G7 HANK Models Global Macro Model v4 Global Macro Model v5 Payrolls Forecasting

Payrolls Forecasting Lab

Name: RoboMacro
Author: RoboMacro

EXPERIMENT

10 independent models — time-series econometrics through gradient boosting — each forecasting Friday's BLS Employment Situation. Methodology

Sep 2026 NFP

BLS Employment Situation — Friday, October 2, 2026

21 days away

Core Ensemble Forecast — Median of 8 Models

Headline consensus excludes VAR and LightGBM (implementation-corrected but weak on recent prints). Full 10-model median shown below for reference.

Nonfarm Payrolls

96k

80% model range: 52k → 352k

Unemployment Rate

4.1%

80% model range: 4.0% → 5.6%

Part-Time Workers

4377k

80% model range: 940k → 4852k

Avg Hourly Earnings YoY

3.27%

80% model range: 3.10% → 3.83%

NFP model disagreement (σ): ±121k · Core range (80%): 52k – 352k · Full 10-model median: 96k

Monthly Track Record — Forecast vs First Print (Last 4 Releases)

All scoring uses the first BLS print (8:30am headline) — not revised figures. Beat/miss compares who was closer; direction shows whether each forecast was too high or too low vs that first print.

Model implementation corrections (Aug 2026)

Some models were producing invalid forecasts due to implementation errors (not weak signal). Historical rows below for corrected models reflect the old broken pipeline until forecasts are re-run.

VAR — Corrected Aug 2026: VAR was fit on mixed units (NFP MoM in thousands vs ICSA ~200k raw claims), so companions dominated and forecasts were systematically wrong (e.g. −155k vs +57k first print in Jun 2026). Fix: per-column z-scoring before VAR estimation plus ±800k forecast cap. Re-run forecasts after this date reflect the corrected pipeline.
LGBM — Corrected Aug 2026: (1) ADPMNUSNERSA was ingested at person scale (~132.7M) beside PAYEMS in thousands. (2) Regime features used raw claims level differences (~200k) beside MoM NFP. (3) LightGBM lacked StandardScaler and used subsample/colsample on tiny tuned feature sets, producing absurd forecasts (e.g. −1,259k Jun 2026). Fix: ADP unit normalization, StandardScaler on base features, percent-change regime signals, adaptive tree hyperparams for small feature sets, ±800k cap. Historical track-record rows before re-run reflect the old pipeline.

Core ensemble vs street on first print: 3/4 releases we beat consensus

Aug 2026

✓ We beat street

Ensemble 87k off (87k too low)

Street 109k off (109k too low)

Jul 2026

✗ Street was closer

Ensemble 104k off (104k too high)

Street 103k off (103k too high)

Jun 2026

✓ We beat street

Ensemble 34k off (34k too high)

Street 53k off (53k too high)

May 2026

✓ We beat street

Ensemble 85k off (85k too low)

Street 87k off (87k too low)

Econometric vs Machine Learning medians (first print): Econ beat street 2/4 · ML beat street 2/4 · Econ closer than ML 2/4

Aug 2026

ML closer

Econ median 121k off (121k too low)

ML median 81k off (81k too low)

Jul 2026

Econometric closer

Econ median 96k off (96k too high)

ML median 112k off (112k too high)

Jun 2026

Econometric closer

Econ median 15k off (15k too high)

ML median 85k off (85k too high)

May 2026

ML closer

Econ median 107k off (107k too low)

ML median 62k off (62k too low)

Model	Sep 2026 NFP Awaiting first print	Aug 2026 First print +162k	Jul 2026 First print -23k Revised to +21k · not scored	Jun 2026 First print +57k Revised to +31k · not scored	May 2026 First print +172k Revised to +63k · not scored
Did we beat street? Core 8-model median vs Dow Jones / FinanceFlowAPI — excludes VAR and LightGBM	Pending	✓ We beat street Ensemble 87k off (87k too low) Street 109k off (109k too low)	✗ Street was closer Ensemble 104k off (104k too high) Street 103k off (103k too high)	✓ We beat street Ensemble 34k off (34k too high) Street 53k off (53k too high)	✓ We beat street Ensemble 85k off (85k too low) Street 87k off (87k too low)
STREETStreet Consensus Economist survey benchmark · first-print miss below	—	Forecast 109k below first print Forecast +53k ↓ First print +162k	Forecast 103k above first print Forecast +80k ↑ First print -23k	Forecast 53k above first print Forecast +110k ↑ First print +57k	Forecast 87k below first print Forecast +85k ↓ First print +172k
CORECore Ensemble (8 models) Excludes VAR + LightGBM · headline median · first-print miss below	Pre-release forecast+96kFirst print not yet released	Forecast 87k below first print Forecast +75k ↓ First print +162k ✓ We beat street Ensemble 87k off (87k too low) Street 109k off (109k too low)	Forecast 104k above first print Forecast +81k ↑ First print -23k ✗ Street was closer Ensemble 104k off (104k too high) Street 103k off (103k too high)	Forecast 34k above first print Forecast +91k ↑ First print +57k ✓ We beat street Ensemble 34k off (34k too high) Street 53k off (53k too high)	Forecast 85k below first print Forecast +87k ↓ First print +172k ✓ We beat street Ensemble 85k off (85k too low) Street 87k off (87k too low)
10Full Ensemble (10 models) Includes VAR + LightGBM · reference only	Pre-release forecast+96kFirst print not yet released	Forecast 87k below first print Forecast +75k ↓ First print +162k	Forecast 104k above first print Forecast +81k ↑ First print -23k	Forecast 34k above first print Forecast +91k ↑ First print +57k	Forecast 85k below first print Forecast +87k ↓ First print +172k
Econometric Models Time-series methods grounded in macroeconomic theory · 5 models
ECONEconometric median (5) Median of econometric models · first-print miss below	Pre-release forecast+93kFirst print not yet released	Forecast 121k below first print Forecast +41k ↓ First print +162k ✗ Street was closer Ensemble 121k off (121k too low) Street 109k off (109k too low)	Forecast 96k above first print Forecast +73k ↑ First print -23k ✓ We beat street Ensemble 96k off (96k too high) Street 103k off (103k too high)	Forecast 15k above first print Forecast +72k ↑ First print +57k ✓ We beat street Ensemble 15k off (15k too high) Street 53k off (53k too high)	Forecast 107k below first print Forecast +65k ↓ First print +172k ✗ Street was closer Ensemble 107k off (107k too low) Street 87k off (87k too low)
ARNaive Autoregressive (AR)	Pre-release forecast+93kFirst print not yet released	Forecast 81k below first print Forecast +81k ↓ First print +162k #4 Beat street	Forecast 96k above first print Forecast +73k ↑ First print -23k #5 Beat street	Forecast 15k above first print Forecast +72k ↑ First print +57k #2 Beat street	Forecast 107k below first print Forecast +65k ↓ First print +172k #4
SARIMASARIMAX	Pre-release forecast+42kFirst print not yet released	Forecast 122k below first print Forecast +40k ↓ First print +162k #7	Forecast 67k above first print Forecast +44k ↑ First print -23k #2 Beat street	Forecast 11k below first print Forecast +46k ↓ First print +57k #1 Beat street	Forecast 133k below first print Forecast +39k ↓ First print +172k #7
VARVector AutoregressionFIXED AUG 2026	Pre-release forecast+88kFirst print not yet released	Forecast 317k below first print Forecast -155k ↓ First print +162k #10	Forecast 122k above first print Forecast +99k ↑ First print -23k #7	Forecast 396k below first print Forecast -339k ↓ First print +57k #10	Forecast 218k below first print Forecast -46k ↓ First print +172k #8
UCUnobserved Components	Pre-release forecast+99kFirst print not yet released	Forecast 121k below first print Forecast +41k ↓ First print +162k #6	Forecast 83k above first print Forecast +60k ↑ First print -23k #3 Beat street	Forecast 32k above first print Forecast +89k ↑ First print +57k #4 Beat street	Forecast 58k below first print Forecast +114k ↓ First print +172k #2 Beat street
MIDASAlmon Distributed Lag	Pre-release forecast+124kFirst print not yet released	Forecast 25k below first print Forecast +137k ↓ First print +162k #2 Beat street	Forecast 130k above first print Forecast +107k ↑ First print -23k #8	Forecast 37k above first print Forecast +94k ↑ First print +57k #5 Beat street	Forecast 21k below first print Forecast +151k ↓ First print +172k #1 Beat street
Machine Learning Models Data-driven methods with non-linear feature interactions · 5 models
MLML median (5) Median of machine-learning models · first-print miss below	Pre-release forecast+233kFirst print not yet released	Forecast 81k below first print Forecast +81k ↓ First print +162k ✓ We beat street Ensemble 81k off (81k too low) Street 109k off (109k too low)	Forecast 112k above first print Forecast +89k ↑ First print -23k ✗ Street was closer Ensemble 112k off (112k too high) Street 103k off (103k too high)	Forecast 85k above first print Forecast +142k ↑ First print +57k ✗ Street was closer Ensemble 85k off (85k too high) Street 53k off (53k too high)	Forecast 62k below first print Forecast +110k ↓ First print +172k ✓ We beat street Ensemble 62k off (62k too low) Street 87k off (87k too low)
ENetElasticNet Linear	Pre-release forecast+354kFirst print not yet released	✓ Hit first print Forecast +163k ✓ First print +162k #1 Beat street	Forecast 417k above first print Forecast +394k ↑ First print -23k #10	Forecast 64k above first print Forecast +121k ↑ First print +57k #6	Forecast 62k below first print Forecast +110k ↓ First print +172k #3 Beat street
RFRandom Forest	Pre-release forecast+65kFirst print not yet released	Forecast 93k below first print Forecast +69k ↓ First print +162k #5 Beat street	Forecast 86k above first print Forecast +63k ↑ First print -23k #4 Beat street	Forecast 30k above first print Forecast +87k ↑ First print +57k #3 Beat street	Forecast 120k below first print Forecast +52k ↓ First print +172k #6
XGBXGBoost	Pre-release forecast+57kFirst print not yet released	Forecast 81k below first print Forecast +81k ↓ First print +162k #3 Beat street	Forecast 112k above first print Forecast +89k ↑ First print -23k #6	Forecast 85k above first print Forecast +142k ↑ First print +57k #7	Forecast 108k below first print Forecast +64k ↓ First print +172k #5
LGBMLightGBM + RegimeFIXED AUG 2026	Pre-release forecast+233kFirst print not yet released	Forecast 217k above first print Forecast +379k ↑ First print +162k #9	Forecast 12k above first print Forecast -11k ↑ First print -23k #1 Beat street	Forecast 388k above first print Forecast +445k ↑ First print +57k #9	Forecast 235k above first print Forecast +407k ↑ First print +172k #9
MLPNeural Network (MLP)	Pre-release forecast+351kFirst print not yet released	Forecast 163k below first print Forecast -1k ↓ First print +162k #8	Forecast 261k above first print Forecast +238k ↑ First print -23k #9	Forecast 176k above first print Forecast +233k ↑ First print +57k #8	Forecast 268k above first print Forecast +440k ↑ First print +172k #10
All models beating street:		5/10	5/10	5/10	3/10
Econometric models beating street:		2/5	3/5	4/5	2/5
ML models beating street:		3/5	2/5	1/5	1/5

“Off” = absolute miss vs first print (lower is better). Green We beat street = core ensemble was closer than consensus. Street row shows how far consensus missed; core ensemble row shows how far we missed. Models marked FIXED AUG 2026 had implementation errors corrected — see note above.