Calibration Ledger

Live track record for every nowcast Macro Sentinel publishes. Each prediction is recorded the moment it's made, scored against realized data when it lands, and aggregated here on every site rebuild. The methodology moat made tangible: instead of claiming "our nowcast is calibrated" via in-sample backtests, every prediction is recorded with model version + as-of date + target date and graded publicly. Updated ; ledger contains rows across forecast targets.

How to read this page. Each section below is one nowcast target. Magnitude targets (Inflation 3m/6m, Wage 6m/12m, CCMI 12m return v1 + v2) are scored by MAE (typical error in pp), bias (systematic over/under), and correlation. Categorical targets (Inflation regime) are scored by hit rate vs the realized regime label. The aggregate metrics sit alongside a 24-month rolling chart that surfaces drift — a model whose recent rolling MAE is materially worse than its full-sample MAE is calibrating poorly to the current regime.

Methodology

Append-only JSONL ledger. Every nowcast made on this site is recorded as a single line in ${data.ledger_path}. Each row contains the target name, the date the prediction was made (as_of), the date it targets (target_date), the prediction value, the realized actual (filled when available), and the model version. Once written, rows are not edited except to fill in actuals.

Backfill from rolling-origin OOS history. For nowcasts with rolling-origin calibration (Inflation 3m/6m, Wage 6m/12m, CCMI v2 12m return), the historical OOS predictions made at each past month are loaded into the ledger so the page has track record from day one. These are as-if-live — the prediction was made using only data available at that month's as_of. Every newer prediction is appended live going forward.

Why this matters. The calibration discipline of Macro Sentinel research depends on every claim being verifiable. Without a public ledger, "our nowcast is calibrated at corr +0.88" is a statement about historical OOS that any subscriber has to take on trust. With the ledger, every prediction is recorded the day it's made and graded the day the realized data lands. Drift, regime instability, and methodology gaps surface immediately — and so do well-calibrated improvements.

Known caveat — CCMI v1 12m return bias (resolved by v2). The original CCMI nowcast was built with static train-only OLS through 2022-12-31. The post-2022 OOS portion shows a systematic bias of roughly pp — the model trained pre-disinflation didn't see the AI-rally-meets-disinflation regime that followed, so it consistently underpredicts forward returns. CCMI v2 was rebuilt with 60-month rolling-origin OLS and the same ledger now shows v2 bias of pp — bias eliminated. Both versions stay in the ledger so the methodology improvement is auditable.

Ledger row count by target: .

Total rows in ledger: .