Global Index Valuation Agent

Where is the world's equity value — and where's the opportunity? A cost-effective research agent that ranks ~90 global market indices by value and fundamental growth, surfaces the handful of calls worth your attention, and learns from whether its past calls actually played out — so you spend a minute, not an afternoon.

 ·  ▶ Live demo · Architecture · Data-ingestion agent · Write-up: the decisions and what broke

[!WARNING] For research/education only — not investment advice. The rankings are experimental and unvalidated — a first stock-level walk-forward backtest has run (2026-07-22) with an encouraging but not yet statistically significant signal (see docs/STATUS.md for the numbers). Do your own research. See DISCLAIMER.md.

---

What it is

Most equity tools are either data terminals you dig through (Bloomberg, Koyfin, TIKR) or opaque ratings that never show whether they worked (Simply Wall St, Stockopedia). This is neither. It is an agent that does the research for you and shows its work:

• It sources every major index, computes valuation + fundamental growth, and ranks markets within their peer group (countries vs countries, sectors vs sectors).
• It surfaces the few things that matter — cheapest value, highest fundamental growth, and the GARP sweet spot (cheap and growing) — at the top.
• It records every call against a fixed date and grades itself against what the market actually did, so the track record is public and falsifiable.

Try it: https://global-index-valuation-agent.vercel.app (the strategist read, the insight cards, and the Value × Growth map are the one-minute answer).

Why it's different (the wedge)

1. Top-down + bottom-up in one view. "Which market is cheap & growing" and (Phase 2) "which stock within it" on one globally-comparable scale. Most tools are stock-only or single-market.
2. A 1-minute answer, not a terminal. Insights rise to the top; filters are there if you want to dig.
3. A transparent, self-validating track record. It grades its own calls in the open.
4. Cost-engineered to run for ~$0. The expensive model never touches the raw data.

---

How it works

flowchart LR
  SRC["Free data sources<br/>(ETF holdings, prices, fundamentals)"] --> ENG
  subgraph ENG["Engine — deterministic, free (weekly)"]
    direction TB
    F["Fetch + normalize"] --> M["Score: value · growth · momentum · opportunity<br/>(within each peer group)"]
    M --> S["Surface the few insights that matter"]
    M --> L["Record predictions → grade past calls → auto-tune weights"]
  end
  ENG --> J["dashboard_data.json (published snapshot)"]
  J --> UI["Static dashboard on Vercel<br/>insights · Value×Growth map · scoreboard"]
  LLM["LLM (optional, cheap tier)<br/>tags + strategist brief"] -.-> S

The deterministic spine. Fetching, parsing, computing P/E and other metrics, scoring, ranking, and grading are all plain Python — free, exact, auditable. The LLM is optional and only ever narrates the final ≤90-row scoreboard. With no API key the product still runs end-to-end for $0 (deterministic fallbacks fill in the text).

The scores (all ranked within a peer-group kind, so adding US sectors never distorts the country comparison):

Score	Question it answers
Value	Is it cheap? (P/E, P/B, P/S, yields — with value-trap guards)
Growth	Are the underlying fundamentals growing? (revenue + earnings growth of the top holdings + forward estimates — not price momentum)
Opportunity (GARP)	Cheap and growing and not overvalued? (value 40% · growth 40% · momentum 12% · mean-reversion 8%)

Two feedback loops make it an agent, not a report:

• Market feedback — every run records a prediction; later runs grade it (rank-IC, hit-rate) and the auto-tuner nudges the weights toward what worked.
• User feedback — pin / dismiss / rate, which re-weights what surfaces next.

Cost model — right model for the right task. Deterministic work is free; LLM work is routed by frequency × stakes: cheap/owned models (self-hosted Qwen, GLM-4-Flash, DeepSeek) for frequent customer-facing calls, frontier models only for the rare weekly synthesis. Full design in docs/ARCHITECTURE.md.

---

Jobs and agents

We keep the terminology precise (see docs/AGENTS.md):

• Jobs = deterministic Python (no LLM) — they do the work, run frequently, ~free.
• Agents = LLM-powered review that improves the jobs — infrequent. (None are active yet — no model is configured.)

Jobs running today:

1. Valuation / scoring (engine/) — sources indices, scores them, surfaces insights, grades itself, publishes the dashboard. (weekly)
2. Data pipeline (engine/datapipeline.py) — one sequenced job: probe sources → ingest fundamentals (SEC EDGAR) → tag stocks by sector → validate + auto-fix the KPI catalog → data-quality checks → recalibration check. Validation runs right after ingestion. (daily) See docs/DATA_INGESTION.md.

This is the beginning of the README. Read the rest at the original. Read the full README ↗