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 temporally constrained, comprehensive reports that general-purpose harnesses and isolated financial NLP benchmarks do not adequately measure. FinanceHarness introduces a practitioner-guided harness and FinanceGym benchmark, improving fixed-backbone performance from 25.3% to 32.4%, while every evaluated system remains below 40% overall.

  • Problem

    General deep-research harnesses and isolated financial NLP benchmarks do not adequately evaluate comprehensive, temporally constrained financial reports.

  • Method

    FinanceHarness combines finance-oriented tools, practitioner-guided workflows, and a point-in-time sandbox with FinanceGym’s expert-validated thesis-driven questions and pre-/post-cutoff rubrics.

  • Results

    25.3% to 32.4%: the fixed-backbone harness ablation shows progressive overall improvement, while every evaluated system remains below 40% overall.

  • Takeaways & Limitations

    Financial deep research remains challenging, and evaluation, tool use, and optimization benefit from sharing a temporally controlled research environment.

  • Takeaways & Limitations

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

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 mismatch between generic deep-research systems and financial research by combining a point-in-time search environment, the expert-validated FinanceGym benchmark, and an expert-guided agent harness. FinanceGym evaluates evidence retrieval and post-cutoff reasoning while preventing future-information leakage.

  • Motivation: FinanceHarness targets financial research, which requires evidence gathering, cross-source validation, report synthesis, entity-event reasoning, and forecasting beyond generic deep research.The introduction identifies specialized financial capabilities as the central motivation for the framework.
  • Motivation: Existing benchmarks fail to provide scalable evaluation for comprehensive, temporally constrained financial reports, while financial NLP benchmarks test isolated capabilities and do not support point-in-time post-cutoff reasoning.Financial research reports must reason about future events without leaking information published after each question’s cutoff date.
  • Contributions: The authors build a point-in-time search sandbox with a dated web corpus, dense retrieval, and cutoff-date access control that excludes post-cutoff information from research reports.The sandbox is designed to support effective finance-related retrieval while enforcing temporal constraints.
  • Contributions: FinanceGym is an expert-validated benchmark of 400 high-quality research questions and rubrics separating pre-cutoff evidence retrieval from post-cutoff reasoning.Questions are grounded in financial situations sampled from an entity graph, with investment theses and paired rubrics generated using practitioner input.
  • Contributions: FinanceHarness is an expert-knowledge-guided harness that runs agents in the point-in-time sandbox and evaluates them with FinanceGym rubrics.Its layered services cover LLM-facing tools, APIs, execution workflows, and evaluation rubrics, integrating environment construction, agent execution, and reward modeling.

2. Related Work

Prior financial benchmarks target isolated skills or short-form question answering, while newer deep-research evaluations often lack per-question publication-date cutoffs and finance-specific criterion-level attribution. FinanceHarness positions FinanceGym to evaluate analyst-style, citation-grounded financial research under reproducible, point-in-time, and verifiable conditions.

  • Financial benchmarks: Financial NLP benchmarks target isolated skills or finance-specialized LLM evaluation, but do not capture analyst-style reports connecting entities, sectors, events, and time.Examples include sentiment analysis, named entity recognition, factual question answering over SEC filings, and broad finance-LLM evaluation.
  • Research agents: Modern research agents combine retrieval-augmented generation with ReAct-style tool use [Yao et al., 2022], while differing in whether research policy is encoded in scaffolds or trained into the backbone.The evaluation includes trained models, a fixed ReAct wrapper with multiple backbones, and agentic search systems on the same point-in-time corpus retriever to separate backbone, scaffold, and tool-distribution effects.
  • Deep-research evaluation: FinanceGym targets reproducible, point-in-time, and verifiable evaluation dimensions, where point-in-time retrieval uses per-question publication-date cutoffs and verification relies on externally checkable signals.The signals include gold answers, binary rubric items, citation matches, and post-cutoff-date outcomes.
  • Deep-research evaluation: Recent deep-research benchmarks emphasize long-form, citation-grounded answers and richer rubrics, but most rely on live-web retrieval or a single global snapshot rather than per-question publication-date cutoffs.LLM-as-judge methods scale evaluation, yet finance requires criterion-level attribution because a report can retrieve the correct historical information without satisfying every evaluation criterion.
  • Deep-research evaluation: FinanceBench [Islam et al., 2023] evaluates short-form question answering over filings, whereas BrowseComp-Plus [Chen et al., 2025] and DeepResearchGym [Coelho et al., 2025] use fixed corpora without per-question publication-date cutoffs.DeepScholar-Bench [Patel et al., 2025] is closer on the point-in-time axis but is tied to the evolving arXiv index.

3. FinanceGym Construction

FinanceGym is built on a point-in-time financial search sandbox and a data-driven pipeline that generates, filters, balances, and expert-annotates thesis-driven questions. Its rubric-based evaluation scores reports against pre-cutoff evidence and post-cutoff outcomes while enforcing cutoff-date access.

  • 3.1 Search Sandbox: The search sandbox contains 100+ million dated web articles, uses Qwen3-Embedding-4B vectors and a FAISS IVF-SQ8 index, and blocks post-cutoff documents.Full texts are stored separately for citation-grounded synthesis, while the API permits historical search and reading only through each assigned cutoff date.
  • 3.2 Benchmark Generation: Cutoff dates maximize event volume, entity diversity, and relation entropy, while the finance graph provides 4.37M filtered edges spanning 1.11M unique entities.The graph is extracted from finance-relevant sources, and cutoff selection penalizes days dominated by a single entity or relation type.
  • 3.2 Benchmark Generation: FinanceGym generates analyst-style questions, investment theses, and two-tier rubrics from mined financial situations without topic, sector, or reasoning prompts.Pre-cutoff criteria test retrievable facts, whereas post-cutoff criteria test outcomes verifiable only after the cutoff, separating evidence retrieval from forward-looking reasoning.
  • 3.2 Benchmark Generation: The pipeline reduces 29,669 unconstrained generations to a 2,078-question publication pool, selects 500 balanced questions with an ILP, and retains 411 after expert annotation.Quality gates assess feasibility, relevance, coherence, evidence grounding, and balance; expert annotators score clarity and feasibility of questions and rubric items.
  • 3.3 Scoring Contract: The outcome score averages each report’s fraction of earned rubric points across questions, with each question contributing equally regardless of rubric length.An LLM judge assigns 0–4 criterion scores from not addressed through fully grounded, using the question, thesis, cutoff date, rubrics, report, and cited URLs.
  • 3.3 Scoring Contract: Evaluation enforces publication dates ≤ the cutoff, prohibits live-web access, and invalidates runs with detected point-in-time leakage.Pre-cutoff rubrics assess retrieval and synthesis, while post-cutoff rubrics assess anticipation of developments verifiable only afterward.

4. FinanceHarness

FinanceHarness is an executable finance-specific research harness built around a point-in-time retrieval contract and rubric-based grading. Its layered orchestration, tooling, grounding, and guardrails align training and evaluation while preserving model-swappability.

  • FinanceHarness combines a point-in-time search sandbox with FinanceGym grading, using identical retrieval tools, orchestration, and rubric signals during optimization and evaluation.This alignment ensures models trained inside the harness encounter the same tool-use distribution at test time.
  • The harness exposes layered finance tools for PIT search, source reading, citation composition, and report finalization rather than relying on generic web search.Its core loop supports connecting entities, sector context, numerical evidence, and forward-looking risks into auditable research reports.
  • The runtime enforces bounded agent execution, dispatch, schema validation, recovery, tool-result chaining, citation finalization, and run-level budget limits, while tiered tools keep prompts compact.Deferred tools load full schemas only after the model selects a tool family, and extension connectors can use MCP or CLI interfaces.
  • FinanceHarness supports research, analytical, and automatic prompt modes over a consistent registry, while reusable skills dynamically activate specialist tools and context.The consistent registry prevents trajectories from being stranded when modes switch, and skill routing uses descriptions rather than hard-coded names.
  • 39.4% versus 2.1%: disabling URL pre-fetch validation raises visit errors, although the end score barely changes because the backbone often self-corrects.The resulting trajectories become substantially more expensive; excessive searching likewise primarily affects cost rather than quality.
  • The harness couples directly to the PIT search server and rubric judge, while backbone swaps leave the search API, document-fetch contract, citation requirements, and scoring rubric unchanged.This separates the learnable model layer from the environment and evaluation stack.

5. Experimental Setup

FinanceGym is built from 400 expert-annotated questions and 2,464 rubric items, balanced across topics, sectors, reasoning types, and monthly cutoff-date buckets. Experiments compare three scaffolding-based baseline groups under controlled retrieval and evaluate agents with point-in-time evidence and rubric-level judging.

  • Dataset: FinanceGym contains 400 expert-annotated questions with 2,464 rubric items, selected from a 500-question ILP-balanced subset and balanced across 9 topics, 9 sector leaves, 6 reasoning types, and 12 monthly cutoff-date buckets.The publication dataset contains 2,078 questions before subset selection, and each question has a mean of 6.16 annotated rubric items.
  • Baselines: Baselines are grouped by scaffolding complexity into fine-tuned open-weight models, foundation models with a search tool, and agentic search systems.The groups share a point-in-time-filtered FAISS retriever; the foundation-model group uses a fixed 30-step ReAct wrapper, while agentic systems make orchestration decisions in code.
  • Experimental comparisons: Identical questions and corpus retrieval isolate tool-distribution transfer, the added value of engineered scaffolding over minimal ReAct, and model-identity effects under fixed scaffolding.Fine-tuned models trained with live-web tools are evaluated through corpus-backed Search →FAISS+PIT and Visit →SQLite equivalents.
  • Evaluation protocol: Agents run once on the full 500-question subset, while headline metrics aggregate judge outputs for FinanceGym’s 400 questions and 2,464 rubric items; the full run supports auxiliary diagnostics.All baselines use identical search infrastructure.
  • Evaluation protocol: A Gemini-3.5-Flash judge scores each rubric criterion on a 0–4 scale using pre- and post-cutoff evidence plus the agent’s cited report, with 1,000-bootstrap standard errors.The judge is distinct from the Gemini-3-Flash baseline backbone.

6. Results and Analysis

FinanceHarness remains on FinanceGym’s performance–cost frontier and is among the strongest systems, leading open-weight models while trailing only two proprietary backbones. Results also show that backbone choice strongly constrains performance, task-matched scaffolding improves efficiency, and post-cutoff forecasting remains substantially harder than pre-cutoff evaluation.

  • Cost-quality Frontier: FinanceHarness stays on the performance–cost frontier, while Opus-5 achieves the best overall performance at 44.9%.The frontier reflects a favorable performance-to-cost trade-off among benchmarked systems.
  • Model and Procedure Effects: Backbone changes produce wider performance differences than scaffold changes around Gemini-3-Flash, while FinanceHarness reaches an efficiency frontier despite its 27B backbone.Fine-tuned open-weight models also face tool-distribution shift when evaluated through the corpus-backed FAISS/PIT interface.
  • Headline Results: On a 27B open-weight backbone, FinanceHarness leads all open-weight models, slightly outperforms the best agentic search system, and trails only two proprietary backbones.Table 2 reports normalized rubric means for Overall, Pre-cutoff, and Post-cutoff scores with bootstrap standard errors.
  • Pre-cutoff/Post-cutoff Analysis: Across systems, pre-cutoff scores are several times higher than post-cutoff scores, indicating that forward-looking rubric items remain largely partially addressed or missed.Because every agent uses the same PIT-filtered corpus, the gap reflects benchmark difficulty rather than retrieval setup.
  • Harness Ablation and Training: GRPO training adds only 0.4 point over the untrained FinanceHarness, making it a refinement rather than a headline result.The training uses 172 separate machine-curated instances and keeps training independent of FinanceGym.

7. Conclusion

FinanceGym is a point-in-time benchmark built from expert-annotated questions and rubric items, while FinanceHarness is an expert-knowledge-guided agent harness operating in the same reproducible sandbox. Across 17 baselines and FinanceHarness, systems remain below 40% overall with a persistent pre-/post-cutoff gap, while the full harness improves the fixed-backbone score from 25.3% to 32.4%.

  • Conclusion: FinanceGym comprises 400 expert-annotated questions and 2,464 rubric items in a reproducible point-in-time search sandbox, while FinanceHarness runs and optimizes agents within that environment.The benchmark combines pre-cutoff synthesis with post-cutoff reasoning rubrics.
  • Conclusion: 32.4% overall rubric score is achieved by FinanceHarness, improving from 25.3% for the fixed-backbone Qwen3.6-27B search-tool model.The ablation compares the fixed-backbone model with the full FinanceHarness.
  • Conclusion: Below 40% overall, every evaluated system shows a persistent pre-/post-cutoff gap across 17 baselines and FinanceHarness.The benchmark separates pre-cutoff evidence retrieval from post-cutoff outcome anticipation, indicating that stronger retrieval alone does not address financial deep research.

Limitations

The benchmark is limited by incomplete source coverage and by evaluating specialized deep-research models under a controlled retriever interface rather than their trained live-web stacks.

  • Limitations: The 2025 English-language web corpus has limited coverage of non-English sources, paywalled professional data, and structured filings.The point-in-time search sandbox uses a corpus collected by the authors.
  • Limitations: Scores for specialized deep-research models are transfer results under a controlled point-in-time interface because evaluation uses the corpus-backed retriever rather than live-web stacks.These models are evaluated out-of-distribution on the authors’ corpus-backed retriever.

A. FinanceHarness Ablation Breakdowns

Across topic, reasoning type, sector, and situation type, FinanceHarness configurations improve FinanceGym scores, with most gains coming from the harness and only a small additional increment from RFT. Crypto and macro/rates are the weakest sectors, while causal and comparative analysis are the weakest reasoning types.

  • A. FinanceHarness Ablation Breakdowns: The ablation evaluates normalized rubric means across topic, reasoning type, sector, and situation type under a common backbone and FinanceGym protocol.Tables 5–8 report the topic, reasoning-type, sector, and situation-type breakdowns, respectively.
  • A. FinanceHarness Ablation Breakdowns: Scores increase from Vanilla through Naive, Harness, and RFT, with harness configurations providing nearly all gains and RFT adding only a small final increment.All configurations use the same Qwen3.6-27B backbone and FinanceGym evaluation protocol.
  • A. FinanceHarness Ablation Breakdowns: Crypto and macro/rates remain the weakest sectors, while causal and comparative analysis remain the weakest reasoning types.The ablation is broken down by topic, reasoning type, sector, and situation type; the reported weakness patterns concern sectors and reasoning types.

B. Cross-Backbone Harness Evaluation

FinanceHarness is evaluated unchanged across Qwen3.6-27B, Gemini-3.5-Flash, and GPT-5.5 using an 11-task coverage suite and dual-LLM judging. Qwen3.6-27B performs comparably to GPT-5.5 on faithfulness and grounding, while risk and market tasks are strongest and relative valuation weakest.

  • Evaluation setup: The coverage suite contains 11 component-focused tasks run at N=3 per cell and scored from 1–5 by Gemini-3.5-Flash and GPT-5.5 judges.Scores assess faithfulness, grounding, coherence, and completeness; grounding accepts cited sources or structured tool data.
  • Capability checks: All three backbones pass capability checks for multi-turn context use, figure-preserving /compact summaries, and clarification of under-specified questions.These checks test deployed-system behaviors beyond end-to-end FinanceGym report quality.
  • Component categories: Relative valuation is weakest because the comparables tool returns a small peer set, prompting weaker backbones to fill tables from prior knowledge and incur grounding penalties.The resulting ungrounded cells are penalized by the judges.
  • Cross-backbone results: Within about 0.3 of GPT-5.5 on faithfulness and grounding, Qwen3.6-27B also outperforms Gemini-3.5-Flash on the unchanged FinanceHarness coverage suite.The comparison uses the same harness, tools, and judges across three backbones.
  • Component categories: Risk and market-data tasks are strongest across backbones because value-at-risk, beta, correlation, rates, and indices compose cleanly and remain grounded.Faithfulness and grounding are broken down by component category in Table 10.

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

FinanceHarness analysis identifies tool-distribution shift, attribution, temporal reasoning, and orchestration fit as major determinants of FinanceGym performance, while case studies trace AI-compute demand across silicon and physical-infrastructure chains. The accelerator-focused analysis positions NVIDIA, AMD, Broadcom, and Marvell as the starting layer of this value chain.

  • C.1. Tool-Distribution Shift: FinanceGym underperformance by fine-tuned open-weight models reflects tool-distribution shift: they were optimized for live-web search but evaluated through a PIT-filtered corpus retriever with fixed records and no live-web access.Their weaker attribution can reflect learned reporting conventions rather than absent knowledge or search ability.
  • C.2. Pre-cutoff and Post-cutoff Remain Different Failure Modes: About 12% versus 46% separates post-cutoff from pre-cutoff performance after training, showing that forward-looking judgment remains harder than evidence recall.The benchmark’s pre-cutoff and post-cutoff rubrics therefore represent distinct failure modes.
  • C.3. Agentic-system Trade-offs: TTD-DR achieves the highest scores among agentic search systems through iterative planning, retrieval, drafting, critique, and revision, whereas GPT-Researcher is slightly weaker but lighter and less iterative.STORM and deepagents show that orchestration comparisons measure fit to FinanceGym’s source-grounded financial rubric rather than universal system quality.
  • C.4. Per-Axis Breakdowns: Tables 11 and 12 break down FinanceGym scores by sector and reasoning type across system groups, complementing the main text’s per-topic breakdown.The reasoning-type columns cover causal, comparative, forecast, risk, effect, and quantitative analysis.
  • D. FinanceHarness Case-Study Reports: FinanceHarness case studies span industry value-chain research, discounted-cash-flow valuation, risk analytics, relative comparables, option pricing, and fixed-income analysis.Each boxed report presents a research question, cited report, and sources, with representative tool trajectories shown separately.
  • Synthesis: AI-compute demand propagates through a silicon chain from accelerators to HBM, advanced packaging and foundries, and semiconductor equipment, alongside a physical-infrastructure chain from servers and networking to power, cooling, construction, generation, and data centers.The synthesis frames these as two interconnected demand chains extending well beyond GPUs.
  • Financial snapshot: Technology, qualification requirements, or physical constraints support durable economics for NVIDIA, Broadcom, TSMC, ASML, Arista, Eaton, and Vertiv, while TSMC and selected power producers offer better valuation asymmetry and memory, optics, and server assemblers offer higher but more cyclical growth.The screening metrics are recent year-over-year revenue growth, operating margins, and forward P/E unless otherwise indicated, and are especially cycle-sensitive for memory, utilities, and hardware assemblers.
  • 1. Accelerators and custom AI silicon: NVIDIA, AMD, Broadcom, and Marvell form the accelerator and custom-silicon starting layer, with NVIDIA supplying accelerators, networking products, systems, and a software platform.NVIDIA’s Q1 fiscal 2027 Data Center revenue reached $75.2 billion and represented approximately 92% of total revenue; AMD is the principal listed merchant alternative, while Broadcom and Marvell provide custom-silicon and networking exposure.

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

The paper maps AI infrastructure across memory, semiconductor manufacturing, systems, networking, power, construction, generation and data-center real estate, then ranks exposures by durability, valuation and risk. It favors scarce proprietary capacity and power-related infrastructure while warning that spending pauses, substitution, bottlenecks and regulation can propagate across the chain.

  • 3. Best physical-infrastructure beneficiaries: Eaton and Vertiv: Power distribution and cooling are durable physical-infrastructure beneficiaries because rack density makes power conversion and heat removal binding constraints; Eaton offers diversification while Vertiv offers higher purity and growth.Vertiv’s Q1 adjusted operating margin rose 430 basis points to 20.8%, while full-year organic growth guidance increased to 29–31%; entry valuation remains important.
  • 7. Least compelling at current structure: data-center REITs: Equinix and Digital Realty have valuable power-secured portfolios, but high leverage, elevated EV/EBITDA multiples and continuing capital requirements limit upside relative to equipment suppliers.Equinix reported 10% Q1 revenue growth, a 51% adjusted EBITDA margin and 12% AFFO growth, with approximately 60% of its largest deals AI-related.
  • 1. Best combination of durability and valuation: TSMC: TSMC combines 60% operating margins, above-40% expected 2026 revenue growth and approximately 18x forward P/E, making it the preferred risk-adjusted exposure.Its foundry and advanced-packaging role reduces dependence on which accelerator architecture wins, although overseas expansion costs and geographic concentration remain risks.
  • 4. Best networking asset: Arista: Arista’s 1.6T systems address both scale-up and scale-out AI fabrics, supported by exceptional margins, balance-sheet strength and an AI-focused roadmap, but its roughly 37x forward valuation and customer concentration constrain the thesis.The broader networking chain includes Broadcom Tomahawk 6 switching silicon and optical components, while Arista’s products target clusters ranging from thousands to hundreds of thousands of accelerators.
  • 5. Attractive but higher-risk value: Vistra and Constellation: Vistra and Constellation offer electricity-generation exposure at less demanding multiples than many AI-equipment beneficiaries, but their returns depend on regulation, regional prices, operations and transmission availability.The IEA expects data-center electricity generation to exceed 1,000 TWh by 2030, while nuclear becomes more important after 2030.
  • Key risks: The principal risks are hyperscaler spending pauses, efficiency and custom-ASIC substitution, power bottlenecks, regulation and coordinated technology transitions across HBM4, N2, 1.6T Ethernet and liquid cooling.A spending pause could propagate through servers, optics, semiconductors, equipment and construction, while transmission projects can take four to eight years.
  • 2. Highest-quality technology and ecosystem assets: ASML, NVIDIA and Broadcom: Scarce proprietary technology currently captures the highest value: ASML’s EUV ecosystem, NVIDIA’s integrated accelerator platform, Broadcom’s custom-accelerator and networking exposure, and TSMC’s advanced manufacturing.ASML’s €38.8 billion backlog, NVIDIA’s current economics, and Broadcom’s custom-accelerator exposure diversify participation across chipmakers and architectures, but large customers retain substantial buyer power.

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

The analyses span Microsoft DCF valuation, TSLA risk analytics, and Alphabet relative valuation, combining explicit modeling assumptions with sensitivity, tail-risk, and earnings-normalization diagnostics. Results indicate Microsoft’s modeled value is below its market price, TSLA exhibits high volatility and fat-tailed risk, and Alphabet’s apparent valuation discount is materially affected by non-operating gains.

  • Microsoft five-year unlevered DCF: $233 per diluted share: Microsoft’s DCF yields $233.19 per diluted share versus a $381.70 market price, implying the modeled value is 38.9% below market.The model produces $1.674 trillion of enterprise value and $1.741 trillion of equity value; no investment recommendation is made.
  • 1. FY2025 operating base: Microsoft’s FY2025 operating base includes $281.724 billion of revenue, $128.528 billion of EBIT, a 45.62% EBIT margin, and $64.551 billion of capital expenditures.D&A rose 53.2% and capex rose 45.1% from FY2024, supporting assumptions that depreciation catches up with infrastructure investment while capex gradually normalizes.
  • Working-capital baseline: The DCF forecasts a 4.0% investment for every dollar of incremental revenue, based on FY2025 working-capital outflows equal to 3.79% of revenue growth.Forecast assumptions are explicitly identified as analyst assumptions rather than company guidance or consensus estimates.
  • Cost of capital: The Microsoft valuation uses an 11.5986% WACC, a 2.5% perpetual-growth rate, and a terminal value representing 71.03% of enterprise value.The largest judgment calls are the peer-beta-based WACC and the pace at which elevated AI infrastructure spending normalizes.
  • 5. Return Distribution Diagnostics: TSLA’s excess kurtosis of 3.04 indicates fat-tailed returns, with historical 99% VaR of 9.5% exceeding normal-parametric VaR of 8.6%.The analysis attributes the divergence to the normal model underestimating high-confidence tail risk and identifies historical VaR as more conservative at 99%.
  • TSLA Risk Analytics — 1-Day VaR, Volatility, Beta, and Correlation: TSLA’s 1-day VaR is $58,000–$61,000 at 95% confidence and $86,000–$95,000 at 99%, while annualized volatility is approximately 60% and beta to SPY is 2.02.The approximately five-year tool window contains 1,254 simple-return observations but is broader than the requested 2023-01-03 to 2025-12-31 range.
  • Executive Synthesis: Alphabet’s trailing GAAP P/E is 16.39x versus a peer median of 25.43x, but normalized trailing P/E is 34.28x after excluding $127.32 billion of earnings from non-operating equity gains.The peer table also reports Alphabet’s forward P/E at 22.21x versus a 21.71x peer median and EV/EBITDA at 22.46x versus 16.25x.
  • 2. Legal Settlements & Real Estate Optimization Charges: Alphabet’s TTM non-operating equity gains totaled $157.38 billion, while prior legal settlements and office-optimization charges were excluded from forward estimates and did not materially change the 34.0% operating-margin baseline.The cited prior items were a $1.40 billion legal settlement in Q2 2025 and a $607 million office-space optimization charge in Q3 2024.

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

The analyzed sections combine relative valuation, option and bond analytics, and a FinanceHarness case study spanning AI infrastructure beneficiaries beyond GPUs. Results include Alphabet’s $2.93T–$6.17T implied equity range, Black-Scholes option Greeks, Treasury repricing accuracy, and multi-company AI value-chain comparisons.

  • Implied Equity Valuation Analysis; Comparability Limitations & Contextual Drivers; Summary Findings; III. SOURCES: $2.93T–$6.17T: Peer-median multiples imply Alphabet equity values from $2.93T ($240.55/share) to $5.26T–$6.17T ($430.83–$506.00/share), while forward P/E is near peer median.The valuation framework applies peer median P/E and EV/EBITDA multiples; Alphabet’s headline trailing P/E is distorted by $157.38B of non-operating equity gains, and AI infrastructure capex suppresses free-cash-flow yield.
  • D.5. Derivatives; Apple (AAPL): Option Pricing — Greeks & Put-Call Parity; II. CITED REPORT Black-Scholes Values; Put-Call Parity: $11.1610/$14.0128: Black-Scholes values the European call/put at those per-share premiums, with corresponding deltas, gamma, vega, theta, and rho reported for the stated AAPL inputs.For 100-share contracts, premiums are approximately $1,116.10 and $1,401.28; put-call parity leaves a residual of approximately $0.00003 per share.
  • Put-Call Parity; D.5. Derivatives: $0.00003: The put-call-parity residual confirms consistency between the calculated AAPL call and put values to rounding.The calculation uses continuous dividends and the stated spot, strike, volatility, rate, yield, and 90/365-year maturity.
  • D.6. Fixed Income; U.S. Treasury Bond: Duration, Convexity & Repricing; U.S. Treasury 4.125% Note Due 2032-11-15 — Bond Analytics; Price, Duration, and Convexity; Percentage price changes: 101.5680 versus 101.5681: Duration plus convexity nearly matches the exact clean-price reprice for a 50-basis-point yield decline, whereas duration-only gives 101.5220.For the +50-basis-point shock, convexity similarly gives 95.9992 versus the exact 95.9991, while duration-only gives 95.9535; the bond’s clean price is 98.7379 and modified duration is 5.64.
  • Bond Terms and Accrued Interest; Key Observations; Price, Duration, and Convexity; D.6. Fixed Income: +2.87% versus −2.77%: The Treasury bond gains more when yields fall than it loses when yields rise by the same amount, demonstrating positive convexity and duration-only underpricing of about $0.046 per $100.The bond terms use a 4.125% semiannual coupon, 2032-11-15 maturity, 2026-07-27 settlement, $100 face value, and 4.35% YTM; accrued interest is $0.8183 and dirty price is 99.5562.
  • E. FinanceHarness Case-Study Trajectories; E.1. Industry Deep Research: FinanceHarness exercises web research, structured market-data retrieval, DCF/WACC and relative-comps valuation, and citation composition through illustrative industry and single-name trajectories.The trajectories are presented as two focused examples rather than one per report and show round-by-round agent execution.
  • AI Infrastructure Value Chain: Beneficiaries Beyond the GPU: 61.6% median-implied upside: NVIDIA’s peer-comps range is $216–$422 per share, while Micron, SMCI, Coherent, and Constellation show 155.3%, 468.5%, 112.6%, and 36.2% median-implied upside, respectively.The comparisons cover GPU, memory, server, optical, and power-related value-chain companies using P/E, EV/EBITDA, EV/Revenue, and P/S implied values.
  • AI Infrastructure Value Chain: Beneficiaries Beyond the GPU; E.1. Industry Deep Research: -47.4% and -17.7%: Vertiv and Equinix have negative median-implied upside, while Microsoft has -17.4%, contrasting with positive implied upside across several semiconductor, infrastructure, and power beneficiaries.The AI value-chain analysis uses peer-relative trading multiples and searches company, industry, energy, and data-center sources for demand, capacity, margins, capex, backlogs, outlook, and risks.

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

The Microsoft valuation seam combines FY2025 financial-data collection, WACC estimation, a five-year Gordon-growth DCF, and sensitivity analysis, producing an intrinsic-value range of $179–$340 per share. The base case estimates $233 per share at an 11.6% discount rate and 2.5% terminal growth.

  • E.2. Valuation Compute Seam: The workflow gathers FY2025 filings and market inputs before deriving WACC, forecasting unlevered FCF, bridging enterprise value to equity value, and comparing valuation with market price.The plan explicitly includes reported financials, peer betas, operating assumptions, terminal cash flows, and WACC/perpetual-growth sensitivity.
  • E.2. Valuation Compute Seam: Microsoft’s TTM fundamentals include $318.27B revenue, $184.46B EBITDA, $37.01B FCF, $170.14B operating cash flow, and 18.3% revenue growth.The historical series shows revenue increasing from $211.91B in 2023 to $281.72B in 2025.
  • Microsoft (MSFT): DCF, WACC & Sensitivity: Microsoft’s market reference lists a $2.84T market cap, 7.43B shares outstanding, 1.13 beta, and $381.70 price, alongside peer-reference collection for valuation inputs.The reference identifies Microsoft as a United States technology/software-infrastructure company.
  • Microsoft (MSFT): DCF, WACC & Sensitivity: WACC is 11.6%, based on a 4.68% risk-free rate, 1.406 beta, 5.0% equity-risk premium, 5.18% cost of debt, 17.6% tax rate, and 1.5% debt-to-equity.CAPM implies 11.7% cost of equity, while after-tax cost of debt is 4.3%.
  • E.2. Valuation Compute Seam: The DCF uses a five-value unlevered FCF schedule, Gordon-growth terminal value, negative $66.819 net debt, and 7.465 billion shares outstanding.The explicit forecast spans five years before terminal-value calculation.
  • Microsoft (MSFT): DCF, WACC & Sensitivity: $233 intrinsic value/share results from a five-year Gordon-growth DCF using an 11.6% discount rate, 2.5% terminal growth, $1,741 equity value, and negative $67 net debt.The terminal value contributes 71.0% of enterprise value.
Loading 2607.27853v2…