quantmine 0.2.0__tar.gz

quantmine-0.2.0/.airflowignore

@@ -0,0 +1,10 @@
+.venv
+.venv-win
+tmp
+notebooks
+__pycache__
+airflow
+quantmine
+scripts
+tests
+data

quantmine-0.2.0/.gitignore

@@ -0,0 +1,40 @@
+# --- market data & run artifacts (never committed: size + data licensing) ---
+data/
+tmp/
+analysis/
+output.txt
+run.log
+*.parquet
+*.pkl
+*.xlsx
+
+# --- airflow runtime (airflow.cfg contains secret keys; db/logs are local state) ---
+airflow/
+*.db
+*.db-shm
+*.db-wal
+
+# --- python ---
+__pycache__/
+*.pyc
+.venv/
+.venv-win/
+.ipynb_checkpoints/
+
+# --- build & test artifacts ---
+dist/
+build/
+*.egg-info/
+.pytest_cache/
+.coverage
+htmlcov/
+
+# --- local scratch / superseded scripts ---
+scripts/
+
+# --- generated dashboards (frontend being rebuilt) ---
+factor_analysis_dashboard.html
+
+# --- editor/os noise ---
+.DS_Store
+Thumbs.db

quantmine-0.2.0/.python-version

@@ -0,0 +1 @@
+3.13

quantmine-0.2.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 b00831205-web
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

quantmine-0.2.0/PKG-INFO

@@ -0,0 +1,222 @@
+Metadata-Version: 2.4
+Name: quantmine
+Version: 0.2.0
+Summary: Equity factor research library with honest statistics: Newey-West IC tests, multiple-testing control, point-in-time universe, turnover-based costs, Carhart attribution
+License: MIT
+License-File: LICENSE
+Requires-Python: >=3.13
+Requires-Dist: numpy>=2.4
+Requires-Dist: pandas>=3.0
+Requires-Dist: pyarrow>=24.0
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: scipy>=1.17
+Requires-Dist: statsmodels>=0.14
+Provides-Extra: data
+Requires-Dist: beautifulsoup4>=4.14; extra == 'data'
+Requires-Dist: curl-cffi>=0.15; extra == 'data'
+Requires-Dist: pandas-datareader>=0.11; extra == 'data'
+Requires-Dist: yfinance>=1.4; extra == 'data'
+Provides-Extra: viz
+Requires-Dist: plotly>=6.8; extra == 'viz'
+Description-Content-Type: text/markdown
+
+# quantmine
+
+An equity factor research library built around one principle: **statistical
+honesty**. Every step that commonly inflates backtest results — survivorship
+bias, overlapping-return autocorrelation, multiple testing, look-ahead in
+orthogonalization, unrealistic transaction costs — is explicitly addressed,
+and conclusions are reported with their uncertainty, not just their point
+estimates.
+
+The repo doubles as a full S&P 500 research case study: the library
+(`quantmine/`), the daily Airflow pipeline (`pipelines/`), and the findings
+below were produced by the same code you can `pip install`.
+
+## Install
+
+```bash
+pip install quantmine          # library only
+pip install "quantmine[data]"  # + yfinance download stack
+```
+
+For development (repo checkout, Python ≥ 3.13, [uv](https://docs.astral.sh/uv/)):
+
+```bash
+uv sync
+python -m pytest test
+```
+
+## Quick start
+
+Bring your own price data — any wide DataFrame (index = trading days,
+columns = tickers) works; nothing is hard-wired to yfinance or the S&P 500:
+
+```python
+import pandas as pd
+import quantmine as qf
+
+# 1) Wrap your data (or use qf.ParquetSource / qf.YFinanceSource)
+data = qf.MarketData(close=close_df, volume=volume_df)
+
+# 2) Register a custom factor — built-in factors register automatically
+@qf.factor_register("my_reversal")
+def my_reversal(close: pd.DataFrame, tickers: list) -> pd.DataFrame:
+    return -close[tickers].pct_change(5)
+
+# 3) Compute all registered factors (dependencies resolve automatically)
+pool = qf.build_param_pool(data, day=5, halflife=10, period=20)
+failed, factors = qf.calculate_all_factors(pool)
+
+# 4) Cross-sectional IC with Newey-West t-stats and multiple-testing control
+fwd = qf.forward_return(data.close, periods=[1, 5, 20])
+cs_ic = qf.CS_Information_Correlation(factors, fwd, output_path="cs_ic.parquet")
+report = qf.multiple_testing(qf.newey_west_summary(cs_ic))
+
+# 5) Quantile backtest on a point-in-time universe, turnover-based costs
+universe = qf.MembershipTableSource(membership_df)   # or qf.StaticUniverse([...])
+results, history = qf.quantile_backtest(universe, factors, ["my_reversal"], fwd)
+daily = qf.expand_all_to_daily_returns(history, data.close)
+
+# 6) Carhart four-factor attribution of the long-short returns (daily, HAC)
+french = qf.load_french_factors("ff3_daily.csv", "momentum_daily.csv")
+model = qf.carhart_attribution(daily[("my_reversal", 20)]["long_short"], french)
+```
+
+### Extension points
+
+| Protocol / hook | Purpose | Ships with |
+|---|---|---|
+| `DataSource.load(...) -> MarketData` | plug in any price/volume source | `ParquetSource`, `CSVSource`, `ExcelSource`, `YFinanceSource` |
+| `ConstituentsSource.get_constituents(date) -> set` | point-in-time universe from any provider | `MembershipTableSource` (interval table), `StaticUniverse` |
+| `@factor_register(name)` | add factors; params injected by name, factor-on-factor dependencies resolved | 8 built-in factors |
+| `config.example.yaml` | every pipeline parameter as validated dataclasses via `qf.load_configs` | defaults documented in-file |
+
+## Headline result (S&P 500 case study)
+
+The 20-day average volume factor (`TwentyDayAvgVol`) is the only candidate
+that survives the full testing gauntlet:
+
+| Stage | Result |
+|---|---|
+| Train IC (2015–2023, Newey-West) | t = 3.78 (20d holding), passes Bonferroni & Benjamini-Hochberg across all 18 factor × horizon tests |
+| Out-of-sample quintiles (2024–2026) | Monotonic (Spearman 0.9); long-short gross ~14 %/yr (Sharpe ~1.5), ~1.4 Sharpe net of turnover-based costs |
+| Carhart 4-factor attribution (daily, HAC, net of costs) | Market beta 0.24 (significant), large-cap tilt (SMB −0.13); momentum & value loadings insignificant |
+| Net alpha | ~10 %/yr net of costs (t = 1.78, p ≈ 0.07, n = 603 daily obs) — economically meaningful, **marginally short of the 5 % significance bar** on 2.4 years of out-of-sample data |
+
+The honest conclusion: the factor's IC is robustly significant in-sample under
+conservative testing; its out-of-sample net alpha is economically meaningful
+but does not clear conventional significance. Live verification over a longer
+window is required — and that is exactly what a research report should say.
+(A low-power period-level regression, n ≈ 30, says nothing either way —
+p ≈ 0.38 with a confidence interval wide enough to hold any conclusion. Test
+power and test bookkeeping move the verdict as much as the signal does.)
+
+The remaining seven candidate factors (momentum, short-term reversal,
+volatility, downside volatility, volume-price correlation, …) fail the
+corrected significance tests. Documenting *why* they fail is part of the
+point.
+
+## Methodology highlights
+
+- **Survivorship-bias correction** — the universe is rebuilt from historical
+  S&P 500 membership (764 tickers over 2015–2026); 569 were recoverable via
+  yfinance, and the residual gap is disclosed rather than hidden.
+- **Point-in-time universe** — each backtest cross-section only contains
+  stocks that were actually index members on that date.
+- **Newey-West IC tests** — daily ICs on overlapping k-day forward returns are
+  autocorrelated; plain `t = IR·√n` overstates significance several-fold. NW
+  (Bartlett kernel, lag = 2(k−1)) uses all daily observations while correcting
+  the standard error. A down-sampled IID test is kept as a robustness control.
+- **Multiple-testing control** — Bonferroni and Benjamini-Hochberg across all
+  factor × holding-period combinations.
+- **Train/test split with embargo** — factors are selected and the
+  orthogonalization is fit on 2015–2023 only; a gap of one month before the
+  test window prevents overlapping forward returns from leaking across the
+  split.
+- **Expanding-window orthogonalization** — correlated factors are residualized
+  with betas estimated on data available up to each date (no full-sample
+  look-ahead).
+- **Turnover-based transaction costs** — costs are charged on actual
+  membership turnover per rebalance, not on a flat 100 %-turnover assumption.
+- **Sanity checks** — factor displacement and cross-sectional shuffling tests
+  confirm the backtest machinery itself is not the source of the returns.
+- **Tested** — the research chain is covered by a unit + golden-value test
+  suite (`test/`), including hand-computed backtest fixtures, determinism
+  checks, and point-in-time universe edge cases.
+
+## Pipeline
+
+```
+yfinance (batch download, retry, blacklist, checkpoints)
+        │  quantmine/data_acquisition.py · pipelines/task_1.py
+        ▼
+cleaning & merge (ffill, dedup)          Airflow DAG: pipelines/DAG_pipeline.py
+        │  pipelines/task_2.py
+        ▼
+factor computation (registry-driven, vectorized pandas)
+        │  quantmine/factor_mining.py · pipelines/task_3.py
+        ▼
+IC testing: cross-sectional & time-series IC, NW t, BH/Bonferroni,
+train/test split, orthogonalization        quantmine/ic_calculator.py
+        ▼
+quintile backtest: PIT universe, monotonicity, turnover costs,
+displacement/shuffle sanity tests          quantmine/back_testing.py
+        ▼
+Carhart 4-factor attribution (daily, HAC)  quantmine/factor_attribution.py
+```
+
+Repository layout:
+
+```
+quantmine/          research library (importable package)
+pipelines/            Airflow DAG + daily CLI tasks
+test/                 pytest suite (unit + golden-value)
+config.example.yaml   all pipeline parameters, documented defaults
+```
+
+## Reproducing the case study
+
+**Market data is not included** (Yahoo Finance terms of service do not permit
+redistribution). To reproduce:
+
+1. Historical S&P 500 membership: provide a CSV with `ticker`, `start_date`,
+   `end_date` columns and point `SP500_MEMBERSHIP_CSV` at it.
+2. Prices/volumes: `python pipelines/task_1.py --date <ds> --batch manual`
+   downloads in batches with checkpointing, then `pipelines/task_2.py` cleans
+   and merges.
+3. Fama-French factors: download the daily FF3 and momentum CSVs from the
+   [Ken French Data Library](https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html)
+   into `tmp/ff3/`.
+4. Run the research chain: `python pipelines/task_3.py ...` for factors, then
+   the IC → backtest → attribution steps as in the quick start (or
+   `python -m quantmine.ic_calculator` for the packaged train/test workflow).
+
+For the Airflow DAG, set `QUANT_PROJECT_ROOT` and `QUANT_PYTHON_BIN` (see
+`DAG_pipeline.py` docstring) and copy `airflow.cfg.example` keys into your own
+config — never commit a real `airflow.cfg`.
+
+## Known limitations
+
+- ~195 of 764 historical members could not be recovered from yfinance (mostly
+  true delistings/acquisitions), so a residual survivorship bias remains and
+  likely flatters the results slightly.
+- The out-of-sample window (2024–2026) covers a single market regime.
+- The long-short portfolio carries a significant 0.24 market beta; a
+  beta-hedged variant is on the roadmap.
+- Transaction cost model is a flat per-turnover rate; no market-impact or
+  borrow-cost modeling.
+
+## Roadmap
+
+- [ ] Migrate storage from parquet files to PostgreSQL/DuckDB
+- [ ] REST API + MCP server exposing the research chain as agent-callable tools
+- [ ] RAG-based automated research reports (ChromaDB + LLM)
+- [ ] Extend the Airflow DAG to cover IC testing → backtest → reporting
+- [ ] Rebuild the analytics dashboard (frontend rewrite in progress)
+- [ ] Scale the data layer (full US market) with PySpark
+- [ ] Beta-hedged long-short variant; GARCH volatility targeting
+
+## License
+
+MIT

quantmine-0.2.0/README.md

@@ -0,0 +1,200 @@
+# quantmine
+
+An equity factor research library built around one principle: **statistical
+honesty**. Every step that commonly inflates backtest results — survivorship
+bias, overlapping-return autocorrelation, multiple testing, look-ahead in
+orthogonalization, unrealistic transaction costs — is explicitly addressed,
+and conclusions are reported with their uncertainty, not just their point
+estimates.
+
+The repo doubles as a full S&P 500 research case study: the library
+(`quantmine/`), the daily Airflow pipeline (`pipelines/`), and the findings
+below were produced by the same code you can `pip install`.
+
+## Install
+
+```bash
+pip install quantmine          # library only
+pip install "quantmine[data]"  # + yfinance download stack
+```
+
+For development (repo checkout, Python ≥ 3.13, [uv](https://docs.astral.sh/uv/)):
+
+```bash
+uv sync
+python -m pytest test
+```
+
+## Quick start
+
+Bring your own price data — any wide DataFrame (index = trading days,
+columns = tickers) works; nothing is hard-wired to yfinance or the S&P 500:
+
+```python
+import pandas as pd
+import quantmine as qf
+
+# 1) Wrap your data (or use qf.ParquetSource / qf.YFinanceSource)
+data = qf.MarketData(close=close_df, volume=volume_df)
+
+# 2) Register a custom factor — built-in factors register automatically
+@qf.factor_register("my_reversal")
+def my_reversal(close: pd.DataFrame, tickers: list) -> pd.DataFrame:
+    return -close[tickers].pct_change(5)
+
+# 3) Compute all registered factors (dependencies resolve automatically)
+pool = qf.build_param_pool(data, day=5, halflife=10, period=20)
+failed, factors = qf.calculate_all_factors(pool)
+
+# 4) Cross-sectional IC with Newey-West t-stats and multiple-testing control
+fwd = qf.forward_return(data.close, periods=[1, 5, 20])
+cs_ic = qf.CS_Information_Correlation(factors, fwd, output_path="cs_ic.parquet")
+report = qf.multiple_testing(qf.newey_west_summary(cs_ic))
+
+# 5) Quantile backtest on a point-in-time universe, turnover-based costs
+universe = qf.MembershipTableSource(membership_df)   # or qf.StaticUniverse([...])
+results, history = qf.quantile_backtest(universe, factors, ["my_reversal"], fwd)
+daily = qf.expand_all_to_daily_returns(history, data.close)
+
+# 6) Carhart four-factor attribution of the long-short returns (daily, HAC)
+french = qf.load_french_factors("ff3_daily.csv", "momentum_daily.csv")
+model = qf.carhart_attribution(daily[("my_reversal", 20)]["long_short"], french)
+```
+
+### Extension points
+
+| Protocol / hook | Purpose | Ships with |
+|---|---|---|
+| `DataSource.load(...) -> MarketData` | plug in any price/volume source | `ParquetSource`, `CSVSource`, `ExcelSource`, `YFinanceSource` |
+| `ConstituentsSource.get_constituents(date) -> set` | point-in-time universe from any provider | `MembershipTableSource` (interval table), `StaticUniverse` |
+| `@factor_register(name)` | add factors; params injected by name, factor-on-factor dependencies resolved | 8 built-in factors |
+| `config.example.yaml` | every pipeline parameter as validated dataclasses via `qf.load_configs` | defaults documented in-file |
+
+## Headline result (S&P 500 case study)
+
+The 20-day average volume factor (`TwentyDayAvgVol`) is the only candidate
+that survives the full testing gauntlet:
+
+| Stage | Result |
+|---|---|
+| Train IC (2015–2023, Newey-West) | t = 3.78 (20d holding), passes Bonferroni & Benjamini-Hochberg across all 18 factor × horizon tests |
+| Out-of-sample quintiles (2024–2026) | Monotonic (Spearman 0.9); long-short gross ~14 %/yr (Sharpe ~1.5), ~1.4 Sharpe net of turnover-based costs |
+| Carhart 4-factor attribution (daily, HAC, net of costs) | Market beta 0.24 (significant), large-cap tilt (SMB −0.13); momentum & value loadings insignificant |
+| Net alpha | ~10 %/yr net of costs (t = 1.78, p ≈ 0.07, n = 603 daily obs) — economically meaningful, **marginally short of the 5 % significance bar** on 2.4 years of out-of-sample data |
+
+The honest conclusion: the factor's IC is robustly significant in-sample under
+conservative testing; its out-of-sample net alpha is economically meaningful
+but does not clear conventional significance. Live verification over a longer
+window is required — and that is exactly what a research report should say.
+(A low-power period-level regression, n ≈ 30, says nothing either way —
+p ≈ 0.38 with a confidence interval wide enough to hold any conclusion. Test
+power and test bookkeeping move the verdict as much as the signal does.)
+
+The remaining seven candidate factors (momentum, short-term reversal,
+volatility, downside volatility, volume-price correlation, …) fail the
+corrected significance tests. Documenting *why* they fail is part of the
+point.
+
+## Methodology highlights
+
+- **Survivorship-bias correction** — the universe is rebuilt from historical
+  S&P 500 membership (764 tickers over 2015–2026); 569 were recoverable via
+  yfinance, and the residual gap is disclosed rather than hidden.
+- **Point-in-time universe** — each backtest cross-section only contains
+  stocks that were actually index members on that date.
+- **Newey-West IC tests** — daily ICs on overlapping k-day forward returns are
+  autocorrelated; plain `t = IR·√n` overstates significance several-fold. NW
+  (Bartlett kernel, lag = 2(k−1)) uses all daily observations while correcting
+  the standard error. A down-sampled IID test is kept as a robustness control.
+- **Multiple-testing control** — Bonferroni and Benjamini-Hochberg across all
+  factor × holding-period combinations.
+- **Train/test split with embargo** — factors are selected and the
+  orthogonalization is fit on 2015–2023 only; a gap of one month before the
+  test window prevents overlapping forward returns from leaking across the
+  split.
+- **Expanding-window orthogonalization** — correlated factors are residualized
+  with betas estimated on data available up to each date (no full-sample
+  look-ahead).
+- **Turnover-based transaction costs** — costs are charged on actual
+  membership turnover per rebalance, not on a flat 100 %-turnover assumption.
+- **Sanity checks** — factor displacement and cross-sectional shuffling tests
+  confirm the backtest machinery itself is not the source of the returns.
+- **Tested** — the research chain is covered by a unit + golden-value test
+  suite (`test/`), including hand-computed backtest fixtures, determinism
+  checks, and point-in-time universe edge cases.
+
+## Pipeline
+
+```
+yfinance (batch download, retry, blacklist, checkpoints)
+        │  quantmine/data_acquisition.py · pipelines/task_1.py
+        ▼
+cleaning & merge (ffill, dedup)          Airflow DAG: pipelines/DAG_pipeline.py
+        │  pipelines/task_2.py
+        ▼
+factor computation (registry-driven, vectorized pandas)
+        │  quantmine/factor_mining.py · pipelines/task_3.py
+        ▼
+IC testing: cross-sectional & time-series IC, NW t, BH/Bonferroni,
+train/test split, orthogonalization        quantmine/ic_calculator.py
+        ▼
+quintile backtest: PIT universe, monotonicity, turnover costs,
+displacement/shuffle sanity tests          quantmine/back_testing.py
+        ▼
+Carhart 4-factor attribution (daily, HAC)  quantmine/factor_attribution.py
+```
+
+Repository layout:
+
+```
+quantmine/          research library (importable package)
+pipelines/            Airflow DAG + daily CLI tasks
+test/                 pytest suite (unit + golden-value)
+config.example.yaml   all pipeline parameters, documented defaults
+```
+
+## Reproducing the case study
+
+**Market data is not included** (Yahoo Finance terms of service do not permit
+redistribution). To reproduce:
+
+1. Historical S&P 500 membership: provide a CSV with `ticker`, `start_date`,
+   `end_date` columns and point `SP500_MEMBERSHIP_CSV` at it.
+2. Prices/volumes: `python pipelines/task_1.py --date <ds> --batch manual`
+   downloads in batches with checkpointing, then `pipelines/task_2.py` cleans
+   and merges.
+3. Fama-French factors: download the daily FF3 and momentum CSVs from the
+   [Ken French Data Library](https://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html)
+   into `tmp/ff3/`.
+4. Run the research chain: `python pipelines/task_3.py ...` for factors, then
+   the IC → backtest → attribution steps as in the quick start (or
+   `python -m quantmine.ic_calculator` for the packaged train/test workflow).
+
+For the Airflow DAG, set `QUANT_PROJECT_ROOT` and `QUANT_PYTHON_BIN` (see
+`DAG_pipeline.py` docstring) and copy `airflow.cfg.example` keys into your own
+config — never commit a real `airflow.cfg`.
+
+## Known limitations
+
+- ~195 of 764 historical members could not be recovered from yfinance (mostly
+  true delistings/acquisitions), so a residual survivorship bias remains and
+  likely flatters the results slightly.
+- The out-of-sample window (2024–2026) covers a single market regime.
+- The long-short portfolio carries a significant 0.24 market beta; a
+  beta-hedged variant is on the roadmap.
+- Transaction cost model is a flat per-turnover rate; no market-impact or
+  borrow-cost modeling.
+
+## Roadmap
+
+- [ ] Migrate storage from parquet files to PostgreSQL/DuckDB
+- [ ] REST API + MCP server exposing the research chain as agent-callable tools
+- [ ] RAG-based automated research reports (ChromaDB + LLM)
+- [ ] Extend the Airflow DAG to cover IC testing → backtest → reporting
+- [ ] Rebuild the analytics dashboard (frontend rewrite in progress)
+- [ ] Scale the data layer (full US market) with PySpark
+- [ ] Beta-hedged long-short variant; GARCH volatility targeting
+
+## License
+
+MIT

quantmine-0.2.0/airflow.cfg.example

@@ -0,0 +1,18 @@
+# Minimal Airflow settings for this project.
+# Copy the relevant keys into your own $AIRFLOW_HOME/airflow.cfg.
+# NEVER commit the real airflow.cfg: it contains fernet_key / secret_key.
+
+[core]
+# Point Airflow at this repo so it picks up DAG_pipeline.py
+dags_folder = /path/to/quant-factor-mining
+
+# Generate your own key:
+#   python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
+fernet_key = REPLACE_ME
+
+[database]
+sql_alchemy_conn = sqlite:////path/to/airflow/airflow.db
+
+[api]
+# Generate your own random secret, e.g.:  openssl rand -base64 16
+secret_key = REPLACE_ME

quantmine-0.2.0/config.example.yaml

@@ -0,0 +1,33 @@
+# quantmine 配置示例 —— 复制为 config.yaml 后按需修改。
+# 所有 key 和字段都可省略, 省略即使用默认值(下方标注的即默认值)。
+
+data_acquisition:
+  checkpoint_dir: tmp/checkpoint
+  max_retries: 3
+  wait: 60                  # 批次下载失败后的重试等待秒数
+
+momentum:
+  day: 5                    # 动量回看天数(5 = 5日动量)
+
+forward_return:
+  periods: [1, 5, 20]       # 持有期(交易日)
+
+newey_west:
+  lag_multiplier: 2         # NW修正的lag = lag_multiplier * (持有期-1)
+
+orthogonalize:
+  threshold: 0.03           # 平均|IR|低于该值的因子被剔除
+  min_period: 60            # expanding beta 的最小估计窗口
+
+time_series_stationary_test:
+  rolling_period: 126       # 滚动IC窗口(约半年)
+  periods: [1, 5, 20]       # ACF的滞后阶
+
+backtest:
+  part: 5                   # 分位组数(5 = 五分组)
+
+transaction_cost:
+  cost_per_trade: 0.001     # 单边费率(按实际换手计费)
+
+carhart_attribution:
+  maxlags: 20               # HAC标准误的最大滞后

quantmine-0.2.0/pipelines/DAG_pipeline.py

@@ -0,0 +1,76 @@
+"""Airflow DAG for the daily data-download / clean / factor-calculation loop.
+
+Configuration comes from environment variables so the DAG file itself is
+portable across machines:
+
+    QUANT_PROJECT_ROOT  Absolute path of this project. Defaults to the parent
+                        of the directory containing this file (repo root).
+    QUANT_PYTHON_BIN    Python interpreter used to run the task scripts.
+                        Defaults to ``python``. On a WSL-hosted Airflow that
+                        drives a Windows venv, point it at
+                        ``.venv-win/Scripts/python.exe`` (WSL interop launches
+                        Windows executables directly, no PowerShell wrapper
+                        needed).
+"""
+import os
+from datetime import datetime, timedelta
+from airflow.providers.standard.operators.bash import BashOperator
+from airflow.utils.trigger_rule import TriggerRule
+from airflow.sdk import DAG
+
+PROJECT_ROOT = os.environ.get("QUANT_PROJECT_ROOT",
+                              os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+PYTHON_BIN = os.environ.get("QUANT_PYTHON_BIN", "python")
+
+
+def task_command(script: str) -> str:
+    """Build the bash command that runs one pipeline script.
+
+    Args:
+        script: Script filename inside the ``pipelines/`` directory.
+
+    Returns:
+        A bash command string with the Airflow ``ds``/``run_id`` templates.
+
+    Notes:
+        The command runs from the repo root so that relative data paths
+        (``tmp/``, ``data/``) resolve correctly.
+    """
+    return (f'cd "{PROJECT_ROOT}" && "{PYTHON_BIN}" pipelines/{script} '
+            '--date {{ ds }} --batch {{ run_id }}')
+
+
+with DAG("quant_factor_mining",
+        default_args={
+            "retries": 1,
+            "retry_delay": timedelta(minutes=5)
+        },
+        description="quant_factor_mining pipeline version 0.1, using S&P 500",
+        schedule=timedelta(days=1),
+        start_date=datetime(2020, 1, 1),
+        catchup=False,
+        tags=['quant_factor_mining'],
+        ) as dag:
+    t1 = BashOperator(
+        task_id="data_downloading",
+        bash_command=task_command("task_1.py"),
+    )
+    t2 = BashOperator(
+        task_id="data_cleaning",
+        depends_on_past=True,
+        bash_command=task_command("task_2.py"),
+        trigger_rule=TriggerRule.ALL_DONE
+    )
+    t3 = BashOperator(
+        task_id="factor_calculation",
+        bash_command=task_command("task_3.py"),
+    )
+    task_retry = BashOperator(
+        task_id="retry_downloading",
+        bash_command=task_command("task_retry.py"),
+        trigger_rule=TriggerRule.ALL_FAILED
+    )
+
+
+t1 >> t2 >> t3
+t1 >> task_retry