stock-checker

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# Stock Checker

A bun-workspaces monorepo that screens US equities with an institutional-flow
signal engine, visualizes them in a web UI, and validates every strategy change
with a backtest.

| Package | What it is |
|---|---|
| `packages/core` | Signal engine, backtest, and CLI (`predict` / `learn` / `optimize` / `backtest`) |
| `apps/api` | Fastify API server (screener, ticker detail, OHLCV) — port 5101 |
| `apps/web` | Next.js 16 screener UI (candlestick + Gaussian Channel band charts, portfolio, light/dark) — port 5100 |

| Screener | Ticker detail (Gaussian Channel band) |
|---|---|
| ![Equity screener table with composite scores, signals, and detected chart patterns](https://github.com/gracefullight/stock-checker/raw/HEAD/docs/images/screener.png) | ![TSLA detail page: candlestick chart with SMA 20/50/200, Bollinger Bands, and trend-colored Gaussian Channel band](https://github.com/gracefullight/stock-checker/raw/HEAD/docs/images/ticker-detail.png) |

## Signal philosophy

The engine follows the principles in [docs/TRADING_PRINCIPLES.md](https://github.com/gracefullight/stock-checker/blob/HEAD/docs/TRADING_PRINCIPLES.md):
price, volume, VWAP, moving averages, liquidity, relative strength, and earnings
revisions over oscillator soup.

- **Trend regime** — Gaussian Channel (green = uptrend, red = downtrend) gates all buys.
- **Institutional flow score** — relative strength vs SPY and the sector ETF,
  VWAP accumulation, breakout volume, dollar-volume liquidity, earnings revisions.
- **Strong-leader pullback entry (주도주 눌림목)** — BUY only when a name that
  is STRONGLY outperforming both the market and its sector (`rs ≥ 0.7`) pulls
  back below its 50-day SMA on a calm bar that closed in the bottom 20% of its
  range (`ibs < 0.2`) with real participation. Backtested (8y incl. the 2020
  crash and the 2022 bear, **546 tickers**, real pipeline, net of a 10bps
  round-trip cost): **60.4% 5-day win rate / 1.28 reward-risk / N=225** vs the
  51.3% / 1.05 ungated baseline (z≈2.6, p≈0.004), every entry year ≥ 50%.
- **SELL = exit discipline, not a downside prediction** — distribution-day
  SELLs are suppressed inside intact uptrends and only fire when the trend
  itself is broken.
- Classic indicators (RSI, Stochastic %K, Bollinger, Donchian, Williams %R,
  MACD, ATR, volume ratio, Fear & Greed) are still computed and displayed, but
  they season the score rather than drive it.
- Volatility-adjusted risk levels per signal: 1.5×ATR stop loss, 2× reward
  take profit, trailing stop that activates after a 0.5×ATR move.

## Validated results

8-year window (entry years 2019–2026, incl. the 2020 COVID crash and the 2022
rate-hike bear), **546-ticker** diversified universe (large + mid + small cap,
all 11 sectors), fixed 5-day exit, evaluated through the real pipeline, **net
of a 10bps round-trip transaction cost** (a "win" = profitable after costs).
Full context and hard-won validation rules in
[docs/TRADING_PRINCIPLES.md](https://github.com/gracefullight/stock-checker/blob/HEAD/docs/TRADING_PRINCIPLES.md).

| Config | WR (5d) | R/R | N | Avg ret/trade |
|---|---|---|---|---|
| **Shipped gate** (`rs≥0.7` + `ibs<0.2` + `scr<400` + below-50d) | **60.4%** | **1.28** | 225 | 1.08% |
| Legacy V7 gate (`rs≥0.5`, `ibs<0.3`, `scr<380`) | 56.3% | 1.32 | 476 | 0.85% |
| + SPY kill-switch + 200d stage (NOT shipped — hurts at scale) | 55.7% | 1.29 | 230 | 0.79% |
| V5 institutional baseline (no quality gate) | 51.3% | 1.05 | 84,541 | 0.20% |

Shipped gate by entry year (WR / N): 2019 66%/41 · 2020 62%/21 · 2021 61%/33 ·
2022 55%/40 · 2023 52%/29 · 2024 55%/22 · 2025 65%/20 · 2026 74%/19 — every
year ≥ 50%, both bear regimes included. Train ≤2024: 58.6% / 1.19 (N=186) ·
holdout ≥2025: 69.2% / 2.35 (N=39). Significant vs baseline (z≈2.6, p≈0.004).

**Market-cap scope.** This is a **large-cap strategy**: ~90% of gate signals
fire on $10B+ names (the `atr%<3.5` calmness and `rs≥0.7` leadership profile
rarely matches smaller names). On mid caps the WR edge disappears (≈52% vs a
50.5% mid baseline, N=25 — winners run bigger but no hit-rate edge); the
ungated small-cap pullback baseline is outright negative (46% WR, −0.07%/trade).
Trade it on liquid large caps only.

**Falsification record.** On the original 122-ticker growth-heavy universe the
same family printed up to **71.7% WR / 1.75 R/R (N=46)** — expanding the
universe 4.5× collapsed it. The 70%+ readings were small-N universe artifacts,
not edge ("universe shapes conclusions"). The SPY-Gaussian market kill-switch
helped at 122 tickers and consistently *hurt* at scale, so it ships as an
optional gate param, off by default. Remaining caveats: as-of-today universe
and cap tiers (survivorship bias), no live forward track record.

## Usage

Tooling is managed by [mise](https://mise.jdx.dev); tasks wrap every common
operation (run `mise tasks` to see them all).

```bash
mise install        # pin runtimes (node 24, bun)
bun install         # install workspace deps + git pre-commit hook
mise run dev        # API (5101) + Web (5100) dev servers in parallel
```

### Environment (optional)

| Variable | Effect |
|---|---|
| `TIINGO_API_KEY` | Enables the [Tiingo](https://www.tiingo.com) daily-OHLCV fallback when Yahoo is rate-limited or down (free tier: 1,000 req/day). Without it, OHLCV degrades to empty on Yahoo failure. |
| `SLACK_WEBHOOK_URL` | Slack notification for BUY/SELL opinions from `predict`. |

### CLI (packages/core)

```bash
# Daily prediction for a ticker list (default command)
mise run predict -- --ticker=TSLA,PLTR --sort=asc

# Slack notification for BUY/SELL opinions (either form)
SLACK_WEBHOOK_URL=https://hooks.slack.com/services/XXX mise run predict -- --ticker=TSLA,PLTR
mise run predict -- --ticker=TSLA,PLTR --slack-webhook=https://hooks.slack.com/services/XXX

# Strategy validation & tuning
mise run backtest        # version comparison, goal search, SELL validation
mise run backtest -- --cost-bps=20   # vary the round-trip cost (default 10bps)
mise run backtest -- --quick         # stop after version comparison + gate tuning
mise run optimize        # parameter optimizer (positional symbol, e.g. TSLA)
mise run learn           # learn from prediction feedback
```

Each `predict` run appends rows to a monthly CSV in `packages/core/public/`
(e.g. `stock_data_202511.csv`), tickers in alphabetical order (`--sort=desc`
reverses).

### Quality gate

```bash
mise run ci          # lint → typecheck → test → build
```

## Automation

- `.github/workflows/daily-data.yml` — runs `predict` after US market close and
  auto-commits the monthly CSV.
- `.github/workflows/weekly-optimize.yml` — weekly parameter optimization,
  results uploaded as a build artifact.

Stock Checker

A bun-workspaces monorepo that screens US equities with an institutional-flow signal engine, visualizes them in a web UI, and validates every strategy change with a backtest.

Package	What it is
`packages/core`	Signal engine, backtest, and CLI (`predict` / `learn` / `optimize` / `backtest`)
`apps/api`	Fastify API server (screener, ticker detail, OHLCV) — port 5101
`apps/web`	Next.js 16 screener UI (candlestick + Gaussian Channel band charts, portfolio, light/dark) — port 5100

Screener	Ticker detail (Gaussian Channel band)

Signal philosophy

The engine follows the principles in docs/TRADING_PRINCIPLES.md: price, volume, VWAP, moving averages, liquidity, relative strength, and earnings revisions over oscillator soup.

Trend regime — Gaussian Channel (green = uptrend, red = downtrend) gates all buys.
Institutional flow score — relative strength vs SPY and the sector ETF, VWAP accumulation, breakout volume, dollar-volume liquidity, earnings revisions.
Strong-leader pullback entry (주도주 눌림목) — BUY only when a name that is STRONGLY outperforming both the market and its sector (rs ≥ 0.7) pulls back below its 50-day SMA on a calm bar that closed in the bottom 20% of its range (ibs < 0.2) with real participation. Backtested (8y incl. the 2020 crash and the 2022 bear, 546 tickers, real pipeline, net of a 10bps round-trip cost): 60.4% 5-day win rate / 1.28 reward-risk / N=225 vs the 51.3% / 1.05 ungated baseline (z≈2.6, p≈0.004), every entry year ≥ 50%.
SELL = exit discipline, not a downside prediction — distribution-day SELLs are suppressed inside intact uptrends and only fire when the trend itself is broken.
Classic indicators (RSI, Stochastic %K, Bollinger, Donchian, Williams %R, MACD, ATR, volume ratio, Fear & Greed) are still computed and displayed, but they season the score rather than drive it.
Volatility-adjusted risk levels per signal: 1.5×ATR stop loss, 2× reward take profit, trailing stop that activates after a 0.5×ATR move.

Validated results

8-year window (entry years 2019–2026, incl. the 2020 COVID crash and the 2022 rate-hike bear), 546-ticker diversified universe (large + mid + small cap, all 11 sectors), fixed 5-day exit, evaluated through the real pipeline, net of a 10bps round-trip transaction cost (a "win" = profitable after costs). Full context and hard-won validation rules in docs/TRADING_PRINCIPLES.md.

Config	WR (5d)	R/R	N	Avg ret/trade
Shipped gate (`rs≥0.7` + `ibs<0.2` + `scr<400` + below-50d)	60.4%	1.28	225	1.08%
Legacy V7 gate (`rs≥0.5`, `ibs<0.3`, `scr<380`)	56.3%	1.32	476	0.85%
+ SPY kill-switch + 200d stage (NOT shipped — hurts at scale)	55.7%	1.29	230	0.79%
V5 institutional baseline (no quality gate)	51.3%	1.05	84,541	0.20%

Shipped gate by entry year (WR / N): 2019 66%/41 · 2020 62%/21 · 2021 61%/33 · 2022 55%/40 · 2023 52%/29 · 2024 55%/22 · 2025 65%/20 · 2026 74%/19 — every year ≥ 50%, both bear regimes included. Train ≤2024: 58.6% / 1.19 (N=186) · holdout ≥2025: 69.2% / 2.35 (N=39). Significant vs baseline (z≈2.6, p≈0.004).

Market-cap scope. This is a large-cap strategy: ~90% of gate signals fire on $10B+ names (the atr%<3.5 calmness and rs≥0.7 leadership profile rarely matches smaller names). On mid caps the WR edge disappears (≈52% vs a 50.5% mid baseline, N=25 — winners run bigger but no hit-rate edge); the ungated small-cap pullback baseline is outright negative (46% WR, −0.07%/trade). Trade it on liquid large caps only.

Falsification record. On the original 122-ticker growth-heavy universe the same family printed up to 71.7% WR / 1.75 R/R (N=46) — expanding the universe 4.5× collapsed it. The 70%+ readings were small-N universe artifacts, not edge ("universe shapes conclusions"). The SPY-Gaussian market kill-switch helped at 122 tickers and consistently hurt at scale, so it ships as an optional gate param, off by default. Remaining caveats: as-of-today universe and cap tiers (survivorship bias), no live forward track record.

Usage

Tooling is managed by mise; tasks wrap every common operation (run mise tasks to see them all).

mise install        # pin runtimes (node 24, bun)
bun install         # install workspace deps + git pre-commit hook
mise run dev        # API (5101) + Web (5100) dev servers in parallel

Environment (optional)

Variable	Effect
`TIINGO_API_KEY`	Enables the Tiingo daily-OHLCV fallback when Yahoo is rate-limited or down (free tier: 1,000 req/day). Without it, OHLCV degrades to empty on Yahoo failure.
`SLACK_WEBHOOK_URL`	Slack notification for BUY/SELL opinions from `predict`.

CLI (packages/core)

# Daily prediction for a ticker list (default command)
mise run predict -- --ticker=TSLA,PLTR --sort=asc

# Slack notification for BUY/SELL opinions (either form)
SLACK_WEBHOOK_URL=https://hooks.slack.com/services/XXX mise run predict -- --ticker=TSLA,PLTR
mise run predict -- --ticker=TSLA,PLTR --slack-webhook=https://hooks.slack.com/services/XXX

# Strategy validation & tuning
mise run backtest        # version comparison, goal search, SELL validation
mise run backtest -- --cost-bps=20   # vary the round-trip cost (default 10bps)
mise run backtest -- --quick         # stop after version comparison + gate tuning
mise run optimize        # parameter optimizer (positional symbol, e.g. TSLA)
mise run learn           # learn from prediction feedback

Each predict run appends rows to a monthly CSV in packages/core/public/ (e.g. stock_data_202511.csv), tickers in alphabetical order (--sort=desc reverses).

Quality gate

mise run ci          # lint → typecheck → test → build

Automation

.github/workflows/daily-data.yml — runs predict after US market close and auto-commits the monthly CSV.
.github/workflows/weekly-optimize.yml — weekly parameter optimization, results uploaded as a build artifact.