Commit Graph

4 Commits

Author SHA1 Message Date
claude_dev e3b688354f fix(data): data_platform硬化(增量merge/verify+raw_redownload/run_daily_update)+测试
merge_increment/verify_increment 增量staging→验证→合并工具; raw_redownload/run_daily_update/import_vnpy_daily 强化; 补 data_platform 与 index_downloader 测试.
2026-07-15 07:12:46 +08:00
claude_dev a372544045 feat(data): 双源5yr全市场部署 + baostock 15min + 断点续传
- raw_redownload 加断点续传(exists/skip已存在, 扩范围重下覆盖)
- baostock_download: 15min双源(qfq+raw)下载器, NAS容器跑, 限速防封
- full_deploy_5yr.sh: 无人值守日线双源pipeline(qfq→raw→rsync→验证)
- verify_dual_source: 容器内双源部署验证(fetch_day/iter_bars/除权日)

实测: 日线双源5yr 29600文件×2(rsync NAS), 浦发除权日 raw-5.9%/qfq-0.7%,
容器内 fetch_day/iter_bars/全市场抽检5只全通过. 15min沪深300 baostock限流121只.
2026-07-09 19:22:36 +08:00
claude_dev c6b19f4244 feat(data): 恢复双源(task#79)—撮合raw+策略qfq, 分红除权准确
用户要模拟=回测准确: raw除权缺口致MA假信号, 必须双源。
- data_source: qfq→qfq_dir(干净qfq), raw→raw_dir; _check_adjust_cfg(cfg提供才校验)
- engine 双bar流: step(raw_bars,qfq_bars)撮合/盯市raw+策略on_bar qfq; run zip(raw,qfq)
- live_orchestrator: warmup用qfq(信号am); 去adjust参数(双源固定)
- raw_redownload --adjust(''raw/'qfq'); config qfq_dir
- 113/113通过
2026-07-08 07:21:33 +08:00
claude_dev 1ed7b72aca feat(data): raw真实价数据源(task#79)—raw_dir+dir_key路由+新浪源重下
根因: daily_dir mixed-adjust(hfq bulk+akshare raw tail)致3-30 -94%假跌。
方案(Linus三问简化单raw, 除权留分期项#3):
- datareader read_parquet_daily/15min 加 dir_key 参数
- data_source iter_bars/fetch_day: adjust=raw→raw_dir(缺配置报错防混源), qfq→daily_dir
- engine PaperEngine 默认 adjust=raw
- config 加 raw_dir; scripts/raw_redownload.py 新浪源adjust='' 直连+单线程限速
- 验证: 浦发606行close 6.5/14.6 mean10.08 0跳变, 撮合成交价9.71-10.25真实
- 测试9/9+trader全量108/108通过
2026-07-07 22:19:11 +08:00