90% 10.64% 4.91% 22.11% 6.38% 11.17% 12.38% 1.83% 10.58% 14.26% 11.32% 12.28% 10.52% 5.59% 11.31% 14.26% 13.26% 10.79% 1.63% 13.10% 12.17% 14.87% 10.77% 13.26% 13.07% 12.20% 10.87% 10.15% 11.08% 14.19% 8.44%

I NVEST L OGIC B ENCH 2026 Fund Name

Manager(s)

Vanguard Total Stock Market Index Admiral Vanguard Value Index Admiral Vanguard Wellington Investor Vanguard Windsor II Investor Vanguard Windsor Investor Vanguard Wellesley Income Investor Vanguard World Stock Index Investor

Vanguard Index Vanguard Index J. Keogh, T. Martin J. Keogh, T. Martin J. Keogh, T. Martin J. Keogh, T. Martin Vanguard Index

18

1Y

3Y

5Y

18.25% 10.19% 10.27% 9.64% 8.67% 6.31% 20.64%

11.25% 9.79% 7.60% 8.53% 8.42% 5.32% 12.12%

12.29% 10.90% 8.60% 9.84% 9.52% 6.19% 12.83%

Related documents

• FinRegQA-EU: Corruption-Based Preference Data for Grounded EU Financial Regulatory Question Answering#1132057
• Kronos: A Foundation Model for the Language of Financial Markets#130734
• Taste-based Investing, Government Policies and Competition in Financial Intermediation#130845
• AI Knows When It's Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models#187357
• Compute Where it Counts: Self Optimizing Language Models#175253