@viviennaBTC: The core argument of this article complements some foundations of my previous "cyclical stock selection method". Our shared view is that the true moat is not the AI model, nor the prompts from celebrity investors, but the underlying "wiring" — how you connect real data, disciplined mathematics, and agents together. The ar…

The core argument of this article complements some foundations of my previous "cyclical stock selection method". Our shared view is that the true moat is not the AI model, nor the prompts from celebrity investors, but the underlying "wiring" — how you connect real data, disciplined mathematics, and agents together. The article's methodology for hedge fund stacks is divided into four layers: L4 — AI Agents (the visible performance) These are the repos flooding social media: virattt/ai-hedge-fund (14 celebrity investor personas + 4 quantitative analysts + risk control + portfolio manager, total 19 agents) and the more "adult" TauricResearch/TradingAgents (replacing celebrities with real trading desk roles, core being bull vs bear adversarial debate). The author explicitly states: this layer is "theater", meant for show; models are never the moat. L3 — Reading Financial Reports (read only what changed) Using edgar-crawler to extract Risk Factors and MD&A into clean JSON, reading only the year-over-year content that actually changed. L2 — Legal/Financial Fraud Screening (deciding what's worth reading) Use FinanceToolkit to run Beneish M-Score, Altman Z-Score, Piotroski F-Score, and Sloan accruals — four scores calibrated by decades of real fraud/bankruptcy cases from academia. L1 — Real Data (models are never allowed to fabricate numbers) SEC EDGAR provides free access to all publicly traded US companies' filings; Edgartools parses 10-K, 8-K, Form 4, 13F into clean Python, and comes with an MCP server that can be directly hooked into Claude Code / Claude Desktop as a tool, letting Claude read real documents instead of "making up plausible-sounding numbers" from training memory. (I deeply resonate with this — even top-tier AI like CC Opus 4.8 suffers from severe data hallucination.) It can be seen that this article is about "how to wire things correctly without errors", while my methodology is "after wiring, what direction to think in". The former addresses garbage in, the latter addresses even with clean data, what's the right question. Both are essential together. Comparison of the Two Frameworks The most interesting part: both are layered architectures, but stacked in opposite directions. My cyclical stock selection is top-down: Macro environment (Layer A) → Equity four-season cycle (Layer B) → Individual stock selection (Layer C). First use the macro regime to "open or close the gate" for the entire market, then zoom down to individual stocks. This article is bottom-up: Real data (L1) → Legal screening (L2) → Read diff (L3) → Agent writes memo (L4). First ensure underlying data is not fabricated, then layer by layer apply intelligence where it matters. One is an epistemological framework (in what macro season, what logic to use for stock selection), The other is an engineering stack (how to correctly connect data, math, and models without errors). But they are highly consistent on several deep values: 1. Shared core belief: Models/intuition are not the moat; underlying discipline is. The article says "models are never the moat; the data you connect is"; in my methodology I repeatedly emphasize analytical honesty, explicitly labeling look-ahead bias and backtesting limitations. It's essentially the same thing — hostility toward "answers that sound right but lack factual support". The article's line "a flag whispers go look, it doesn't shout fraud" shares the same restraint as my handling of forensic signals as "probability hints rather than evidence". 2. My Layer C is exactly where this article's L2 lands. My Layer C selectively absorbs Peter Lynch's toolbox (six company types mapped to four seasons, cyclical stock P/E reversals, PEG), but explicitly rejects his anti-macro philosophy. The four forensic scores in the article's L2 (Beneish/Altman/Piotroski/Sloan) are purely individual stock quality/fraud screening, naturally belonging to Layer C's stock selection phase. They can insert an automatic gate between "selecting candidate stocks" and "deep-diving into financial reports": First use accruals, F-Score to filter out or flag names in the candidate pool where "profits are outpacing cash / fundamentals are deteriorating", then manually deep-read. This aligns with my work habits of "careful distinction between buckets, batch processing with scripts". 3. My Layer A is exactly what TradingAgents' "macro analyst / bull-bear debate" tries but fails to do well. The article's agent layer has a "news analyst on the macro", but it lacks regime-gating discipline — it reads macro, but does not switch the entire stock selection logic on/off based on macro seasons. My Layer A (drawing from Tao Chuan's macro analysis + Oppenheimer's "The Long Good Buy" partition of cycle phases) precisely provides what this agent stack lacks: An upper-level gate that decides "whether and what type of stocks to pick now". In other words, my methodology can serve as the "L5 / regime gate" for this agent stack, sitting above L4.