Fresta Finance — S&P 500 Structural Entropy Ranking

Applies the Fresta Lens Framework to financial markets.
Ranks every S&P 500 company by structural entropy across five orders.

---

What It Does

Most financial risk models score companies in isolation — debt ratio, volatility, margins — and aggregate them additively. This misses critical layers:

• How metrics interact — a company with great margins but extreme supply chain concentration may be more fragile than its individual scores suggest
• What the system inherits — sector dependencies, macro stress, and infrastructure concentration that no balance sheet will show you
• Where the roots are — geographic and structural roots of the supply chain, single points of failure, and circular dependencies
• What politics decides — tariff exposure, regulatory risk, geopolitical concentration, and policy dependency

Fresta Finance computes a five-order entropy score for every S&P 500 company:

E_unified = 0.40 × E_total + 0.35 × E_tree + 0.25 × E_political

where E_total = E0 + E_upstream + E_inherited

Order	Script	What it captures
E0 (1st)	fresta_finance.py	Local financial health — margins, debt, volatility
E_upstream (2nd)	fresta_finance.py	Propagated entropy through sector dependency graph
E_inherited (3rd)	fresta_finance.py	Systemic stress from macro, rates, and concentration
E_tree (4th)	fresta_tree.py	Supply chain dependency tree — root diversity, SPOFs, circular dependencies
E_political (5th)	fresta_political.py	Political-economic risk — tariffs, regulation, geopolitical concentration

Lower score = less structural entropy = more resilient company.

---

Pipeline

The three scripts run in sequence. Each builds on the previous output:

fresta_finance.py   →  sp500_entropy_ranked.csv
        ↓
fresta_tree.py      →  sp500_tree_analysis.csv
        ↓
fresta_political.py →  sp500_unified_report.html

---

Usage

Requirements

pip install pandas numpy yfinance requests anthropic

For orders 4 and 5, you need either a local LLM via LM Studio (API-compatible, any model) or the Anthropic API. Set your key if using Claude:

set ANTHROPIC_API_KEY=sk-ant-...    # Windows
export ANTHROPIC_API_KEY=sk-ant-... # Linux/Mac

Run (automated)

run_fresta.bat   # Windows — runs all three scripts in sequence

Run (manual)

python fresta_finance.py    # ~15–30 min (downloads market data)
python fresta_tree.py       # ~10–15 min with Claude API / ~2–4h with local LLM
python fresta_political.py  # ~10–15 min with Claude API / ~1–2h with local LLM

The first run of fresta_finance.py downloads price data for ~500 tickers. Results are cached locally for 7 days. Tree and political analyses are cached per company as individual JSON files — runs can be interrupted and will resume exactly where they stopped.

---

Output

File	Description
output/sp500_entropy_ranked.csv	1st–3rd order financial scores
output/sp500_entropy_report.html	Interactive 3-order financial report
output/sp500_tree_analysis.csv	4th order supply chain tree metrics
output/sp500_unified_report.html	Full interactive 5-order unified report

Open sp500_unified_report.html in any browser. It is fully self-contained — sortable by any column, filterable by text or single-root risk flag, with rank change indicators showing which companies look structurally different when supply chain and political risk are added.

---

How Scoring Works

E0 — 1st Order (Local Entropy)

Three blocks, each scored as a percentile rank across all S&P 500 companies:

Block	Metrics	Weight
Recycling capacity	Profit margins, operating margins, FCF, ROE	45%
Structural fragility	Debt/equity, current ratio	35%
Noise/stress	Annualised volatility, max drawdown	20%

E_upstream — 2nd Order (Propagated Entropy)

Builds a weighted dependency graph from sector groupings, market cap, and price correlations (threshold: r > 0.70). Runs iterative propagation (α = 0.30, 2 iterations): each company inherits a fraction of its dependencies' entropy. Adds a Herfindahl concentration penalty and a circular dependency penalty.

E_inherited — 3rd Order (Systemic Stress)

Models macro overlays (SPY, VIX, RATES) as virtual nodes. Financial sector companies inherit rate stress; communication companies inherit macro stress. Adds a saturation penalty when infrastructure entropy exceeds a critical threshold (65.0).

E_tree — 4th Order (Supply Chain Tree)

For each company, an LLM extracts the real dependency tree:

• Customers — who buys, with sector and geography
• Suppliers — who sells, with type, geography, and number of alternatives
• Geographic roots — countries/regions weighted by actual exposure
• Single points of failure — critical dependencies with no viable alternative
• Circular dependencies — feedback loops in the supply chain

Key metrics computed from the tree:

Metric	What it measures
root_diversity	Shannon entropy of geographic root distribution
single_root_risk	True if any single root > 60% of supply chain weight
supplier_hhi	Herfindahl index of supplier concentration
customer_hhi	Herfindahl index of customer concentration
spof_count	Number of single points of failure
circular_count	Number of circular dependency loops

Adaptive depth: before calling the LLM, each company is assigned a complexity score (0–10) from its sector, market cap, and business description. Simple companies (local utilities, domestic retailers) get a short shallow prompt and small token budget (~350 tokens). Complex companies (semiconductors, global tech, defense) get a deep structured prompt (~850 tokens). This makes the analysis 3–5× faster on average without sacrificing quality where it matters.

Anti-hallucination: all LLM outputs are validated against a strict JSON schema. Invalid or out-of-range fields are replaced with conservative defaults — never invented. Up to 3 retry attempts per company before falling back to safe defaults. Results are saved as individual validated JSON files and never re-analysed.

E_political — 5th Order (Political-Economic Risk)

Two-stage analysis. First, sector-level political risk is assessed once and cached (~12 LLM calls total). Then each company is scored against that baseline using its tree data as context:

Component	What it captures
tariff_risk	Exposure to import/export tariffs and trade wars
regulatory_risk	Antitrust, data privacy, environmental, financial regulation
geopolitical_risk	Conflict zones, political instability, sanctions exposure
china_exposure	Supply chain or revenue dependence on China
taiwan_exposure	Supply chain dependence on Taiwan (critical for semiconductors)
policy_dependency	Reliance on government subsidies, contracts, or favorable policy

The company-level score adjusts the sector baseline using tree data as context — a semiconductor company with TSMC as its sole supplier is scored worse than a peer with diversified Asian manufacturing, even within the same sector.

---

Key Findings (S&P 500, March 2026)

The five-order analysis reveals structural fragilities invisible to financial-only models.

Most resilient (unified): BRK-B, CBOE, CME, BLK, V — financial market infrastructure with diversified roots, regulatory moats, and no geographic concentration.

Biggest rank falls when supply chain and political risk are added:

Ticker	Financial rank	Unified rank	Δ	Why
AMD	#3	#495	−492	100% TSMC dependency, Taiwan existential risk
SMCI	#331	#503	−172	NVIDIA allocation + Taiwan components + accounting risk
ALB	#456	#502	−46	Chile Atacama lithium concentration, China processing monopoly
QCOM	top quartile	bottom quartile	—	Fabless + ~60% China revenue exposure

Biggest rank improvements:

Ticker	Why
VICI	Domestic real estate, no supply chain concentration
NTRS, AMP, CINF	Domestic financial services, no physical supply chain risk
BRK-B	Most diversified company in the index — confirmed at all five orders

The NVDA paradox: NVDA ranks #1 by financial health (E0 = 13.6 — best in the entire index). But it carries one of the highest E_tree scores. The most financially resilient company is simultaneously one of the most structurally fragile — a cash machine whose entire product depends on a single foundry (TSMC) in a single geopolitical hotspot (Taiwan Strait). The unified score captures what financial analysis cannot.

---

Theoretical Grounding

This tool is a financial application of the Fresta Lens Framework — a five-volume theoretical work (~500 pages) on structural evaluation, entropy, and system coherence.

• Full framework: doi.org/10.5281/zenodo.18251304
• EDGE (domain evaluation tool): github.com/EviAmarates/fresta-edge

---

License

MIT

Support

If this is useful: ko-fi.com/tiagosantos20582