方法与数据说明Methodology & Data Policy

每个指标先定义独立预测对象，再在外层严格留出的滚动起点上评估；内层只用当时历史选择朴素基线、动态回归、因子桥接与稳健状态空间候选，并按验证误差决定是否集成。Each indicator has an explicit forecast target and is evaluated on untouched rolling outer origins. Inner history alone selects among simple baselines, dynamic regression, factor bridges, and robust state-space candidates, with ensembles determined by validation errors.

跨指标比较以 MASE、RMSSE 和相对本轮最强朴素基线的技能为主；方向准确率同时报告 Wilson 区间和多数方向基线，近期窗口单独披露。未稳定战胜基线的模型不会标成正式优势。Cross-indicator comparisons center on MASE, RMSSE, and skill against the strongest simple baseline in each cycle. Directional accuracy includes a Wilson interval and majority-direction baseline, with recent performance shown separately. Models without stable baseline outperformance are not presented as demonstrated edge.

宏观序列优先使用 ALFRED 首次发布值及 realtime_start；代理发布日期只在缺少精确历史时使用，并受陈旧期限约束。每个预测以追加事件冻结目标、信息截止、区间、模型、数据哈希、代码哈希和后续首次发布实际值；只有稳定发布确认后的事件进入公开账本，失败或本地尝试仍可审计但会被排除。Macro series prefer ALFRED initial releases and realtime_start. Release-date proxies are used only where exact history is unavailable and are bounded by stale-data limits. Every forecast is frozen as an append-only event with target, information cutoff, intervals, model, data and code hashes, and its later first-release outcome. Only confirmed stable-publication events enter the public ledger; failed or local-only attempts remain auditable but are excluded.

研究候选与公开服务模型分开记录：正式状态可直接服务；实验性候选只有全样本 MASE 小于 1 且技能为正时保留，否则回退到同一评估窗口的基线。50% / 80% / 95% 区间使用预测原点之前的扩展绝对残差校准，前 20 个点为预热，95% 覆盖率低于 90% 会明确显示警告。Research candidates and public-serving models are recorded separately. Production models may serve directly; experimental candidates serve only when full-sample MASE is below 1 with positive skill, otherwise the same-window baseline protects the public forecast. 50% / 80% / 95% intervals use expanding absolute residual calibration available before each origin; the first 20 points warm up, and observed 95% coverage below 90% is shown as an explicit warning.

日频、周频、月频指标按各自官方发布节奏更新。月频指标的最新观测往往早于站点更新时间，这属于正常现象，不代表数据异常。Daily, weekly, and monthly indicators update on their own official release cadence. Monthly series often have an observation date that precedes the site update date, which is normal and not a data anomaly.

公开站点以月度配置摘要形成统一判断，再用混频观察面保留真实节奏，通过单指标详情页承接深读，再用发布时间工作台和研究流连接持续跟踪。The public site starts with a monthly allocation summary, keeps real cadence through the mixed-cadence surface, deepens through indicator detail pages, and then connects ongoing monitoring through the release workbench and research stream.

页面聚焦核心指标与其预测窗口，不直接展示内部调度日志或维护字段。更完整的工程说明请见系统说明页。This page focuses on flagship indicators and their forecast windows rather than internal scheduler logs or maintenance fields. Deeper engineering notes live in the System Guide.