Chronos forecast ledger. One row per forecast step: series, units, origin date, horizon step, target date, median and 80% interval quantiles (q10/q90), seasonal-naive baseline, action receipt, and the realized value once published. Scored against the baseline with abs_error, coverage (inside_80), and status.
| Column | Range | Mean | Filled |
|---|---|---|---|
| abs_error | — | — | 0% |
| baseline | 4.1–233,000 | 78,920 | 97% |
| baseline_abs_error | — | — | 0% |
| baseline_method | 2 distinct | — | 97% |
| horizon_step | 0–12 | 6.324 | 100% |
| inside_80 | — | — | 0% |
| issued_at | 2026-09-11 13:15:00+00–2026-09-13 01:03:25.751+00 | — | 100% |
| median | 4.047–206,000 | 72,630 | 97% |
| model | 1 distinct | — | 97% |
| notes | 4 distinct | — | 41% |
| origin_date | 2026-07-01 05:00:00+00–2026-09-05 05:00:00+00 | — | 97% |
| q10 | 3.625–197,400 | 68,240 | 97% |
| q90 | 4.221–222,800 | 77,840 | 97% |
| realized_value | — | — | 0% |
| run_id | 4 distinct | — | 100% |
| scored_at | — | — | 0% |
| series_id | 3 distinct | — | 97% |
| source | 1 distinct | — | 100% |
| status | 2 distinct | — | 100% |
| target_date | 2026-08-01 05:00:00+00–2027-08-01 05:00:00+00 | — | 97% |
| units | 1 distinct | — | 97% |
Ledger audit — 2026-09-11
Full audit of all 25 rows (24 forecast rows, 1 void schema-seed row). Checks run programmatically against the row set, plus action-receipt verification.
Counts: checked 25, changed 0, voided 0.
Per-check results:
Required fields: all 24 forecast rows have run_id, source, series_id, units, origin_date, issued_at, horizon_step, target_date, median, q10, q90, baseline, baseline_method, model, action_id. The seed row (01a09102-1d0f-7a83-a094-5efb94bc8f18) is correctly void with an explanatory note and carries no forecast fields.
Target-date alignment: every row satisfies target_date = origin_date + horizon_step months. PCOPPUSDM: origin 2026-07-01, steps 1–12 → 2026-08-01 … 2027-07-01, all aligned. UNRATE: origin 2026-08-01, steps 1–12 → 2026-09-01 … 2027-08-01, all aligned.
Interval ordering: q10 < median < q90 strictly holds on all 24 forecast rows. No inverted, zero-width, or degenerate intervals.
Baseline method: seasonal_naive_12 on all 24 rows, matching the fixed watchlist baselines (no post-hoc method switching).
Action references: both receipts resolve and are successful runs of the TimesFM Forecast route (route 9693eaa2-0302-4340-bb73-4b02eddcb9bd): copper run 01a09094-47a1-7c33-80bc-81b6b274f1f4, unemployment run 01a09094-386a-71ac-ae7b-9675b7056980.
Gap/rounding notes: present where policy requires them — UNRATE step 2 documents the 2025-10 interpolation into its baseline; PCOPPUSDM step 1 documents the 0.1 USD/t input rounding.
No open row carries realized_value or scored_at, so nothing is pre-scored. No corrections were needed; nothing was voided.
Due-score pass — 2026-09-11 (this pass is repeatable; no rows were modified)
Checked every open row whose target_date has passed (today is 2026-09-11): 3 rows found, 0 scored, 3 deferred.
Scored: none — no actual for a past target date is published yet.
Deferred (left open, release lag is normal, absence is not failure):
rows (run_id / horizon_step) | target_date | why deferred |
|---|---|---|
| 2026-08-01 | FRED has not published the August 2026 monthly average. Fresh pull 2026-09-11 via Fred Series ( |
| 2026-09-01 | Target month not complete — the September monthly average cannot exist until after 2026-09-30 (publishes ~mid-October). |
| 2026-09-01 | BLS September Employment Situation releases 2026-10-02. Actual unavailable by construction. |
Leakage check (no action needed yet): both forecast runs were issued 2026-09-11T13:15Z with origins 2026-07-01 (PCOPPUSDM) and 2026-08-01 (UNRATE). The August copper obs was not published at issue time and still is not, so no row is at leakage risk from a post-publication origin.
Next scoring opportunity: re-pull PCOPPUSDM (lin, m) next tick; score step 1 when the 2026-08-01 obs lands. Ledger forecast-ledger.
Reproduction audit — run pcoppusdm-2026-09-11 (PCOPPUSDM, 12 steps) — PASS 12/12
Reproduced the full run from its recorded origin history on 2026-09-12. Re-pulled PCOPPUSDM (units=lin, frequency=m, Fred Series action 01a09622-b878-7aca-8ae2-c34a2850bfaf; latest published obs still 2026-07-01 = 13542.82, so the origin window is unrevised), truncated to the recorded 139-month context (2015-01 → 2026-07, matching the original run's context_length: 139), rounded to 0.1 USD/t per the gap/rounding policy, and re-ran the Forecast route once (action 01a09625-cca5-7fbc-b131-63aca8f88865, model google/timesfm-3.0-pytorch, horizon 12, freq MS). Every baseline was recomputed independently from the fresh pull as the value 12 months before each target, not copied from the ledger.
step | target | row id | date | median | q10 | q90 | baseline | result |
|---|---|---|---|---|---|---|---|---|
1 | 2026-08-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
2 | 2026-09-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
3 | 2026-10-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
4 | 2026-11-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
5 | 2026-12-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
6 | 2027-01-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
7 | 2027-02-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
8 | 2027-03-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
9 | 2027-04-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
10 | 2027-05-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
11 | 2027-06-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
12 | 2027-07-01 |
| PASS | PASS | PASS | PASS | PASS | PASS |
Checks per row: (1) target date equals origin + step months and matches the route's dated step; (2) stored median equals the reproduction's 0.5 quantile; (3) stored q10 equals the 0.1 quantile; (4) stored q90 equals the 0.9 quantile; (5) stored baseline equals the independently recomputed seasonal-naive-12 value. All quantile comparisons agree to within 0.005 USD/t (values are bit-identical at 4 decimals); strict ordering q10 < median < q90 holds on all 12 rows. Recomputed baselines (9671.9, 9994.8, 10739.9, 10812.0, 11791.0, 12986.6, 12951.3, 12528.7, 12890.7, 13512.2, 13552.0, 13542.8) match the stored column exactly.
Rows checked: 12. Rows changed: 0. Rows voided: 0.
Reproduction receipt: View run. Original receipt of record: View run. The reproduction action is an audit artifact only — no ledger rows reference it.
Refresh check, 2026-09-12 — no series updated, no rows added.
Checked both represented series against their release windows. Neither has a newly published complete observation, so no refreshed forecasts were issued and the ledger's 24 open rows are unchanged.
PCOPPUSDM (units lin, frequency m): fresh Fred Series pull on 2026-09-12 shows the latest published observation is still 2026-07-01 (13542.82). The 2026-08 observation is not yet published (expected ~mid-September per the release-lag calendar). The open step-1 row (target 2026-08-01) therefore has no actual to score against, and August is not a complete target month, so it is not a refresh candidate.
UNRATE (units lin, frequency m): fresh Fred Series pull on 2026-09-12 shows the latest published observation is 2026-08-01 (4.1), which is the origin of the existing forecast. The 2026-09 observation releases 2026-10-02, so UNRATE cannot refresh before then.
Next refresh window: PCOPPUSDM August, mid-September.