Forecast track record
Archived, never edited, scored against official data. The first round was archived retroactively.
Scoreboard
Every scored forecast against its ONS outturn.
Last scored ; data store updated 16 Nov 2026.
Too few to show skillLimits
The record
Error = archived median minus outturn, in pp; positive means the forecast ran high.
Against the benchmarks
| Target | This round | No change | BoE Feb 2026 MPR | OBR March 2026 EFO | BoE survey of external forecasters (Apr 2026) | BoE Nov 2026 MPR | OBR October 2026 EFO |
|---|---|---|---|---|---|---|---|
| 2026Q2 UK CPI, y/y | 0.12 | 0.303.10% | 0.702.10% | 0.662.14% | 0.102.90% | — | — |
| 2026Q2 UK real GDP, y/yround 2026-07-21 | 0.27 | 0.530.84% | 0.570.80% | 0.450.92% | — | — | — |
| 2026Q2 UK real GDP, y/yround 2026-07-29 | 0.23 | 0.530.84% | 0.570.80% | 0.450.92% | — | — | — |
| 2026Q2 UK unemployment rate | 0.07 | 0.105.00% | no published quarterly path | ||||
| 2026Q3 UK CPI, y/yround 2026-07-21 | 0.39 | 0.003.10% | 1.102.00% | 1.022.08% | — | — | — |
| 2026Q3 UK CPI, y/yround 2026-07-29 | 0.56 | 0.003.10% | 1.102.00% | 1.022.08% | — | — | — |
| 2026Q3 UK real GDP, y/yround 2026-07-21 | 0.58 | 1.100.84% | 0.941.00% | 0.721.22% | — | — | — |
| 2026Q3 UK real GDP, y/yround 2026-07-29 | 0.48 | 1.100.84% | 0.941.00% | 0.721.22% | — | — | — |
| 2026Q3 UK real GDP, y/yround 2026-11-12 | 0.37 | 0.571.37% | 0.941.00% | 0.721.22% | — | 0.041.90% | 0.571.37% |
No change is the naive baseline: the last outturn at the round's data edge, held flat (read from the current vintage, so revisions can shift it slightly). This is the driftless walk; the limits say what happens against a walk with drift.
Open rounds
On the record. Dots are published ONS outturns, scored above where the round came first; later quarters wait.
GDP growth from the 2026-11-12 round: our median and bands beside the OBR's path.
UK real GDP, % change on a year earlier
- BoE SVAR median
- 68% band (re-calibrated)
- 90% band (re-calibrated)
- OBR October 2026 EFO
- ONS outturn
Source: BoE SVAR round, 12 Nov 2026 · OBR October 2026 EFO · ONS
Details
Numbers by quarter
| Quarter | Median | 68% band (re-calibrated) | 90% band (re-calibrated) | EFO |
|---|---|---|---|---|
| 2026Q3 | 1.57% | 1.09%–2.06% | 0.54%–2.58% | 1.37% |
| 2026Q4 | 1.88% | 1.13%–2.64% | 0.38%–3.45% | 1.51% |
| 2027Q1 | 1.57% | 0.26%–2.67% | −0.63%–3.66% | 1.11% |
| 2027Q2 | 1.28% | −0.01%–2.64% | −1.44%–4.04% | 0.95% |
| 2027Q3 | 1.16% | −0.20%–2.61% | −1.50%–4.27% | 0.96% |
| 2027Q4 | 1.16% | −0.26%–2.66% | −1.42%–4.06% | 1.16% |
| 2028Q1 | 1.24% | −0.21%–2.79% | −1.64%–4.19% | 1.37% |
| 2028Q2 | 1.29% | −0.11%–2.78% | −1.82%–4.47% | 1.58% |
| 2028Q3 | 1.30% | −0.16%–2.73% | −1.85%–4.37% | 1.75% |
| 2028Q4 | 1.40% | −0.17%–2.69% | −1.92%–4.49% | 1.81% |
| 2029Q1 | 1.35% | −0.28%–2.74% | −1.94%–4.47% | 1.83% |
| 2029Q2 | 1.34% | −0.15%–2.86% | −1.90%–4.65% | 1.82% |
| 2029Q3 | 1.36% | −0.19%–2.96% | −1.88%–4.94% | 1.80% |
Sources
Source: archived round 2026-11-12/boe-svar.json; OBR October 2026 EFO path: four-quarter growth of Table 1.1 real GDP, detailed economy tables, stored in forecasts/official/obr_efo_oct_2026.csv; outturns from the data store
Same round, CPI: the model tracks the EFO through 2027Q3, then stays near 3% while the EFO returns towards 2%. Nothing in the model pulls inflation back to target (a unit root, no inflation anchor), so its later quarters extend recent inflation (limits).
UK CPI, % change on a year earlier
- BoE SVAR median
- 68% band (re-calibrated)
- 90% band (re-calibrated)
- OBR October 2026 EFO
- ONS outturn
Source: BoE SVAR round, 12 Nov 2026 · OBR October 2026 EFO · ONS
Details
Numbers by quarter
| Quarter | Median | 68% band (re-calibrated) | 90% band (re-calibrated) | EFO |
|---|---|---|---|---|
| 2026Q3 | 2.97% | 2.43%–3.56% | 1.94%–4.02% | 3.05% |
| 2026Q4 | 3.48% | 2.40%–4.50% | 1.54%–5.63% | 3.52% |
| 2027Q1 | 3.62% | 2.12%–5.16% | 0.57%–6.74% | 3.79% |
| 2027Q2 | 3.58% | 1.38%–5.81% | −0.50%–7.96% | 3.64% |
| 2027Q3 | 3.30% | 0.69%–5.94% | −1.52%–8.16% | 3.38% |
| 2027Q4 | 3.15% | 0.72%–5.81% | −1.86%–8.40% | 2.89% |
| 2028Q1 | 3.07% | 0.54%–5.82% | −1.64%–8.07% | 2.37% |
| 2028Q2 | 3.08% | 0.78%–5.47% | −2.61%–8.44% | 2.14% |
| 2028Q3 | 3.07% | 0.80%–5.53% | −1.98%–8.84% | 1.99% |
| 2028Q4 | 3.15% | 0.80%–5.53% | −2.30%–8.95% | 1.91% |
| 2029Q1 | 3.06% | 0.72%–5.50% | −2.30%–8.62% | 2.02% |
| 2029Q2 | 3.02% | 0.72%–5.48% | −2.02%–8.65% | 2.13% |
| 2029Q3 | 3.06% | 0.63%–5.55% | −2.34%–8.69% | 2.14% |
Sources
Source: archived round 2026-11-12/boe-svar.json; EFO path from Table 1.7 of the October 2026 EFO detailed economy tables, stored in forecasts/official/obr_efo_oct_2026.csv; outturns from the data store
Unemployment, satellite round 2026-07-28; no official path to overlay. Bands carry GDP uncertainty only and run narrow: in a backtest excluding Covid quarters, 68% bands built this way held 22–48% of outturns and 90% bands 30–61% (coverage: the share inside a band).
UK unemployment rate, %
- BoE SVAR unemployment satellite median
- 68% band (GDP uncertainty only)
- 90% band (GDP uncertainty only)
- ONS outturn
Source: BoE SVAR unemployment satellite, 28 Jul 2026 · ONS
Details
Numbers by quarter
| Quarter | Median | 68% band | 90% band |
|---|---|---|---|
| 2026Q2 | 4.97% | 4.94%–5.00% | 4.93%–5.01% |
| 2026Q3 | 4.96% | 4.89%–5.03% | 4.85%–5.08% |
| 2026Q4 | 4.94% | 4.81%–5.07% | 4.74%–5.16% |
| 2027Q1 | 4.93% | 4.74%–5.14% | 4.61%–5.27% |
Sources
Source: archived round 2026-07-28/okun-unemployment.json; outturn, ONS series MGSX from the data store
Band calibration and the EFO comparison
- The October 2026 EFO was fixed on 28 Oct, before the 12 Nov round, on less data: the charts show where views differ. The forecast evaluation is on the BoE SVAR validation page.
- A round archived after an outturn is drawn beside it but never scored on it.
- GDP and CPI bands are rescaled per horizon to measured coverage: on average they under-cover, but UK GDP over-covers at short horizons (68% factor 0.85 one quarter ahead, 1.13 at eight). Factors:
band_calibration.json. - Factors come from the model shown (with its Covid dummies) over 50 overlapping origins, so they are imprecise beyond two decimals.
- The EFO growth path is four-quarter growth of Table 1.1 real GDP in the October 2026 EFO detailed economy tables.
- Raw quantiles stay untouched in the round files. Re-calibrated numbers are tabled under each fan; raw medians and bands are on the Growth and Inflation tabs. The BoE MPR and the survey of external forecasters are in the official CSVs.
Archive
What was claimed, and when. Each round is frozen when archived.
- 7 days later
- 1 day later
- 106 days later
First round weaker: archived after the CPI release from a file committed ~16 h before it, with thinner provenance.
Data vintages and scoring inputs
The data edge is the last observation the model was allowed to see; outturns are read from the dated, never-edited snapshots in the data store.
A CI job fails any pull request that edits or deletes a round, so a correction creates a new round and the superseded one stays scored.
Scores are computed from outturns.json,
which is vintage-versioned so revisions never silently rewrite a
score, and written to scorecard.json;
the method is in the README.
Method and limits
What this record does not show, and how it is kept.
Limits
- Too few to judge: no skill claim before ten or more outturns, a year or more away.
- Bands run narrow in hindsight tests, worse with horizon; open fans are re-calibrated.
- Unemployment bands are too narrow: the archived round carries GDP uncertainty only (measured coverage). The model now adds the satellite’s own uncertainty; no round with those bands is archived yet.
- No demonstrated skill against a drifting random walk: no gain on any variable survives multiple-testing correction.
- Three variables, no fixed calendar; coefficients are not re-estimated every round.
- Every row here is BoE SVAR; the microsimulation does not forecast. How a score is built
How we score six rules, each enforced in code
- 1Archive first. The round's JSON is committed before the outturn is published (the first round is the exception).
- 2Never edit. CI rejects any edit or deletion; a correction is a new round.
- 3No hindsight. A round archived after the outturn's snapshot is never scored on it.
- 4Dated outturns. Read from a dated ONS snapshot; revisions are appended, never overwritten.
- 5Error and coverage. Signed error of the median; coverage of the bands as archived.
- 6Benchmarks. Beside a no-change forecast and any published path for that quarter.
The full method
- Archive and freeze
- One JSON file per model under forecasts/rounds/<date>/: median, 68% and 90% bands, data edge, source hash.
- The forecast-archive workflow fails any PR that edits, renames or deletes one.
- So the Git timestamp proves when each claim was made.
- No hindsight
- Outturns
- ingest_outturns.py converts stored ONS series to the forecasts’ units: GDP growth from the ABMI level; CPI and unemployment as published.
- Each value is appended with its snapshot date; revisions are appended, never overwritten.
- The scoreboard uses the newest; earlier scores stay reproducible from outturns.json.
- Errors and bands
- Error is the archived median minus the outturn; positive means it ran high.
- Kept signed, because bias and random misses differ.
- Coverage uses the bands as archived, not the re-calibrated bands in the open-round charts.
- Benchmarks
- A no-change forecast: the last outturn at the data edge, held flat.
- Any stored quarterly path published before the round: BoE MPR (Feb and Nov 2026), OBR EFO (Mar and Oct 2026), BoE survey of external forecasters (Apr 2026; the Oct 2026 survey is stored but scores no quarter yet), from the official CSVs.
- Everything lands in scorecard.json, generated, never edited by hand.
- Why publish this early
- It makes early failures part of the evidence.
- Bank Rate comes closest to an accuracy gain, which does not survive the multiple-testing adjustment; it is not forecast here yet.
- Rounds follow data-edge refreshes, not a calendar.
Release notes
One written note and the automated digests of each stored data release. Neither is evidence of skill.