AlphaView

Your market, in focus. A local-first workspace for US stock discovery, portfolio monitoring, and reproducible strategy research.

Powered by Tentenai.com · 繁體中文使用指南

AlphaView brings market screening, transparent technical signals, and historical backtests into one focused dashboard. Your positions and research stay in a local SQLite database. Quotes are daily market data, not a real-time trading feed. Stale valuations retain their dates and show coverage warnings; unclosed future quotes and unavailable daily changes remain blank.

Market overview captured on September 5, 2026. The screenshot shows a saved research snapshot, not live quotes or a personal portfolio.

What you can do

• Discover candidates beyond your holdings. Choose a separate candidate pool of up to 250, 500, or 1,000 liquid US-listed equities, including ADRs, retrieved in market-cap order. Requested and accepted counts are shown separately.
• Refine and reuse screens. Filter by RSI, relative volume, RPS, price, and signal count; sort results, save named browser presets, and export every matching row to CSV.
• Follow daily changes. Compare distinct stored dates; new matches, exits, missing data, and universe membership changes are labeled separately.
• Keep a research journal. Save local notes and tags with version checks that prevent silent overwrites. Download a consistent local backup of saved workspace data and this browser’s supported settings.
• Understand every signal. Inspect the rules behind Turtle breakout, moving-average trend, RSI pullback, and relative-strength screens.
• Review historical results. Browse the latest 60 screening dates with separate market and personal-watchlist scopes.
• Track your own portfolio. Manage quantities and average costs, monitor allocation and unrealized P&L, and export positions to CSV. Preview a local CSV import before atomically merging additions and updates; positions omitted from the file are kept.
• Check concentration and co-movement. Review current holdings over 60 or 120 sessions, with pairwise sample counts and explicit quote coverage; incomplete valuations hide weights rather than implying a complete portfolio.
• Compare research candidates. Select 2–5 known stocks over 60, 120, or 252 trading-day intervals, with a shared starting date, adjusted-price percentage curves, and explicit coverage. Incomplete series remain listed with reasons and are not plotted. Export the submitted summary and daily percentage series with source, version and stale-context flags.
• Test a hypothesis. Run three long-only, single-stock strategies using prior-close signals and next-open execution with configurable capital, fees and dates, transparent diagnostics, and explicit transaction costs.
• Refresh after the close. Opt into a local end-of-day schedule for your watchlist or a 250/500/1,000-stock market pool; monitor its last attempt from the data page.
• Maintain local storage. Preview table usage and superseded scan snapshots, then explicitly confirm cleanup; the latest snapshot for each scope and date is retained.
• Inspect the data. See quote dates, calendar-aware missing sessions, malformed OHLC, provider failures, and job progress. Retry selected symbols or cancel a job. Failed downloads preserve previously stored data; corrupt histories produce no research signals. Missing or invalid chart observations remain gaps, and incomplete daily P&L is shown as unavailable with coverage counts.

The interface is in Traditional Chinese and uses a restrained dark design, keyboard search, responsive tables, and detailed stock views.

Local end-of-day schedule

Scheduling is off by default. In the data page, choose a scope and market-pool limit, enable the schedule, and save. The local server must remain running and the computer awake; this is not a cloud service. The scheduler checks approximately every 60 seconds and uses the latest XNYS session whose official close was at least 15 minutes ago, including holidays and early closes.

There is one automatic attempt per eligible session for the workspace, including failed, partially completed, cancelled, or interrupted attempts. Waking or restarting the server catches up only the latest eligible session; it does not replay every missed date. Retry failed symbols or refresh manually when needed. Disabling the schedule prevents future launches and does not cancel an active job; use the job’s cancel control separately. A busy workspace defers an unclaimed attempt until a later check. Changing the scope does not grant another automatic attempt for an already attempted session.

Quick start

Requirements: Python 3.12, uv, and Node.js 22.12+.

git clone https://github.com/tentenco/AlphaView.git
cd AlphaView
uv sync --locked --extra web --extra dev
npm ci --prefix web
npm run build --prefix web
uv run --extra web python -m alphaview.panel serve

Open http://127.0.0.1:8876. In 每日選股, choose 股票池上限 (250, 500, or 1,000), then run a market scan to download the candidate universe and daily history. Changing the selector alone does not download data or replace the current pool. Initial downloads depend on provider availability and connection speed.

An optional starter watchlist contains symbols only—no personal quantities, cost basis, or account balances:

uv run --extra web python -m alphaview.panel seed
uv run --extra web python -m alphaview.panel refresh

For market-wide discovery within the supported candidate pool:

uv run --extra web python -m alphaview.panel refresh --scope market --universe-limit 500

The CLI defaults to 250 candidates when --universe-limit is omitted. The option applies only to refresh --scope market. A targeted retry downloads only the selected symbols and retains the current pool and its configured limit.

In 資料管理 → 續跑行情更新, choose the existing market or personal pool. Resume retains its membership and skips only successfully sourced, current, structurally valid histories with consistent metadata. Missing, stale, failed, or inconsistent histories get one download attempt each; the selected pool is then rescanned even if every history was reusable. It does not rediscover market members or change the pool limit. Full refresh and selected-symbol retry remain separate actions.

Import your portfolio

In 我的持股, open 匯入 CSV, paste or select a CSV, review additions and changes, then confirm. The import stays local and never downloads prices or deletes positions omitted from the file. At most 100 rows may be imported, and the resulting personal list must remain within 100 symbols.

Required headers: symbol,shares,cost. Optional: name,sector. Symbols are normalized to uppercase. Use decimal points without currency signs or thousands separators; shares and costs must be finite numbers from 0 to 1,000,000,000. A positive share count requires a cost; a zero-share watchlist entry may leave cost blank. When optional headers are absent, existing names and sectors remain unchanged; new symbols use their code as the name and 自訂清單 as the sector.

This three-row example uses fictional quantities and costs, not current quotes or an actual portfolio:

symbol,shares,cost,name,sector
AAPL,2,100,Apple,Technology
MSFT,1.5,200,Microsoft,Technology
NVDA,0,,NVIDIA,Technology

Duplicate symbols or any invalid row block the entire import. Exported quote/computed columns are ignored with a preview warning. If holdings or the CSV change after preview, the commit is rejected; preview again to review the current changes. Database failures roll back the whole batch.

Research methodology

Screen	Core rule
Turtle breakout	Close above the prior 20-day high, an up candle, and volume confirmation
Trend following	Close > MA50 > MA200, with relative volume at least 1.2×
RSI pullback	Close above MA200, Wilder RSI(14) between 30 and 45, and an up day
Relative strength	Average percentile rank of 120-session returns ≥ 80 within the eligible same-date pool, and close ≥ 90% of its 120-session high (including today)

Relative volume divides current volume by the preceding 20-session simple average, excluding today. Moving averages include today and require their full 20/50/200 observations. Wilder RSI(14) seeds average gains/losses from 14 changes (15 closes), then uses (13 × previous average + current change) / 14; flat history is 50, gains without losses are 100. RPS uses ascending average ranks for ties divided by eligible peer count × 100; peers require 121 prices, a current same-date close, valid history and finite required metrics, with at least three peers. A score ≥ 80 does not necessarily select exactly 20% of a small or tied pool.

Indicators use dividend-adjusted daily prices. Portfolio valuation uses unadjusted daily closes and your entered holdings. Backtests use next-open execution, an initial $10,000, and a default 0.1% cost per side (configurable from 0 to 100 basis points). Cost applies symmetrically to executed notional: entry notional is available cash divided by (1 + cost rate), and exit cost is gross proceeds multiplied by the rate. This combines fees and slippage into a cost estimate without separately adjusting execution prices. Transaction amounts retain full precision; the UI formats them for display. The comparison is buy-and-hold of the same stock over the same period, without benchmark fees. An external SPY benchmark is not implemented; adding one requires separately validated, date-aligned adjusted data. Open positions are marked to the final close. CAGR, daily-return volatility, zero-risk-free-rate Sharpe, win rate, profit factor, and exposure include sample warnings; undefined statistics remain blank. Saved results are labeled stale when their input fingerprint or engine version changes. The current engine is alphaview-backtest-v5, which rejects nonfinite strategy-required metrics; older saved reports remain inspectable and are marked for recalculation.

Portfolio risk diagnostics describe current holdings, not historical portfolio performance. Valuation requires valid quotes for the latest completed XNYS session; incomplete coverage hides weights and concentration while labeling the available-value subtotal. Correlations use adjusted-close daily returns over 60 or 120 sessions, require at least 40 common observations per pair, and never bridge missing or invalid sessions. Stale series, zero-variance pairs, and insufficient samples remain unavailable. Pairwise date sets may differ; these results are not a covariance matrix, VaR, or a forecast.

Each new scan stores the scanner engine version and input revision captured with its price histories and membership. The current scanner is alphaview-scan-v1; changes to signal semantics require a new version under the scanner version policy. Publishing a scan does not change that revision; an input change during calculation prevents publication. Legacy snapshots without complete version provenance remain available for review but require recalculation before their signals can be paired with current charts. A workspace-wide input counter is deliberately conservative: a change outside the selected scope can also require a rescan.

Stock details show the source snapshot and withhold signals when its date or input revision does not match the viewed data. Signal-change reports compare strategy entries and exits only between snapshots with the same recorded input revision; unknown or different versions are marked unavailable, while membership additions and removals remain visible. Matching older snapshots can still be compared as historical records with an explicit warning. Open backtest reports are revalidated when the workspace data version changes.

The candidate comparison uses N+1 valid prices for N daily-return intervals and the same exact dates for every eligible stock. It does not shorten the period for recent listings, fill missing prices, or represent portfolio performance. Its adjusted-price series are not an independently reconstructed total-return dataset. Changes to inputs or calculation versions mark an open comparison for recalculation.

A historical screen recomputes the current candidate universe on earlier dates; it is not a point-in-time index membership dataset and has selection/survivorship bias. Relative strength is a pool-relative ranking, not a rank across every US stock. Expanding or changing the universe can change RPS even when a stock’s price history is unchanged; compare the recorded universe as well as the signal. Research signals are not automated orders.

Architecture

alphaview/panel/    FastAPI API, market ingestion, SQLite storage, research engine
web/               React + TypeScript + Vite dashboard
tests/             Deterministic tests with isolated databases
docs/              Product research and development documentation

Background polling checks a lightweight workspace revision and public job state. Full overview data is fetched only after a data revision changes, on initial load, or on an explicit refresh. Transactional SQLite counters detect same-timestamp edits; active jobs poll every four seconds, visible idle pages every thirty seconds, and hidden pages pause until focus.

The repository also retains an independent optional A-share CLI in main.py and supporting Python modules. It is not invoked by the US-stock web panel. See local setup for the separate command and configuration.

Development

# Backend
uv run --extra web python -m alphaview.panel serve

# Frontend dev server (API proxy to port 8876)
npm run dev --prefix web

# Validation
uv run --extra web --extra dev pytest -q
npm test --prefix web
npm run format:check --prefix web
npm run build --prefix web

API documentation is available locally at http://127.0.0.1:8876/docs.

Data and privacy

• Daily market data: Yahoo Finance through yfinance.
• Market universe: US region, Nasdaq/NYSE, USD equities; market cap ≥ $2B, price ≥ $5, three-month average daily volume ≥ 200,000; a selectable maximum of 250, 500, or 1,000 in market-cap order. Multiple share classes and ADRs may appear. This remains a filtered subset, not all US stocks.
• Discovery uses sequential pages of at most 250 results through yfinance’s documented offset and size parameters. A complete discovery replaces the pool and its metadata together; a page failure or cancellation before publication preserves the prior pool. Provider totals, identity filtering, and deduplication can produce fewer accepted symbols than requested. Accepted members must also report a finite numeric market cap of at least $2B; missing or malformed market caps are excluded. Multi-page responses are not an atomic market snapshot.
• Quote coverage: invalid latest closes are not replaced by older prices; missing adjacent sessions make daily changes unavailable. Portfolio value and unrealized P&L may be partial subtotals with explicit coverage warnings. Allocation weights require valid valuations for all holdings on the latest completed session; mixed-date or incomplete coverage leaves weights unavailable. Historical stock views apply current quantities and costs to the selected date’s price, not a historical holdings ledger.
• Workspace: data/panel.db; override with PANEL_DB_PATH.
• Databases, personal holdings, .env, logs, and local review artifacts are excluded from Git.
• The server binds to loopback. Authentication, multi-user authorization, broker execution, and public hosting are outside the current local-workspace design.
• Respect the data provider's terms and permitted usage. AlphaView does not grant redistribution rights to downloaded market data.

Local backups

The data workspace can download a ZIP containing alphaview.db, research-notes.json, browser-settings.json, and manifest.json. The database is copied using SQLite’s online backup API from a pinned read snapshot, including committed WAL data. Notes and the manifest are derived from that same copy; normal workspace writes may continue while it is created. Only one backup is prepared at a time, with a bounded creation deadline and temporary-file cleanup.

The ZIP includes saved positions, cost basis, notes, market data, scans, backtests, job records, and validated browser screening presets/universe-limit preferences. It is not encrypted and stays a user-triggered local download; store it somewhere appropriate for personal financial information. Unsaved note drafts, application code, environment variables, and unrelated browser storage are excluded. The manifest records versions, table counts, and SHA-256 hashes for integrity checks.

A backup is not a restore operation. AlphaView does not automatically import the ZIP or overwrite the active database. Do not copy just a live .db file while WAL writes are active, and do not replace a running workspace’s files. A restore workflow remains separate work. The read-only backup preflight CLI validates supported ZIP structure, hashes, SQLite integrity and known schemas without restoring data; unknown schemas require separate compatibility review. CSV import covers portfolio rows only and is not a substitute for a full workspace backup.

SQLite Online Backup API · Python Connection.backup

Project direction

AlphaView is developed by Tenten as a practical research workspace: discover, inspect, compare, and keep a clear record of the evidence behind a decision. Feature claims in this README describe implemented behavior; experimental work and follow-up tasks are documented separately. Read the official-source competitor benchmark and Harness workflow.

Powered by Tentenai.com

Storage maintenance and editing conflicts

The data page previews logical payload bytes, row counts, physical database/WAL sizes, and reusable SQLite pages. Logical payload bytes count the listed text fields in UTF-8; they are not the table’s total disk footprint. Physical file measurements can change as other work commits.

Cleanup requires a current preview and explicit confirmation. It removes only superseded scan snapshots, keeping the latest ID for every (scope, date); positions, notes, bars, backtests, jobs, and scheduler attempt records stay intact. If the snapshot set changes or another data job is running, reload the preview. A local backup download is available before cleanup. There is no automatic cleanup or VACUUM: freed pages can be reused, so deleting rows need not shrink the file.

The portfolio editor sends expected_updated_at with each save and rejects conflicting edits with HTTP 409; reopen the editor to review the latest record before saving. New entries send a null expectation and also detect a concurrent addition. Legacy API clients that omit this field retain last-write-compatible behavior and do not receive this conflict protection.

Data recovery options and their validation requirements are documented in the data-provider evaluation.

Next-Harness proposals are separate from shipped features: data recovery and benchmark, workspace restore and research continuity, and ledger and methodology acceptance cases.