PaperlayerView exact revision on arXiv

Source-linked AI summary

FinanceHarness: Autonomous Financial Deep Research Framework

Yijia Xiao, Rujun Han, Yanfei Chen, Zifeng Wang, Ke Jiang, Zhongying CuiZhu, Vishy Tirumalashetty, Wei Wang, Burak Gokturk, Tomas Pfister, Chen-Yu Lee

arXiv:2607.27853v2cs.CLcs.AIq-fin.CP

TL;DR

Financial deep research requires capabilities and temporally constrained evaluation beyond generic deep-research systems and isolated financial NLP benchmarks. FinanceHarness and FinanceGym address this with an expert-guided, point-in-time environment and benchmark, while every evaluated system remains below 40% overall.

  • Problem

    Generic deep-research systems and isolated financial NLP benchmarks do not adequately evaluate comprehensive, temporally constrained financial reports requiring cross-source validation and outcome anticipation.

    Source: §1 · pp. 1–2

  • Method

    The paper builds FinanceGym and FinanceHarness around an expert-validated, point-in-time sandbox with thesis-driven questions, paired pre-/post-cutoff rubrics, agent execution, and reward modeling.

    Source: §1 · p. 2, §7 · p. 13

  • Results

    Every evaluated system stays below 40% overall, while the fixed-backbone progression improves from 25.3% to 32.4% with the full FinanceHarness.

    Source: §7 · p. 14

  • Takeaways & Limitations

    FinanceGym remains a challenging target for financial deep research, with results indicating that stronger retrieval alone does not address the task.

    Source: §1 · p. 2, §7 · p. 14

    The point-in-time sandbox uses a 2025 English-language web corpus, limiting coverage of non-English sources, paywalled professional data, and structured filings.

    Source: Limitations · p. 14

Abstract

from arXiv · show

Powered by advances in LLMs and autonomous agents, deep research has become one of the most widely adopted agentic products. However, most deep research systems write general-purpose reports, which are inadequate for financial deep research. Financial research demands specialized knowledge to analyze historical patterns and forecast upcoming events. Automating financial deep research therefore requires both a layered harness to drive the research agent and a verifiable, point-in-time benchmark that prevents leakage of future information. We present FinanceHarness, a harness that runs finance-oriented tools and practitioner-guided workflows, automating financial deep research end to end: environment and data construction, the agent execution loop, and reward modeling. We further propose FinanceGym, comprising thesis-driven research questions and rubrics that combine pre-cutoff and post-cutoff criteria. Professional expert validation yields an 82% pass rate. With the same open-weight backbone, FinanceHarness improves the overall rubric score from 25.3% to 32.4%, demonstrating the effectiveness of our specialized harness design. However, even pairing FinanceHarness with the most cutting edge LLM (e.g. Opus-5), the FinanceGym score is below 45%, showing that it is a challenging benchmark for financial deep research. Leaderboard is available at: https://financegym.github.io/ and FinanceHarness code is available at: https://github.com/Yijia-Xiao/FinanceHarness.

1. Introduction

FinanceHarness addresses the specialized, temporally constrained demands of financial deep research with a point-in-time sandbox, the FinanceGym benchmark, and an expert-guided agent harness. FinanceGym contains 400 expert-validated questions and remains challenging, with every evaluated system scoring below 40%.

Motivation

  • Financial deep research requires evidence gathering, cross-source validation, report synthesis, financial entity and event relationships, and forecasting their evolution over time.Generic deep research systems may lack the financial tools and expertise needed to analyze earnings reports, economic events, competitors, demand, and macroeconomic conditions.Source: §1 · p. 1
  • Existing deep research benchmarks lack scalable financial evaluation data, while financial NLP benchmarks test isolated capabilities rather than comprehensive, temporally constrained reports.Financial research commonly requires reasoning about future events, making point-in-time evaluation necessary to prevent future-information leakage.Source: §1 · pp. 1–2

Benchmark and environment

  • The point-in-time search sandbox uses real publication dates, dense retrieval, and cutoff-date access control to exclude information published after each question’s cutoff date.This design supports leakage-resistant evaluation of post-cutoff reasoning in financial research reports.Source: §1 · pp. 1–2

Contributions

  • FinanceGym comprises 400 expert-validated research questions and rubrics separating pre-cutoff evidence retrieval from post-cutoff reasoning, with every evaluated system scoring below 40%.The benchmark is grounded in financial situations sampled from an entity graph, investment theses, paired rubrics, multi-stage filtering, and expert validation.Source: §1 · p. 2
  • FinanceHarness automates environment and data construction, agent execution, and reward modeling under a strict point-in-time contract.It provides layered LLM-facing tools, APIs, execution workflows, and evaluation rubrics, running agents in the point-in-time sandbox and scoring them with FinanceGym.Source: §1 · p. 2

2. Related Work

Prior financial benchmarks miss analyst-style, long-form research, while existing deep-research evaluations often lack per-question temporal controls and criterion-level attribution. FinanceGym addresses these gaps with entity-graph questions and evidence-aware rubrics separating pre-cutoff retrieval from post-cutoff outcome anticipation.

Financial benchmarks

  • FinanceBench (Islam et al., 2023) evaluates short-form filing-based question answering, whereas FinanceGym targets analyst-style reports connecting entities, sectors, events, and time.FinanceGym builds questions from a finance entity graph and uses a two-tier rubric separating pre-cutoff evidence retrieval from post-cutoff outcome anticipation.Source: §2 · pp. 2–3

Research agents

  • Research agents combine retrieval-augmented generation with ReAct-style tool use (Yao et al., 2022), differing in whether policy is scaffolded in code or learned by specialized models.Examples include TTD-DR (Han et al., 2025), GPT-Researcher, STORM (Shao et al., 2024), Tongyi-DeepResearch-30B-A3B (Tongyi-DR) (Team et al., 2025b), and MiroThinker-1.7-mini (Team et al., 2025a, 2026).Source: §2 · pp. 2–3
  • Within finance, prior LLM agents largely target trading decisions, whereas FinanceGym evaluates analyst-style research and report generation with verifiable rubrics.The evaluation compares trained models, a fixed ReAct search wrapper with multiple backbones, and agentic search systems on the same point-in-time corpus retriever to separate backbone, scaffold, and tool-distribution effects.Source: §2 · p. 3

Deep-research evaluation

  • Recent deep-research benchmarks support long-form, citation-grounded answers and richer rubrics, but most lack per-question publication-date cutoffs.FinanceBench is the closest finance-domain benchmark; BrowseComp-Plus and DeepResearchGym use fixed corpora without per-question cutoffs, while DeepScholar-Bench is tied to the evolving arXiv index.Source: §2 · p. 3
  • FinanceGym uses evidence-aware rubric judging because financial reports must separately demonstrate historical evidence retrieval and later-outcome anticipation.A report can retrieve correct historical facts yet miss the later outcome, or reason plausibly while omitting source-grounded evidence.Source: §2 · p. 3

3. FinanceGym Construction

FinanceGym combines a point-in-time financial search sandbox with a situation-driven benchmark generated from a finance entity graph, expert-annotated, and evaluated under a rubric-based scoring contract. Its design separates pre-cutoff evidence retrieval from auditable post-cutoff forecasting while preventing temporal leakage.

3.1. Point-in-Time Financial Search Sandbox

  • The search sandbox contains 100+ million dated web articles, uses normalized embeddings and a FAISS IVF-SQ8 index, and enforces access only to documents published by each cutoff date.Full texts are stored separately for citation-grounded synthesis, while agents cannot access post-cutoff articles or the live web.Source: §3.1 · pp. 4–5

3.2. FinanceGym: Situation-Driven Benchmark Construction

  • FinanceGym mines analyst-like situations from a finance graph, selects diverse cutoff dates, and generates unconstrained questions with pre-cutoff and post-cutoff rubric tiers.The graph contains 4.37M filtered edges spanning 1.11M entities and 1.20M source articles; cutoff selection favors event volume, entity diversity, and relation entropy, while the two rubric tiers separate retrievable evidence from forward-looking outcomes.Source: §3.2 · p. 5
  • The construction pipeline filters 29,669 generations to a 2,078-question pool, selects 500 balanced candidates by ILP, and releases 400 expert-annotated questions after an 82% professional-annotation pass rate.The released benchmark contains 2,464 annotated rubric items balanced across 9 topics, 11 sectors, 6 reasoning types, and 12 monthly cutoff-date buckets; release review removes questions that fail quality or privacy and provenance checks.Source: §3.2 · pp. 6–7

3.3. Evaluation Contract

  • The evaluation contract scores each report with question-specific 0–4 rubric tiers and averages the fraction of earned rubric points per question, while PIT leakage invalidates the run.The judge evaluates against the question, thesis, cutoff date, rubrics, and cited report URLs; scores can be disaggregated by criterion type, topic, sector, reasoning type, or cutoff-date bucket.Source: §3.3 · p. 7

4. FinanceHarness

FinanceHarness couples a point-in-time search sandbox with FinanceGym’s grading rubric to produce specialized, auditable financial research reports. Its shared retrieval, orchestration, and rubric interfaces support both evaluation and optimization under a consistent tool-use distribution.

4. FinanceHarness

  • FinanceHarness couples point-in-time retrieval, source reading, citation composition, and rubric-based grading into an executable contract for specialized financial research.The same retrieval tools, orchestration loop, and rubric signal are used during evaluation and optimization, so training and testing share the tool-use distribution.Source: §4 · pp. 7–8

4.1. Expert-guided Harness Design

  • The expert-guided design layers finance-specific tools, deferred schemas, reusable skills, grounding review, and recovery policies over a bounded runtime loop.Capabilities cover exact data retrieval, calculation, valuation, comparables, risk analysis, and scenario reasoning while preserving schema, citation, chaining, and budget invariants.Source: §4.1 · pp. 8–9
  • Disabling URL pre-fetch validation increases visit errors from 2.1% to 39.4%, while excessive searching mainly raises cost because search volume tracks question difficulty rather than score.The end score barely changes without validation because the backbone often self-corrects, whereas search-budget caps primarily control expense.Source: §4.1 · p. 9

4.2. Evaluation and Optimization

  • The evaluation environment fixes the search API, document-fetch contract, citation requirements, and scoring rubric, allowing backbone models to be swapped without changing orchestration.The PIT search sandbox and rubric judge are directly integrated with the harness, and appendices provide cited reports and underlying tool trajectories.Source: §4.2 · p. 9

5. Experimental Setup

The experiments use FinanceGym, a balanced benchmark of 400 expert-annotated questions, and compare agent designs under shared retrieval infrastructure and blind, rubric-based evaluation. The protocol separates scaffolding, backbone, and tool-distribution effects while retaining broader diagnostics on 500 questions.

5.1. Benchmark Dataset

  • FinanceGym contains 400 expert-annotated questions and 2,464 rubric items, balanced across 9 topics, 9 sector leaves, 6 reasoning types, and 12 monthly cutoff buckets.The dataset originates from a 2,078-question publication set, narrowed through a 500-question ILP-balanced subset and subsequent expert annotation and LLM-assisted curation.Source: §5.1 · p. 9

5.2. Agent Baselines

  • Baselines are grouped by scaffolding complexity: fine-tuned open-weight models, fixed ReAct foundation models with search, and engineered agentic search systems.All groups share the PIT-filtered FAISS retriever; FinanceHarness is evaluated as an open-weight layered harness on Qwen3.6-27B, while agentic search systems use Gemini-3-Flash.Source: §5.2 · pp. 9–10
  • The comparison design isolates tool-distribution transfer, engineered scaffolding gains over minimal ReAct, and backbone effects under fixed scaffolding.Fine-tuned models trained with live-web tools are evaluated using corpus-backed Search →FAISS+PIT and Visit →SQLite substitutes.Source: §5.2 · p. 10

5.3. Evaluation Protocol

  • Agents run once on 500 questions, while headline metrics subset outputs to 400 expert-annotated questions and 2,464 rubric items under identical retrieval infrastructure.Agents receive only the question and cutoff date; judges score blinded reports on a 0–4, five-tier rubric, with bootstrap standard errors from 1,000 resamples.Source: §5.3 · p. 10

6. Results and Analysis

FinanceHarness achieves strong FinanceGym performance and a favorable cost-quality trade-off, while results indicate backbone choice remains a major constraint. Task-matched scaffolding improves evidence organization, but GRPO adds only a modest gain.

6. Results and Analysis

  • 44.9% overall performance places FinanceHarness with Opus-5 on the cost-quality frontier; on a 27B backbone, it leads open-weight models, beats the best agentic search system, and trails two proprietary backbones.Figure 5 reports the frontier result, while Table 2 summarizes the system-category comparison.Source: §6 · p. 11, Figure 5 · p. 11, Table 2 · p. 12

6.1. What Drives Performance

  • Backbone changes produce wider performance variation than scaffold changes around Gemini-3-Flash, while task-matched scaffolding recovers and organizes evidence more reliably under a shared point-in-time corpus.Fine-tuned open-weight models also face a tool-distribution shift when evaluated through the corpus-backed FAISS/PIT interface, where several general-purpose search backbones outperform them.Source: §6.1 · p. 11
  • Across systems, pre-cutoff scores are several times higher than post-cutoff scores, indicating that forward-looking rubric items remain difficult even with the same PIT-filtered corpus.FinanceHarness also matches much larger open-weight backbones and approaches the proprietary frontier despite using a 27B backbone.Source: §6.1 · p. 12

6.2. In-Environment Training

  • GRPO training on 172 separate machine-curated instances adds only 0.4 point over the untrained harness, making it a refinement rather than a headline result.The reward combines rubric-style evaluation with an LLM judge of coherence, grounding, and trajectory quality while keeping training and FinanceGym evaluation separate.Source: §6.2 · p. 12, Table 4 · p. 13

7. Conclusion

FinanceGym provides a temporally controlled benchmark, while FinanceHarness integrates expert-guided tools and optimization in the same environment. Across systems, scores remain below 40%, but the full harness improves the fixed-backbone score from 25.3% to 32.4%, with GRPO adding 0.4 point.

Conclusion

  • FinanceGym contains 400 expert-annotated questions and 2,464 rubric items in a reproducible point-in-time search sandbox.Its rubrics separate pre-cutoff evidence retrieval from post-cutoff outcome anticipation.Source: §7 · pp. 13–14
  • On a 27B open-weight backbone, FinanceHarness reaches a score competitive with much larger open-weight models and approaches the proprietary frontier.Figure 6 indicates a favorable performance-to-scale trade-off.Source: Figure 6 · p. 14
  • 32.4% overall score for full FinanceHarness improves over 25.3% for the Qwen3.6-27B search-tool model, with GRPO adding a further 0.4 point.This supports sharing evaluation, tool use, and optimization within the same temporally controlled financial research environment.Source: §7 · p. 14
  • Every system remains below 40% overall across 17 baselines and FinanceHarness, with a persistent pre-/post-cutoff gap.The gap suggests that financial deep research is not solved by stronger retrieval alone.Source: §7 · p. 14

Limitations

The benchmark is limited by its self-collected 2025 English-language corpus and by evaluating specialized deep-research models on a corpus-backed retriever rather than their live-web training stacks.

Corpus scope

  • The self-collected 2025 English-language corpus provides limited coverage of non-English sources, paywalled professional data, and structured filings.These corpus-scope constraints limit the breadth of available financial research evidence.Source: Limitations · p. 14

Tool-distribution shift

  • Specialized deep-research models are evaluated out-of-distribution on the corpus-backed retriever, so scores measure transfer under a controlled point-in-time interface.The evaluation retriever differs from the live-web stacks on which these models were trained.Source: Limitations · p. 14

A. FinanceHarness Ablation Breakdowns

Across identical Qwen3.6-27B and FinanceGym evaluations, scores improve from Vanilla through Naive, Harness, and RFT, with most gains attributable to the harness and only a small increment from RFT. Breakdown results identify crypto, macro/rates, causal analysis, and comparative analysis as the hardest areas.

Evaluation setup

Ablation results

  • The harness configurations provide nearly all score gains across Vanilla, Naive, Harness, and RFT, while RFT adds only a small final increment.All configurations use the same Qwen3.6-27B backbone and FinanceGym evaluation protocol.Source: A. FinanceHarness Ablation Breakdowns · p. 18

Breakdowns

B. Cross-Backbone Harness Evaluation

FinanceHarness remains model-agnostic across Qwen3.6-27B, Gemini-3.5-Flash, and GPT-5.5, with the compact open-weight backbone matching larger models on key reliability dimensions. The harness also routes tools reliably, while risk and market tasks outperform relative valuation because sparse comparables encourage ungrounded peer-table completion.

Evaluation setup

  • The coverage suite evaluates 11 component-focused tasks at N=3 per cell using dual-LLM judges scoring faithfulness, grounding, coherence, and completeness on a 1–5 scale.Grounding can be credited to either a cited source or structured tool data fetched or computed by the agent.Source: B. Cross-Backbone Harness Evaluation · p. 19

Routing and tool behavior

  • Data routing reaches 95.8% and golden-answer computation 90.9%, with 0.99 tool coverage, 0.94 skill-hit rate, and zero unknown-tool errors.These are rule-scored exact rates and coverage measures observed across runs on the same harness.Source: B. Cross-Backbone Harness Evaluation · p. 19

Capability checks

  • All three backbones pass capability checks for multi-turn context use, dollar-preserving /compact summarization, and clarification of under-specified questions.The checks test follow-up context, preservation of numerical content during compression, and appropriate clarification behavior.Source: B. Cross-Backbone Harness Evaluation · p. 19

Cross-backbone results

  • Within about 0.3 of GPT-5.5 on faithfulness and grounding, Qwen3.6-27B outperforms Gemini-3.5-Flash, while coherence and completeness reach ≥4.68 for every backbone.Faithfulness and grounding therefore distinguish the backbones; the comparison uses the unchanged harness, tools, and judges across the 11-task coverage suite.Source: B. Cross-Backbone Harness Evaluation · p. 19, Table 9 · p. 20

Component-category analysis

  • Risk and market tasks are strongest, whereas relative valuation is weakest because small peer sets cause weaker backbones to add ungrounded comparables.The component breakdown covers faithfulness and grounding across backbones; judges penalize unsupported peer-table cells.Source: B. Cross-Backbone Harness Evaluation · p. 19, Table 10 · p. 20

C. Additional Analysis … 1. Accelerators and custom AI silicon

FinanceHarness reveals that financial deep research remains limited by tool-distribution mismatch and forward-looking judgment, while its case studies trace AI infrastructure demand from accelerators through power and data-center assets. The accelerator layer currently combines the strongest growth, margins, and technical barriers, but depends heavily on hyperscaler spending and faces customer-concentration risk.

C.1. Tool-Distribution Shift

  • FinanceGym underperformance by fine-tuned open-weight models reflects tool-distribution shift: systems trained for live-web search must instead use a PIT-filtered corpus with fixed records and a different report format.Their weaker inline attribution reflects a mismatched reporting interface, not necessarily missing knowledge or search ability.Source: §C.1 · p. 20

C.2. Pre-cutoff and Post-cutoff Remain Different Failure Modes

  • About 12% versus 46%: post-cutoff performance remains far below pre-cutoff performance after training, indicating that forward-looking judgment—not evidence recall—is the harder task.Better search, reading, and citation discipline improve pre-cutoff evidence recovery, but post-cutoff rubrics require historical evidence to support judgments verifiable only later.Source: §C.2 · p. 20

C.3. Agentic-system Trade-offs

  • Iterative TTD-DR achieves the highest agentic-search scores, while lighter GPT-Researcher is slightly weaker; STORM and deepagents show that orchestration fit matters more than universal system ranking.TTD-DR repeatedly plans, retrieves, drafts, critiques, and revises, whereas other default output styles are not always aligned with source-grounded financial claims.Source: §C.3 · pp. 20–21

C.4. Per-Axis Breakdowns

  • Tables 11 and 12 extend the FinanceGym analysis with per-sector and per-reasoning-type scores across system groups, complementing the main text’s per-topic breakdown.The tables organize results by nine sector leaves and by causal, comparative, forecast, risk, effectiveness, and quantitative reasoning types.Source: Table 11 · p. 21, §C.4 · p. 21, Table 12 · p. 22

D. FinanceHarness Case-Study Reports

  • FinanceHarness case studies cover six financial research tasks, from industry value-chain analysis and discounted-cash-flow valuation to risk, comparables, options, and fixed income.These are live-web demonstrations separate from FinanceGym’s point-in-time benchmark; their claims are not independently verified and are not investment advice.Source: D. FinanceHarness Case-Study Reports · p. 23

Synthesis

  • The AI buildout creates two linked demand chains: accelerators → HBM → packaging/foundries → equipment, and servers → networking/optics → power, cooling, construction, generation, and data-center real estate.The synthesis identifies durable economics in constrained businesses such as NVIDIA, Broadcom, TSMC, ASML, Arista, Eaton, and Vertiv, while memory, optics, and server assemblers offer faster but more cyclical or commoditized growth.Source: Synthesis · p. 23

Demand anchor

  • Approximately $734 billion of indicated 2026 capital expenditure and finance leases from Alphabet, Amazon, Microsoft, and Meta establish a large demand envelope, while electricity use is projected to rise from 415 TWh in 2024 to about 945 TWh in 2030.Transmission, transformers, cables, and generation constraints could delay around 20% of planned projects, shifting the question from spending growth to whether revenue earns adequate returns.Source: Demand anchor · p. 23

1. Accelerators and custom AI silicon

  • NVIDIA’s Q1 fiscal 2027 data-center revenue reached $75.2 billion, about 92% of total revenue, while Broadcom’s fiscal-Q2 AI revenue reached $10.8 billion and more than doubled year over year.Accelerators and custom silicon offer the strongest current combination of growth, margins, and technical barriers, but customer concentration leaves the layer exposed to hyperscalers developing competing silicon.Source: §1 · p. 24

2. High-bandwidth memory and data-center memory … III. SOURCES

The paper presents AI infrastructure as a hierarchy of profit pools, favoring scarce technology, advanced manufacturing, networking, power and cooling while treating memory, optics, servers and real estate as more cyclical or capital-intensive. Key risks include demand pauses, substitution, customer concentration, capacity overshoot, geopolitics, power constraints, technology transitions and valuation compression.

4. Wafer-fabrication and process-control equipment; 8. Data-center construction and grid interconnection

  • Construction and grid-interconnection businesses gain backlog visibility from data-center, utility-load, and generation investment but remain labor- and execution-intensive, with delays and shortages able to erode that advantage.Quanta reported Q1 revenue of $7.87 billion, up 26%, and record total backlog of $48.5 billion.Source: §8 · p. 26

1. Best combination of durability and valuation: TSMC

  • TSMC is the preferred risk-adjusted exposure, combining approximately 18x forward P/E, above-40% expected 2026 revenue growth and a 60% operating margin despite geographic concentration and overseas-expansion risks.Its foundry and advanced-packaging role reduces dependence on which accelerator architecture wins.Source: §2 · pp. 24–25, §1 · p. 27

2. Highest-quality technology and ecosystem assets: ASML, NVIDIA and Broadcom

  • ASML, NVIDIA and Broadcom provide scarce technology exposure across lithography, integrated data-center platforms, custom accelerators and networking, but face concentrated-customer buyer power.ASML reported a €38.8 billion backlog, while NVIDIA’s data-center platform and Broadcom’s custom-chip exposure span multiple ecosystem layers [1,3,11].Source: §2 · p. 27

3. Best physical-infrastructure beneficiaries: Eaton and Vertiv; 4. Best networking asset: Arista

  • Arista, Eaton and Vertiv are favored physical and networking beneficiaries: Arista combines differentiated economics with 1.6T systems, while power and cooling remain necessary across competing accelerator architectures [17].Arista trades near 37x forward earnings; Eaton offers broader diversification and a lower multiple, whereas Vertiv offers higher purity, faster growth and greater operating leverage.Source: §6 · p. 25, §7 · p. 26, §4 · p. 27

5. Attractive but higher-risk value: Vistra and Constellation

  • Vistra and Constellation offer electricity-generation exposure at less demanding multiples, but local power markets, regulation, transmission constraints and operating performance make the opportunity indirect and higher risk.They screen at approximately 14x and 19x forward earnings, respectively.Source: §9 · p. 26, §5 · p. 27

2. High-bandwidth memory and data-center memory; 6. Tactical rather than durable: Micron, Coherent, Lumentum, Dell and Super Micro

  • Memory, optical-component and server-assembly companies can deliver strong scarcity-driven earnings surprises, but supply response, rapid standards changes, limited assembler economics, working capital and customer concentration create margin-normalization risk.Micron’s fiscal Q3 revenue rose from $9.3 billion to $41.5 billion year over year, while gross margin reached 84.6% and operating margin 80.4%; these margins should not be treated as normal through-cycle economics.Source: §2 · p. 24, §6 · p. 27

7. Least compelling at current structure: data-center REITs; 10. Data-center landlords and interconnection platforms

  • Data-center landlords monetize power-secured sites and interconnection, but roughly 29x EV/EBITDA valuations, 5–6x net debt/EBITDA and continuing capital requirements limit their attractiveness.Equinix reported a 51% adjusted EBITDA margin and expects roughly $4.1 billion of 2026 capital expenditure [7].Source: §10 · p. 26, §7 · p. 27

Key risks

  • The main risks are hyperscaler spending pauses, efficiency and substitution, customer concentration, capacity overshoot, geopolitics, power bottlenecks, technology transitions and valuation compression.A spending pause could propagate through servers, optics, semiconductors, equipment and construction, while grid projects can take four to eight years [6].Source: Key risks · pp. 27–28

D.2. Valuation Compute Seam … 2. Legal Settlements & Real Estate Optimization Charges

Across valuation, risk, and comparables analyses, the block finds that Microsoft’s DCF is below market, TSLA exhibits substantial fat-tailed risk, and Alphabet’s headline valuation discount largely disappears after normalizing non-operating gains. The analyses also document their principal assumptions, data limitations, and adjustments.

Microsoft five-year unlevered DCF: $233 per diluted share

  • The base model values Microsoft at $233.19 per diluted share versus its $381.70 market price, 38.9% below, with $1.674 trillion enterprise value and $1.741 trillion equity value.This is a valuation result, not an investment recommendation; its largest judgment calls are the 11.60% peer-beta-based WACC and the pace at which elevated AI-infrastructure spending normalizes.Source: Microsoft five-year unlevered DCF:… · p. 29, Bottom line · p. 31

1. FY2025 operating base

  • Microsoft’s FY2025 operating base includes $281.724B revenue, $128.528B EBIT, a 45.62% EBIT margin, $64.551B capex, and $136.162B operating cash flow.D&A rose 53.2% and capex 45.1% year over year, supporting assumptions that depreciation catches up with infrastructure investment while capex normalizes.Source: §1 · p. 29

2. Forecast assumptions

  • The Microsoft forecast uses an 11.38% five-year revenue CAGR, modest margin expansion to 46.5%, capex above D&A through FY2029E, and a 4.0% incremental-revenue working-capital investment.These are explicit analyst assumptions rather than company guidance or consensus estimates.Source: §2 · pp. 29–30, Working-capital baseline · p. 29

Cost of capital

  • Microsoft’s WACC is 11.5986%, rounded to 11.60%, based on a 4.68% Treasury yield, 5.00% equity-risk premium, 1.406 peer beta, and 17.63% tax rate.The DCF applies a 2.5% perpetual-growth rate; terminal value contributes 71.03% of enterprise value, making the valuation sensitive to WACC and growth.Source: §4 · p. 30, Cost of capital · p. 31, §5 · p. 31

Data Window and Conventions

  • TSLA’s risk estimates use 1,254 simple daily returns from 1,255 adjusted-price bars and annualize volatility with √252, but the available five-year window extends beyond the requested dates.TSLA’s beta is estimated by OLS against SPY, producing correlation 0.58 and R^2 0.34, so most return variance remains idiosyncratic.Source: Data Window and Conventions · p. 33, §3 · p. 34, §4 · p. 34

5. Return Distribution Diagnostics

  • TSLA’s 1-day VaR is $58,000–$61,000 at 95% and $86,000–$95,000 at 99%, while approximately 60% annualized volatility and beta 2.02 indicate high market sensitivity.Historical 99% VaR is 9.5% versus 8.6% parametric VaR because excess kurtosis is 3.04, indicating fat tails and underestimated normal-model tail risk.Source: §1 · p. 33, §5 · p. 34, §3 · p. 34

Executive Synthesis

2. Legal Settlements & Real Estate Optimization Charges

  • Alphabet’s comparison uses July 28, 2026 market data and TTM results ended June 30, 2026, with a 34.0% TTM operating-margin baseline.Historical legal settlements of $1.40 billion and office space optimization charges of $607 million are excluded from forward earnings estimates and do not materially alter the baseline.Source: Data Snapshots & Estimate Dates · p. 35, §2 · p. 36

Implied Equity Valuation Analysis … AI Infrastructure Value Chain: Beneficiaries Beyond the GPU

This section analyzes AI-infrastructure beneficiaries beyond GPUs through peer-based valuation, option-pricing and bond-risk calculations, and company research across equipment, power, data centers, networking, and components. It combines reported financial and operational metrics with AI-demand linkages and qualifies valuation comparisons based on business mix, capital intensity, balance-sheet strength, and regulatory risk.

Comparability Limitations & Contextual Drivers

  • Alphabet’s valuation comparability is qualified by its mixed monetization, $175 billion–$190 billion projected 2026 AI capex, $121.68 billion net cash, and regulatory overhang.EV-based multiples credit net cash more directly than P/E, while infrastructure investment suppresses free-cash-flow conversion.Source: Comparability Limitations & Contextual… · p. 37

Summary Findings

  • Alphabet’s forward P/E is near peer median at 22.21x versus 21.71x, while normalized trailing P/E is 34.28x versus an unadjusted 16.39x.The unadjusted trailing multiple reflects $157.38 billion in pre-tax non-operating equity gains, while growth and Google Cloud execution support forward valuation parity.Source: Summary Findings · p. 37
  • Peer-median multiples imply GOOGL equity value from $2.93 trillion ($240.55 per share) to $5.26 trillion–$6.17 trillion ($430.83–$506.00 per share).The range uses TTM EBITDA with net cash, forward earnings, and composite EV/EBITDA multiples.Source: Implied Equity Valuation Analysis · pp. 36–37, Summary Findings · p. 38

II. CITED REPORT Black-Scholes Values

  • Black-Scholes pricing produces $11.1610 call and $14.0128 put values, with deltas 0.4910 and −0.5079 and identical gamma and vega of 0.01273 and $0.4451.For 100-share contracts, the premiums are approximately $1,116.10 and $1,401.28; put-call parity holds to a residual of approximately $0.00003 per share.Source: §II · p. 39, Put-Call Parity · p. 39

Price, Duration, and Convexity

Key Observations

E. FinanceHarness Case-Study Trajectories

  • FinanceHarness case studies exercise web research, structured market-data retrieval, DCF/WACC and relative-comps valuation, and citation composition through round-by-round agent trajectories.The appendix presents one broad industry deep-dive and one single-name discounted-cash-flow valuation, with abbreviated repeated tool outputs and scratchpad content.Source: E. FinanceHarness Case-Study… · p. 42

AI Infrastructure Value Chain: Beneficiaries Beyond the GPU

  • The cited relative-comps outputs report median-implied upsides of 61.6%, 3840.1%, 155.3%, 468.5%, and 112.6%, against current prices of $196, $383, $809, $28, and $231, respectively.The corresponding median implied values per share are $316, $15,089, $2,065, $158, and $491.Source: AI Infrastructure Value Chain:… · pp. 42–44
  • The analysis also investigates electricity supply, utilities, construction, power equipment, and data-center operators as beneficiaries of AI-driven capacity expansion.The planned sources include the IEA, Quanta Services, EMCOR, Comfort Systems USA, GE Vernova, Constellation, Vistra, Digital Realty, Equinix, and CoreWeave.Source: AI Infrastructure Value Chain:… · pp. 48–49
  • The research targets semiconductor equipment, foundry, memory, packaging, server, networking, and optical-component companies to assess their exposure to AI infrastructure demand.Searches cover ASML, Applied Materials, Lam Research, KLA, TSMC, Micron, Broadcom, Dell, Arista, Coherent, Lumentum, and Marvell.Source: AI Infrastructure Value Chain:… · pp. 48–50
  • The source-collection plan seeks current financial results and management outlooks, linking AI or data-center demand to growth, margins, orders, backlog, capacity, capex, pricing, and risks.The plan includes company investor-relations sites and other financial sources, with explicit goals to extract these operating and valuation-relevant indicators.Source: AI Infrastructure Value Chain:… · pp. 49–51

E.2. Valuation Compute Seam · Microsoft (MSFT): DCF, WACC & Sensitivity · Range (intrinsic value/share): $179 – $340

Microsoft’s valuation workflow combines FY2025 fundamentals, market inputs, WACC estimation, a five-year DCF, and Gordon-growth sensitivity analysis. The resulting DCF estimates intrinsic value at $233 per share using an 11.6% discount rate and 2.5% terminal growth.

E.2. Valuation Compute Seam

  • The workflow collects reported FY2025 financial statements, market inputs, peer betas, operating assumptions, terminal cash flows, and WACC/perpetual-growth sensitivity inputs.The plan explicitly includes comparing the resulting valuation with Microsoft’s market price and producing an auditable report with sources and assumptions.Source: Microsoft (MSFT): DCF, WACC &… · pp. 52, 55

Microsoft (MSFT): DCF, WACC & Sensitivity

  • Microsoft’s TTM fundamentals include $318.27B revenue, $184.46B EBITDA, $37.01B FCF, and 18.3% revenue growth, alongside $78.23B cash and $125.43B debt.Historical revenue increased from $211.91B in 2023 to $281.72B in 2025, while operating income increased from $88.52B to $128.53B.Source: Microsoft (MSFT): DCF, WACC &… · p. 52

E.2. Valuation Compute Seam

  • Microsoft’s reference data reports a $2.84T market cap, 7.43B shares outstanding, 1.13 beta, and $381.70 current price.The valuation workflow uses these reference fields alongside market rates and fundamentals to identify and classify the equity.Source: Microsoft (MSFT): DCF, WACC &… · pp. 52–53

Microsoft (MSFT): DCF, WACC & Sensitivity

  • 11.6% WACC is derived from a 4.68% risk-free rate, 1.41 beta, 5.0% equity risk premium, 5.2% cost of debt, and 17.6% tax rate.The resulting capital structure uses 98.5% equity and 1.5% debt, with a 6.9% premium over the risk-free rate.Source: Microsoft (MSFT): DCF, WACC &… · p. 54
  • The DCF projects five years of unlevered free cash flow and bridges enterprise value to equity value using net debt of -$67 and 7.465 billion shares outstanding.The terminal value is calculated with the Gordon-growth method and a 2.5% growth rate.Source: Microsoft (MSFT): DCF, WACC &… · p. 55

Range (intrinsic value/share): $179 – $340

  • $233 intrinsic value/share results from a five-year DCF using an 11.6% discount rate, 2.5% Gordon-growth terminal rate, and $1,741 equity value.The model reports $485 in present value of explicit FCF, $1,189 in terminal-value present value, and $1,674 enterprise value.Source: Microsoft (MSFT): DCF, WACC &… · p. 55

Microsoft (MSFT): DCF, WACC & Sensitivity

  • Terminal value contributes 71.0% of enterprise value, making the valuation materially dependent on terminal cash-flow assumptions.The reported terminal value is $2,058 at year N, with a present value of $1,189.Source: Microsoft (MSFT): DCF, WACC &… · p. 55

Related papers

Loading 2607.27853v2…