finance-oracle

# finance-oracle — WallenT **WallenT is a learning-loop analyst, not a predictor and not a dashboard.** The product is the *loop*: every investing decision logs a falsifiable thesis **before** the outcome is known, then gets reviewed against reality, then yields a lesson. The portfolio is the context; getting good at the AI system is the goal. ไม่รีบ — prove the mechanism first, scale later. > Identity, hard gates, and the Analyst Output Contract live in [`CLAUDE.md`](CLAUDE.md). > The locked spec is [`spec.md`](spec.md). Current state is always in [`status.md`](status.md). ## The core loop ``` 1. THESIS (before outcome) → reason + assumptions + disconfirmers + Q-conditions + review_later ❌ no buy/sell verdict · ❌ no confidence% 2. TRACK (price ingest) → real market closes (yfinance) — never fabricated numbers 3. REVIEW (at review_later) → thesis held? assumption failed? disconfirmer appeared? action ONLY as a pre-declared Q-condition (logic, not a new verdict) 4. LESSON (at close) → P&L% + would-repeat? → feeds the next thesis ``` One deal proves the **mechanism**. Expectancy/strategy needs **N deals** — never claim a strategy works from one. ### Scripts (deterministic, rerunnable — real data only) | Script | What | Notes | |--------|------|-------| | `scripts/fetch_prices.py` | pull daily closes for the watchlist (yfinance) | market-date stamped (audit-safe), saves `data/prices/<market-date>.json` | | `scripts/review_deal.py` | review one deal: real current close → P&L%, Q-condition check | `--as-of` for reproducible reviews; emits numbers only, no verdict | | `scripts/expectancy.py` | expectancy report from **closed** `kind:paper` deals | excludes mechanism-proof replays; flags small-N as noise | ```bash python scripts/fetch_prices.py --ticker CPALL.BK python scripts/review_deal.py --ticker CPALL.BK --entry-date 2026-05-12 --entry-price 43.75 --qty 100 --as-of 2026-05-29 python scripts/expectancy.py ``` ## Discipline scaffold (`ψ/`) The original failure mode was "งานหายเงียบ 3 สัปดาห์" — empty state, lost work. The scaffold closes it: state lives in files, not in the agent's head. - `status.md` — single source of now / pending / next (updated every session) - `ψ/journal/` — one typed markdown file per deal (the loop's source of truth) - `ψ/inbox/handoff/` — end-of-session handoff (no handoff = session not closed) - `ψ/memory/` — retros / lessons (the loop applied to the agent itself) - `ψ/observations/` — auto tool-telemetry (gitignored) See [`ψ/README.md`](ψ/README.md) and [`ψ/journal/README.md`](ψ/journal/README.md). ## Hard rules (see `CLAUDE.md`) - **R0:** paper portfolio only — no real buy/sell/money/broker until TINE approves. - **Analyst Output Contract:** logged thesis only · no buy/sell verdict · no confidence%. - **API ได้อย่าด้นสด** — numbers come from a real API or they are not stated. No magic. ## Deferred subsystems (not the core — do not let them pull scope) - **Supabase** (`mytgtdurcpqefpkserae`) — currently a keyless read-only stub. Storage is **markdown/ψ first**; activate Supabase (key + schema + write path, **R1**) only when the loop is proven and typed expectancy needs SQL. Tables the dashboard expects: `prices`, `portfolio`, `transactions`, `watchlist`, `ai_insights`, `trade_journal`, `weekly_reviews`. - **Dashboard** (`dashboard/`) — a static shell built earlier (Codex). It reads sample data and is **frozen** until the loop produces real data; showing a dial without an engine is what stalled this repo before. Run locally only for reference: ```bash cd dashboard && python3 -m http.server 8091 # http://localhost:8091 ``` - **News / 17-factor / screener** — future thesis-generators that must feed the journal, not float free. Out of the core's first round. - **notebooklm-py** — `NO-GO` until LOKI security vet passes (holds real Google OAuth). ## Changelog See `CLAUDE.md` changelog. README rewritten 2026-05-31 to lead with the learning-loop core (was dashboard-first) per Judge repo review.