海外信号前向封存协议 v1
10 月 18 日安装探针不合格，不计入前向观察；2026 年 10 月 19 日上海时间 08:44 首次定时归档已成功，现保留 1 个合格观察。
首次定时验收：2026 年 10 月 19 日
定时器于上海时间 08:44:00 启动服务，08:44:02 成功退出，未手工触发或重跑。实际采集从 2026-10-19T00:44:00.373965Z 开始，至 00:44:02.093013Z 完成，预测于 00:44:02.236136Z 封存。当日 manifest 的 eligible=true，无弃样原因；8 份归档 CSV、receipt 及 prediction 的 SHA256 均经独立核验。
这 1 个合格观察只验证采集与封存链路，不证明预测效果或收益。仍须累计至少 60 个合格目标日再做首次评估；本次未自动评分、发送消息或交易。
稳定核验标识：资格 manifest SHA256 为 7e2a388e925d0b46d846793781e055c1338a73afbf3d87958d07c6c163cef139；原始 receipt SHA256 为 6827020df51f990fb8fce1c74454c5a349fa5574f3b677c4c54ec337251ff3a4。这里只公布证据哈希，不公开原始行、当次预测方向或本机路径。
本协议只从真实 receipt 开始累计证据。既有历史数据用于研究和训练，不能再称作 untouched holdout，也不能凭现在下载证明历史时点可得。
首版固定信号 spx-last-session-direction-csi300-open-next-open-v1 预测沪深300当日开盘至下一 SSE 交易日开盘的涨跌方向：SPX 最近已闭市两个收盘价的收益率为正预测 +1，为负预测 -1，零为 0。不拟合、不再调参，不是既有年度 expanding 三臂模型，也不是交易或收益保证。选择该简单基线用于验证真实时点链路，不根据未来结果切换模型。
上海工作日 08:44 发起新请求，08:45 截止，完成 receipt 与预测封存都必须不晚于截止。采集开始须落在 08:43–08:45。SSE 当日交易必须由同次 receipt 的 trade_cal 确认。SPX 最晚来源日期必须等于截止在纽约的日期向前退至工作日；若美国假期导致此日期没有数据，保守弃样，不把前一日价格当作新信息。两个 SPX bar 之间不超过七个自然日，价格有限且为正、日期不重复，闭市加一小时不晚于截止。常规闭市时间仅为假设，真实 receipt 时间才是本轮接收证据。亚洲、FX 数据同时保存但不参与 v1 预测。
python -m market_research.studies.overseas_leadlag.forward --receipt-directory <MDP-run> --output <QMR-owner-data>/forward/overseas-v1
读取时核对原始 CSV hash；先独占写入 .prediction.json 并 fsync，再读取实际写入完成时钟。随后独占写入 .json manifest，其 eligible 是资格判定依据，包含 payload hash 和实际 sealed_utc，跨截止一律弃样。只有 payload 而无 manifest 表示封存中断。派生 JSON 保留 receipt hash、代码 hash、固定信号身份、选用 bar、预测和不合格原因。scheduled 每个目标日只允许独占创建一个文件，包括迟到/失败弃样，禁止覆写或事后挑选重跑。probe 独立封存，永不计入前向观察。若进程在 CSV/receipt/预测封存前失败，保留现有证据，次日人工核对缺口；不回填成合格记录。完整 started 而无 receipt 表示采集中断，完整 receipt 而无 forecast 表示封存失败。
v1 不消费任何目标日之后的国内价格。不自动打分、不发送消息、不下单。未来评分须另加 append-only outcome 记录，以真实后续获取的目标开盘价标注标签与接收时间；不改写预测。累计至至少 60 个合格 SSE 目标日再做首次方向命中率及不确定性评估，同时列出所有缺失、迟到和美国假期弃样；期间不因结果改规则。变更信号、截止或依赖必须启用新协议版本并重新起算。
稳定发布安装两个 owner 后，可以用 scripts/overseas_forward.sh [scheduled|capture_probe] 串联采集与封存。所需环境变量为 MARKETDATA_BIN（已安装的 owner launcher）、DATA_PLATFORM_CONFIG、MDP_CAPTURE_PYTHON、RESEARCH_PYTHON、OVERSEAS_RAW_ROOT、OVERSEAS_DERIVED_ROOT。两个 Python 和脚本自身必须指向已合并的不可变 release。单用途 systemd timer 采用 OnCalendar=Mon..Fri *-*-* 08:44:00 Asia/Shanghai、AccuracySec=1s、RandomizedDelaySec=0、Persistent=false，服务不自动重试，防止重跑替换当日弃样。日志保留服务退出和产物路径。
Overseas signals: forward archive protocol v1
Evidence starts with an actual receipt. Existing historical data remain research inputs; downloading them today cannot establish historical availability or create an untouched holdout. The October 18 installation probe is ineligible. The first scheduled archive completed on October 19, 2026 at 08:44 Shanghai time; one eligible observation is now retained.
First scheduled acceptance: October 19, 2026
The timer started the service at 08:44:00 Shanghai time and it exited successfully at 08:44:02, without a manual trigger or rerun. Collection ran from 2026-10-19T00:44:00.373965Z to 00:44:02.093013Z; the prediction was sealed at 00:44:02.236136Z. The dated manifest records eligible=true with no exclusion reasons. All eight archived CSV hashes, the receipt hash and the prediction hash were independently checked.
One eligible observation verifies this collection and sealing chain only. It does not establish predictive accuracy or returns. The first evaluation still requires at least 60 eligible target days; no outcome was automatically scored, no message was sent and no trade was placed.
Stable acceptance hashes: qualification manifest SHA256 7e2a388e925d0b46d846793781e055c1338a73afbf3d87958d07c6c163cef139; raw receipt SHA256 6827020df51f990fb8fce1c74454c5a349fa5574f3b677c4c54ec337251ff3a4. These identify the retained evidence without publishing source rows, signal direction or machine-local paths.
Fixed signal and timing
Signal spx-last-session-direction-csi300-open-next-open-v1 uses the sign of the return between the last two completed SPX closes to predict the CSI 300 direction from today’s open to the next SSE session’s open. Positive, negative and zero returns produce +1, −1 and 0. There is no fitting. This baseline checks the live information chain; it is separate from the historical expanding-window models and does not establish an executable strategy.
New requests start at 08:44 on Shanghai weekdays, within the permitted 08:43–08:45 window. Collection and prediction sealing must finish by 08:45. The calendar in the same receipt must confirm that today is an SSE session. The latest SPX source date must equal the cutoff’s New York date rolled back to a weekday. A US holiday without that date’s data causes exclusion. The two bars must be at most seven calendar days apart, have unique dates and finite positive prices, and satisfy assumed close plus one hour before cutoff. Assumed exchange close is not proof of the provider’s historical publication time. Actual receipt clocks provide the new evidence. Asian indices and FX are archived but are not v1 predictors.
Immutable records
python -m market_research.studies.overseas_leadlag.forward --receipt-directory <MDP-run> --output <QMR-owner-data>/forward/overseas-v1The consumer verifies raw CSV hashes, exclusively writes and fsyncs a .prediction.json payload, then reads the actual completion clock. It exclusively writes a .json qualification manifest containing the payload hash, actual sealed_utc and eligible decision. Crossing the deadline excludes the observation. A payload without a manifest indicates interrupted sealing.
Derived records retain receipt and code hashes, signal identity, selected bars, prediction and exclusion reasons. Scheduled mode exclusively creates one record per target day, including failed or late exclusions. It cannot overwrite or select a later rerun. Probes are stored separately and are always ineligible. Failures retain existing evidence; missing observations are investigated rather than backfilled as eligible. A started record without a receipt indicates interrupted collection; a receipt without a forecast indicates failed sealing.
Evaluation and operation
V1 consumes no domestic prices after the target date. It does not automatically score outcomes, send messages or place orders. Future evaluation needs separate append-only outcome records with subsequently received opening prices and receipt times; predictions remain unchanged. The first direction-accuracy evaluation requires at least 60 eligible SSE target days, uncertainty estimates and a list of all missing, late and US-holiday exclusions. Changing the signal, cutoff or dependencies starts a new protocol version.
After both owners are installed, scripts/overseas_forward.sh [scheduled|capture_probe] chains collection and sealing. Required environment names are MARKETDATA_BIN, DATA_PLATFORM_CONFIG, MDP_CAPTURE_PYTHON, RESEARCH_PYTHON, OVERSEAS_RAW_ROOT and OVERSEAS_DERIVED_ROOT. Both Python environments and the wrapper point to merged immutable releases. The dedicated timer uses OnCalendar=Mon..Fri *-*-* 08:44:00 Asia/Shanghai, AccuracySec=1s, RandomizedDelaySec=0 and Persistent=false. There is no automatic retry. Logs retain service status and artifact paths.