InvestorBench: Financial Benchmarking Framework

• InvestorBench is a comprehensive benchmarking suite that rigorously evaluates financial performance and decision-making using statistical, algorithmic, and qualitative metrics.
• It integrates quantitative manager assessments, risk attribution via multi-factor models, and long-only benchmark construction to compare human and AI-driven investment strategies.
• The framework extends to LLM-based agent evaluation, deep research report assessment, and simulation of ESG investment dilemmas for reproducible, multi-faceted analysis.

InvestorBench is a set of principled, extensible benchmarking frameworks designed for evaluating performance, skill, and decision-making in financial contexts. Its scope now spans quantitative benchmarking of active managers, explicit long-only risk-aware portfolio construction, evaluation of LLM-based financial agents, deep-research report assessment, and simulation-based investigation of ESG investment dilemmas. The unifying aim is to provide rigorous, reproducible baselines—statistical, algorithmic, and qualitative—against which both human and AI competitors can be judged on meaningful financial, operational, and explanatory metrics.

1. Quantitative Manager Benchmarking and Portfolio Evaluation

InvestorBench implements a comprehensive evaluation toolkit for active managers, as formalized by Schneider et al. The workflow integrates return-path calculation, metric extraction, factor-model analysis, and systematic comparison to both passive and random baselines. Six core metrics are compulsory:

• Information Ratio (IR)—the core competition statistic, based on log returns.
• Sharpe Ratio (SR)—annualized mean return over volatility.
• Maximum Drawdown (MDD)—maximum observed capital loss from peak to trough.
• Calmar Ratio (CR)—annualized return scaled by MDD.
• Ulcer Index (UI)/Ulcer Performance Index (UPI)—captures drawdown depth and persistence.
• Upside/Downside Volatility—separates positive vs. negative daily return variances.

InvestorBench mandates benchmarking against a well-defined universe:

1. Conventional long-only indices (S&P 500, Nasdaq-100, CTA, crypto, market-neutral, equity long-short, event-driven).
2. An equal-weighted portfolio (the “competition universe”).
3. Random long-only, short-only, and long-short portfolios (“Malkiel’s darts”, 1,000 samples each).

Empirical analysis reveals that randomly constructed long-short portfolios can yield extreme Sharpe ratios, matching those of the winning competition entries, while also exhibiting better drawdown control. In contrast, most managers cannot reliably outperform random long-only/short-only portfolios, but many do outperform these on a short-only basis (Schneider et al., 2024).

2. Risk Attribution and Factor Models

Central to InvestorBench’s method is the isolation of genuine skill (“alpha”) via multi-factor regression:

$$r_{i,t} - r_{f,t} = \alpha_i + \sum_{j=1}^J \beta_{i,j}\bigl(f_{j,t}-r_{f,t}\bigr) + \epsilon_{i,t},$$

where $f_{j,t}$ spans standard market and alternative indices, and $\alpha_i$ (annualized) measures skill while $\beta_{i,j}$ encodes systematic risk exposure. Residual tests (Durbin–Watson, Prais–Winsten) are required to address autocorrelation, with joint alpha significance determined via Gibbons–Ross–Shanken (GRS) tests.

The Appraisal Ratio (AR) is employed for risk-adjusted performance:

$$\mathrm{AR}_i = \frac{\alpha_i^{\text{ann}}}{\sigma(\epsilon_{i,t})\sqrt{252}}$$

where $\alpha_i^{\text{ann}} = \alpha_i \times 238$, paralleling a z-statistic for economically meaningful skill.

Empirically, only a small minority of competitors display statistically significant positive alpha, with most exhibiting negative or statistically insignificant alpha in both single- and multi-factor models (Schneider et al., 2024).

3. Momentum, Selection Heuristics, and Incentive Alignment

InvestorBench introduces endogenous manager-selection rules capturing “performance-chasing” and mean-reverting effects. Two canonical strategies are evaluated monthly:

• Superstars: Equal-weight top-10 managers based on prior month return; exhibits strong mean reversion (NAV $\approx \$87.46 $$).</li> <li><strong>Superlosers</strong>: Equal-weight bottom-10; evidences momentum among recent underperformers (NAV$$\approx \$106.95 $$).</li> </ul> <p>Repeated empirical analysis on competition data indicates that “chasing last month’s winners” is not a robust long-term allocator policy, while short-term momentum among laggards can yield transitory outperformance (<a href="/papers/2406.19105" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Schneider et al., 2024</a>).</p> <p>InvestorBench is explicit about incentive misalignments: competition/tournament structures (one-off, high-variance, tail-risk seeking) contrast sharply with the professional asset allocation environment, which is governed by fee-maximization, longer time horizons, and aversion to extreme downside (<a href="/papers/2406.19105" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Schneider et al., 2024</a>).</p> <h2 class='paper-heading' id='long-only-benchmark-construction-algorithmic-foundation'>4. Long-Only Benchmark Construction: Algorithmic Foundation</h2> <p>The InvestorBench algorithm for equity universe benchmarks eschews <a href="https://www.emergentmind.com/topics/principal-component-analysis-pca" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">principal component analysis</a> and negative weights, enforcing strictly positive allocations via a multifactor “Russian-doll” risk model. The construction proceeds as follows (<a href="/papers/1807.09919" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">Kakushadze et al., 2018</a>):</p> <ol> <li><strong>Target Betas and Specific Variance</strong>: For$$N$stocks, with positive target betas$\beta_i$and specific variances$f_{j,t}$0, preliminary weights are $f_{j,t}$1.
• Multilevel Industry Clustering: Covariance is modeled in $f_{j,t}$2 nested levels (e.g., sector $f_{j,t}$3 industry $f_{j,t}$4 stock).
• Closed-Form Solution: At each level, one-factor fits yield scalar normalizations, ensuring all weights remain positive and functionally interpretable.
• No PCA, No Iterative Bounds: All calculations are direct; the method avoids the instability and sign ambiguity of PCA-based portfolios.
• Dollar-Neutral Outperformance Sleeve: Overlays can be constructed atop the long-only portfolio by adding a dollar-neutral alpha vector $f_{j,t}$5, solved via quadratic programming under risk constraints.

Empirical backtesting demonstrates InvestorBench systematically corrects for volatility skews and sector imbalances found in raw cap-weighted or PCA-1 benchmarks, offering both economic clarity and out-of-sample stability (Kakushadze et al., 2018).

5. LLM-based Financial Decision-Making Benchmarks

InvestorBench extends to LLM-based agents and multi-modal decision tasks, introducing a unified evaluation suite for natural-language and structured-environment financial agents (Li et al., 2024). The framework models trading as a partially observable Markov decision process (POMDP) with infinite horizon and discount factor $f_{j,t}$6. For each supported environment (StockEnv, CryptoEnv, ETFEnv)—backed by timestamp-aligned, open-source datasets—the core modules are:

• Brain/Backbone: LLM core (GPT-4, Llama3, Qwen, etc.)
• Perception: Processes raw OHLCV, news, reports into prompts.
• Profile: Exposes agent “character” and risk preferences.
• Memory: Multi-scale memory layers (shallow, intermediate, deep).
• Action: Emits $f_{j,t}$7 with rationale and key memory context.

The evaluation protocol applies established finance metrics: cumulative return (CR), Sharpe ratio (SR), annualized volatility (AV), and maximum drawdown (MDD). Each model is run five times over the public Gym-style environment, reporting the median-SR epoch for reproducibility.

Key findings:

• Proprietary LLMs (GPT-4/4o/o1) dominate on both CR and SR.
• Performance in cryptocurrencies is highly model-size dependent; smaller LLMs cannot reliably outperform baseline Buy & Hold.
• Domain fine-tuning does not guarantee outperformance in trading tasks (Li et al., 2024).

6. Qualitative and Explanatory Benchmarking of Research Agents

InvestorBench can be extended, following Deep FinResearch Bench (Haque et al., 22 Apr 2026), to assess deep-research (DR) capabilities and report generation. The workflow mandates identical “sell-side” research structures for AI and professional reports, scored on:

• Qualitative rigor: Comprehensiveness, coherence, assumption quality, and analytical depth, each on a 1–4 scale.
• Quantitative accuracy: Symmetric mean absolute percentage error (SMAPE) for forecasted financials, Mean Absolute Error (MAE), bias, Hit Rate, Directional Accuracy (DA), and Signed Recommendation Loss (SRL) for price target and recommendation quality.
• Claim verifiability and credibility: Extraction and verification of atomic claims and citation sources; measurement of factuality, hallucination, and domain trust level.

Automated scoring leverages LLM-judge alignment (notably GPT-5), and all code is standardized for modular, scalable application. DR agents are currently outperformed by human professionals on all qualitative and major quantitative benchmarks (Haque et al., 22 Apr 2026).

7. Simulation-Based ESG-Investment Dilemma Benchmarking

InvestorBench, via the InvestESG MARL environment, enables MARL research into ESG-aware investment and corporate strategy. The scenario simulates $f_{j,t}$8 firms and $f_{j,t}$9 investors over $\alpha_i$0 years, modeling actions (mitigation, greenwashing, resilience), reward functions (short-term profit, mitigation expense, climate-risk benefit), and intertemporal state transitions (climate shocks, wealth, ESG-score, capital).

Major findings:

• Cooperation (genuine mitigation) is only achieved above a threshold fraction of ESG-conscious investor capital.
• Greenwashing strategies fail under recurrent learning; disclosure of climate risks independently increases mitigation.
• Full code and environments are available in PyTorch and JAX for reproducible large-scale experiments (Hou et al., 2024).

---

InvestorBench thus operationalizes a multi-faceted benchmarking paradigm. Across all its layers—manager performance, portfolio construction, LLM-based agent evaluation, deep research output, and simulation-based policy studies—the frameworks provide unified, reproducible methodologies to expose skill, risk, and veracity in financial decision-making environments (Schneider et al., 2024, Li et al., 2024, Kakushadze et al., 2018, Haque et al., 22 Apr 2026, Hou et al., 2024).