convez.ai
PluginFinanceiroAvançado
theodorklink

UBS Valuation Tool — Análise de Pair Trade

theodorklink/UBS_Challenge_2

Comparação quantitativa de pair trades com análise de spreads e sensibilidade de valuation

Ver no GitHub
InicialPluginScore · de 0 a 100

Ferramenta Bloomberg-style para comparação de pair trades, desenvolvida para o UBS Finance Challenge 2026. Analisa Siemens Energy (LONG) vs Sieyuan Electric (SHORT) através de cestas de pares globais e chinesas, integrando dados de yfinance, análise de spreads de valuation e sensibilidade de reset com comentários gerados por IA.

Instalação

$npx convez-ai theodorklink/UBS_Challenge_2

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 6 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Análise comparativa de valuações entre mercados (A-shares vs Europa)
  • Identificação de assimetrias estruturais em pares de ações
  • Cálculo de upside/downside baseado em reset de múltiplos
  • Geração de painéis Bloomberg-style para decisões de trading

Veredito

Ferramenta especializada e rigorosa para análise de pair trades com foco em dados estruturados e sensibilidade de valuation. Adequada para traders quantitativos e analistas que compreendem assimetrias contábeis entre mercados; exige integração manual de dados de segmento e contexto de hipótese direcional pré-definida.

README

A tradução para o português deste README ainda está na fila.

UBS Valuation Tool

Bloomberg-style pair-trade comparison tool for the 2026 UBS Finance Challenge. Built as a quantitative complement to a separate AI research module — the two tools are intentionally decoupled.

Pair Trade

  • LONG Siemens Energy (ENR.DE) — focus on the Grid Technologies segment
  • SHORT Sieyuan Electric (002028.SZ) — Chinese private-sector grid equipment maker

Direction is committed; the tool itself is data-only. The narrative emerges from how spreads are framed (premium / discount to peer median, trajectory slopes, valuation-reset sensitivity), never from tool-generated opinions.


What it produces

A single self-contained HTML file with three Bloomberg-style panels:

PanelContent
LeftSieyuan + China T&D peer basket (NARI, XJ, Pinggao, TBEA, China XD)
MiddleSiemens Energy + Global Power peer basket (GE Vernova, Schneider, Hitachi, Eaton, MHI) — with a separate Grid Technologies segment row populated from manually-transcribed IR data
RightCross-basket spread analysis: direct pair, basket-vs-basket median spread, relative positioning (RICH / NEUTRAL / CHEAP), valuation-reset sensitivity, policy exposure overlay, AI-drafted commentary

Plus a structured cross_market_caveats block surfacing A-share vs European reporting / accounting / fiscal-year-end / liquidity asymmetries — never papered over.


Sample output (from a recent run)

Live data, fetched 2026-04-28:

MetricSieyuan (002028.SZ)Siemens Energy (ENR.DE)
PriceCNY 195.25EUR 172.98
Mcap (USD bn)22.3173.4
YTD %+29.2%+40.9%
Forward P/E30.3x30.3x
vs peer median+59% (RICH)−17% (NEUTRAL)
EBIT margin18.0%4.0%
EBIT margin slope (3Y)+1.3 pp/yr+2.95 pp/yr

Cross-basket spread highlights:

  • Sieyuan sits at the 100th percentile of its China T&D basket on EV/Sales, EV/EBITDA, revenue growth, EBIT margin, and ROE — labelled RICH on every metric.
  • Siemens Energy sits broadly NEUTRAL within Global Power despite a materially lower margin base; consolidated forward P/E is below basket median (Global Power median is being pulled up by GE Vernova post-spinoff multiples).
  • Valuation-reset sensitivity: if Sieyuan's forward P/E reverts to its China T&D basket median, implied move is −37%. If Siemens Energy's reverts to its Global Power median, implied move is +20%. The spread is the trade.
  • Trajectory asymmetry: Siemens Energy's EBIT margin is expanding at +2.95pp/yr — roughly 2.2× the Global Power basket median slope. Sieyuan is at +1.3pp/yr against a flatter China T&D basket.

These observations are produced by the tool from yfinance + manually curated segment data — no qualitative judgement is applied. The asymmetry is structural in the data.


Architecture

ubs-valuation-tool/
├── src/
│   ├── data_sources/
│   │   ├── yfinance_adapter.py        # Stage 1a: live fundamentals
│   │   ├── siemens_energy_segments.py # Stage 1c: segment YAML loader
│   │   └── models.py                  # Provenance Pydantic models
│   ├── comparables/
│   │   └── comp_builder.py            # Stage 3: dual-basket comp tables
│   ├── spread/
│   │   └── cross_basket.py            # Stage 4: spread + reset sensitivity
│   ├── summary/
│   │   └── commentary.py              # Stage 5: Anthropic-drafted commentary
│   ├── render/
│   │   └── bloomberg_view.py          # Stage 6: 3-panel HTML
│   └── cli.py                         # Orchestrator
├── data/
│   ├── universes/
│   │   ├── china_td.yaml              # Sieyuan + 5 peer basket
│   │   └── global_power.yaml          # Siemens Energy + 5 peer basket
│   ├── raw/siemens_energy/            # Manual segment YAMLs (FY2024, FY2025-Q1/Q2)
│   └── policy/exposure_2026.yaml      # CBAM / FSR / Section 232 / IRA / Stock Connect
├── outputs/
│   ├── comp_tables/                   # Per-basket JSON (gitignored)
│   ├── spread_analysis/               # Cross-basket JSON
│   ├── valuation_drafts/              # Commentary drafts
│   ├── views/                         # Rendered HTML
│   └── lineage/                       # Audit-trail JSONs (planned)
├── scripts/
│   ├── smoke_test.py                  # Verify data sources
│   └── rerender.py                    # Re-render HTML from existing JSONs
├── tests/
└── CLAUDE.md                          # Project context + hard rules

Setup

git clone https://github.com/theodorklink/UBS_Challenge_2.git
cd UBS_Challenge_2

python3 -m venv .venv
source .venv/bin/activate           # Windows: .venv\Scripts\activate
pip install -r requirements.txt

cp .env.example .env
# Edit .env: paste ANTHROPIC_API_KEY (for the Commentary drafter, Stage 5)
#            FMP_API_KEY is optional (cross-validation, free tier)

Usage

Run the full pipeline

python -m src.cli run

Pipeline (5 stages, ~90s end-to-end):

  1. Build China T&D comp table — Sieyuan + 5 peers via yfinance
  2. Build Global Power comp table — Siemens Energy + 5 peers + Grid Tech segment row
  3. Cross-basket spread analysis — direct pair, basket-vs-basket, relative positioning, reset sensitivity, policy overlay
  4. Commentary draft — Anthropic Claude Sonnet 4.6, with strict safety perimeter
  5. Render Bloomberg HTML — single self-contained file

Output: outputs/views/sieyuan_vs_siemens_<timestamp>.html (open in any browser).

Skip the LLM step

python -m src.cli run --no-commentary

Re-render only (after fixing render bugs without re-fetching data)

python scripts/rerender.py

Preview server (local viewing)

cd outputs/views && python -m http.server 8001
# Open http://localhost:8001/

Deploy to Vercel (static-only)

The pipeline runs locally (yfinance + Anthropic + ~90s), and the rendered HTML is published as a static page on Vercel. The CLI automatically writes the latest run to public/index.html; pushing to GitHub triggers a Vercel auto-deploy.

One-time setup

  1. Sign up at https://vercel.com → "Continue with GitHub".
  2. "Add New" → "Project" → import theodorklink/UBS_Challenge_2.
  3. Framework Preset: leave on "Other" (Vercel auto-detects static via vercel.json).
  4. Root Directory: ./ (default).
  5. Build Settings: leave at defaults — vercel.json overrides them.
  6. Environment Variables: NONE NEEDED. The deploy is pure static; no API keys at runtime.
  7. Deploy.

You'll get a URL like https://ubs-challenge-2.vercel.app.

Per-presentation refresh workflow

# Generate fresh data + HTML locally
python -m src.cli run

# This writes BOTH:
#   outputs/views/sieyuan_vs_siemens_<timestamp>.html  (timestamped archive)
#   public/index.html                                  (overwritten with latest)

# Commit and push — Vercel auto-deploys in ~30s
git add public/index.html
git commit -m "Refresh Bloomberg view ($(date +%Y-%m-%d))"
git push

That's the whole workflow. No serverless complexity, no API blocking, no 60s timeouts.

Why static-only and not serverless?

  • yfinance is blocked from most cloud IPs (Yahoo blocks AWS/Vercel ranges)
  • Pipeline takes ~90s — Vercel Hobby tier function timeout is 60s
  • Anthropic call alone is ~45s — too tight for live serverless
  • Snapshot is meant to be reproducible — frozen at run time is a feature, not a bug, for a competition submission

Hard rules (encoded in code)

These are non-negotiable safety constraints, defined in CLAUDE.md:

RuleWhere enforced
No DCF generated by the LLMSystem prompt + postcheck
No fabricated tickersyfinance verify_ticker()
Every numeric output carries source + retrieved_at + source_urlPydantic Provenance mixin
Commentary drafter cannot output long, short, buy, sell, overweight, underweight, outperform, underperform, target price, fair value, we recommend, trade idea — case-insensitive postcheck blocks savesrc/summary/commentary.py:_FORBIDDEN_RE
Cross-market comparisons display the structured cross_market_caveats objectbloomberg_view._render_cross_market_warning()
Siemens Energy comparisons to Sieyuan default to Grid Technologies segmentcomp_builder._build_grid_tech_segment_row()
Missing data fails loudly — never imputedrenderer shows "—" / "data pending"

Narrative-supporting framing

The tool surfaces these as standard pair-trade analytics — never as opinions:

  • Premium / discount to peer median as the pair-trade-relevant valuation view
  • 3Y trajectory slope on margins (linear regression, annual obs; semi-annual for Sieyuan)
  • Valuation-reset sensitivity (revert each target to its basket median forward P/E)
  • Order quality composite (book-to-bill × backlog/revenue, where disclosed)
  • Policy exposure overlay with citations (CBAM, FSR, Section 232, IRA, Stock Connect)

If the data points the other way on any single metric, the tool shows it. The output is honest.


Data sources

SourceUsed forCost
yfinanceAll 12 basket members (incl. A-shares via .SS / .SZ)Free
akshareOptional A-share fallback (planned)Free
Anthropic Claude Sonnet 4.6Commentary drafter~$0.20-0.30 per run
Manual YAMLSiemens Energy Grid Tech segment data with page citations from IR PDFsManual entry
Financial Modeling Prep (optional)Cross-validation of yfinance valuesFree tier 250/day

Known limitations

  • Sieyuan reports semi-annually (CSRC permits) — trajectory slopes computed on 6 semi-annual observations rather than 12 quarterly. Auto-flagged in CompanyInfo.reporting_frequency.
  • A-share forward estimates are sparse on yfinance — Western analyst coverage is limited. Forward P/E shows "n/a — limited Western analyst coverage" for affected peers.
  • EV multiples are not segment-allocable — the Grid Tech segment row shows operating metrics (margin, growth, book-to-bill) only; EV-based multiples are blank with the explicit caveat.
  • Mitsubishi Heavy is a stretch peer — power business is one of three segments; flagged in basket validator output.

License

MIT — see LICENSE file (or use freely for academic / competition purposes).


Acknowledgements

Built for the 2026 UBS Finance Challenge submission. Sister project: ubs-ai-module (narrative research tool, Tavily + Claude, Vercel-deployed).

Ver o original no GitHub

Bom para

Relacionados