Our trading system is primarily quantitative — signals generated by algorithms, executed by agents, managed by risk rules. But quantitative signals have a blind spot: they process price and volume data well, but they struggle with the qualitative dimensions of companies that drive long-term fundamental value.
We built a Fundamental Analyst agent to address this. Its job is to generate comprehensive fundamental analysis reports on equities — the kind of work a human sell-side analyst would do, but automated, faster, and scalable across more companies than any individual analyst could cover.
We tested it on two of the most widely-analyzed companies in the market: Apple (AAPL) and Nvidia (NVDA). This is what we found.
What the Agent Does
The Fundamental Analyst is a multi-phase LangGraph agent that orchestrates several research modules:
Phase 1: Financial Statement Analysis. The agent retrieves quarterly and annual financial statements — income statement, balance sheet, cash flow statement — going back 8 quarters. It computes key financial ratios, trend analyses, and quality metrics: gross margin trend, operating leverage, free cash flow conversion, return on invested capital, debt-to-equity trajectory, and working capital dynamics.
Phase 2: Business Quality Assessment. Beyond the numbers, the agent evaluates qualitative dimensions: competitive moat (is the business defensible?), management quality (capital allocation history), earnings quality (cash conversion vs accounting earnings), and the key risk factors that could impair the business model.
Phase 3: Valuation. The agent builds a DCF model using analyst consensus estimates, constructs a comparable company valuation (P/E, EV/EBITDA, P/FCF relative to sector peers), and computes an intrinsic value range.
Phase 4: Synthesis. The final report integrates all three phases into an investment thesis: bull case, base case, and bear case, with an explicit assessment of what assumptions must hold for each scenario.
The full report runs to 15–25 pages and is generated in approximately 4–6 minutes per company.
AAPL Analysis: What the Agent Found
The Apple report generated in January 2026 identified several findings that align with current analyst consensus and some that were more nuanced.
On financial quality (strong): The agent correctly identified Apple's exceptional free cash flow generation — FCF margin consistently above 25%, FCF conversion above 100% of GAAP net income in most quarters. It correctly noted the buyback program's effect on EPS growth despite modest revenue growth: Apple's share count has declined approximately 35% over the past decade, mechanically boosting per-share metrics even in quarters with flat revenue.
On revenue growth (balanced): The agent flagged iPhone unit saturation as the primary concern — unit volumes have been roughly flat for several years, with revenue growth coming from ASP increases and services attachment rather than unit expansion. It correctly identified services (App Store, Apple TV+, Apple Pay) as the key re-rating catalyst — services revenue carries higher margins and more predictable growth than hardware.
On valuation (nuanced): The agent's DCF produced an intrinsic value range of $160–$200 per share under base case assumptions, with the bull case (services penetration acceleration, India market growth) reaching $240. The bear case (regulatory pressure on App Store economics, China market risk) was modeled at $130. Current market price at time of analysis was approximately $215, placing the stock slightly above the base case midpoint — a reasonable conclusion that aligns with professional consensus.
Where the agent struggled: The agent's treatment of Apple's AI strategy was thin. It correctly identified Apple Intelligence as a product initiative but could not assess its competitive positioning versus Google's Gemini or Microsoft's Copilot integration in any depth. Qualitative competitive analysis of rapidly-evolving technology products is a genuine limitation of the current implementation.
NVDA Analysis: The Hard Case
Nvidia was the more interesting test because it is a company where the fundamental picture has been changing faster than any static valuation framework can easily capture.
On the growth story: The agent correctly identified the three-part Nvidia thesis: data center dominance (H100/H200 GPU market share), the software moat (CUDA ecosystem lock-in), and the inference scaling effect (AI model inference growing faster than training). The financial quality analysis was strong — Nvidia's gross margins expanded from 65% to 73%+ in the data center cycle, and the agent correctly flagged this margin expansion as structurally driven rather than temporary.
On the valuation challenge: This is where the agent produced its most interesting output. The standard DCF framework breaks down for a company growing at 120%+ annually. The agent's base case DCF under conservative assumptions (25% revenue CAGR for 5 years, then normalization to 12%) produced a value significantly below current market price. The bull case (40% CAGR for 5 years) roughly matched market price. But neither case adequately modeled the possibility of a step-change in addressable market if AI inference becomes ubiquitous infrastructure.
The agent identified this limitation explicitly in its report: "DCF analysis assumes a path to normalized growth. For companies at Nvidia's current growth rate with potential TAM expansion driven by infrastructure-level technology adoption, DCF undervalues optionality. The appropriate framework may be more similar to early-stage infrastructure companies than mature technology businesses."
This self-aware limitation is one of the more impressive aspects of the system — it recognized the boundaries of its own analytical framework.
On competitive risk: The agent correctly flagged AMD as a credible alternative GPU supplier and identified the risk of custom silicon (AWS Trainium, Google TPU, Microsoft Maia) reducing Nvidia's addressable market for cloud providers. It underweighted the risk of Chinese market restrictions on GPU exports, which became more significant in the months following the analysis.
The Grading System
The agent produces an output that includes a proprietary grading system across five dimensions:
| Dimension | AAPL | NVDA |
|---|---|---|
| Financial quality | A | A |
| Competitive moat | A+ | A |
| Management quality | A | A- |
| Earnings quality | A | A- |
| Valuation attractiveness | C+ | C |
The grading methodology went through three iterations during development. The initial grading produced compressed scores — most companies received B to B+ grades, which was not useful for differentiation. The second iteration added percentile-relative grading (how does this company compare to sector peers?), which improved differentiation but introduced a problem: outstanding companies in weak sectors scored highly while mediocre companies in strong sectors scored poorly.
The current iteration combines absolute quality thresholds with peer-relative positioning, weighted by sector cyclicality. This produces more consistent grades across different market environments.
What Automated Fundamental Analysis Can and Cannot Do
Where it excels: - Financial statement analysis and ratio computation — faster and more comprehensive than human analysts for the quantitative dimensions - Consistency — human analysts are subject to anchoring, recency bias, and narrative bias; the agent applies the same framework every time - Coverage breadth — the agent can produce reports on 20 companies per day; a human analyst can cover 3–5 companies in depth - Identification of financial red flags — the agent is particularly good at spotting deteriorating quality metrics (rising accounts receivable relative to revenue, declining FCF conversion, working capital buildups) that human analysts sometimes discount in strong growth narratives
Where it struggles: - Competitive positioning in fast-moving technology markets — the agent lacks the real-time industry context that experienced analysts develop through conference attendance, management conversations, and channel checks - Qualitative assessment of management quality — the agent can analyze capital allocation history but cannot assess the strategic clarity or execution capability of a management team the way an experienced analyst can - Black swan risks — novel regulatory changes, geopolitical disruptions, and paradigm-shifting technology developments are inherently difficult to model from historical financial data - Very early-stage companies — without a meaningful track record of financial statements, the agent's quantitative foundation is thin and the analysis is correspondingly less reliable
Integration with the Trading System
The fundamental analyst sits in a separate pipeline from our quantitative trading desk. Its reports inform position sizing and sector allocation decisions at the portfolio level rather than generating individual trade signals. The distinction matters — a fundamental analysis report is relevant on a weeks-to-months timescale, while our quantitative signals operate on hours-to-days timescales.
The practical integration: when our quantitative signals generate a consistent directional view on an equity or sector, the fundamental analysis layer acts as a veto or amplifier. A strong quantitative signal on an AAPL long, combined with a fundamental analysis that grades Apple A in quality and A+ in moat, gets larger position sizing than the same quantitative signal on a company with a C+ valuation attractiveness grade. The fundamental layer is a sizing overlay, not a trade trigger.
Takeaways
- Automated fundamental analysis can match or exceed human analysts on quantitative dimensions — financial ratio computation, trend analysis, FCF quality — but cannot fully replace the qualitative judgment of experienced analysts
- The Nvidia test case revealed the key limitation of DCF frameworks applied to hypergrowth companies — the agent's self-awareness about this limitation is actually one of its better features
- A proprietary grading system requires multiple iterations to be useful — compressed scores that cluster in a narrow range provide no decision-relevant information
- Fundamental analysis belongs at the portfolio level (position sizing, sector allocation) rather than the signal level (individual trade triggers) — it operates on a different time horizon than quantitative signals
- Coverage breadth is the primary advantage of automation — 20 companies per day versus 3–5 for a human analyst enables systematic screening of broad opportunity sets