Track record
Backtest and live accuracy
Two different things, kept separate on purpose: a pre-launch backtest that validated the model against 2016–2025 history before it went live, and a live scorecard that accrues automatically from the actual forecasts GDPcast has published since launch.
Backtest, 2016Q1–2025Q4
Before publishing, GDPcast was replayed pseudo-real-time over 40 quarters — 725 weekly snapshots, re-estimated each quarter on data as it actually existed at the time (ALFRED vintages), truth taken from first-release GDP. Full methodology on the About the Model page. This is a backtest, not a live track record — it shows what the model would have done, run on this validated spec, had it existed since 2016.
Headline accuracy by horizon (ex-2020)
| h | RMSE | MAE | Bias | AR(2) RMSE | Direction hit |
|---|---|---|---|---|---|
| h0 | 1.05 | 0.77 | +0.27pp | 1.75 | 94% |
| h1 | 1.62 | 1.25 | -0.27pp | 1.64 | 72% |
| h2 | 2.02 | 1.42 | -0.46pp | 2.19 | 69% |
| h3 | 2.20 | 1.55 | -0.50pp | 2.12 | 66% |
| h4 | 2.00 | 1.48 | -0.57pp | 1.84 | 65% |
| h5 | 2.16 | 1.54 | -0.66pp | 1.94 | 67% |
| h6 | 2.08 | 1.52 | -0.70pp | 1.79 | 62% |
| h7 | 1.97 | 1.51 | -0.71pp | 1.80 | 68% |
| h8 | 1.96 | 1.48 | -0.73pp | 1.88 | 70% |
RMSE, MAE, and bias in percentage points of annualized real GDP growth. AR(2) is a simple autoregressive benchmark estimated the same way, same sample. 2020 (4 quarters) is reported separately rather than folded into these averages: h=0 RMSE 3.99, h≥1 RMSE ~23pp — no model, ours or anyone's, saw the COVID collapse a quarter ahead.
Nowcast vs the Atlanta Fed's GDPNow, matched days
| Segment | n | GDPcast RMSE | GDPNow RMSE |
|---|---|---|---|
| All matched prints | 689 | 1.43 | 1.40 |
| Early in quarter (<6 weeks) | 112 | 1.64 | 1.91 |
| Mid-quarter (6–9 weeks) | 209 | 1.53 | 1.52 |
| Late-quarter (10–13 weeks) | 202 | 1.39 | 1.20 |
| Endgame (14+ weeks) | 166 | 1.18 | 0.98 |
| Final print per quarter | 35 | 1.10 | 0.94 |
Over all 689 matched prints: 1.43 vs GDPNow's 1.40 RMSE — within the ~0.2pp bar set before launch. GDPcast leads early in the quarter (1.64 vs 1.91); GDPNow leads by the final print (1.10 vs 0.94). Matched at ±4 days on a weekly grid against GDPNow's published archive.
vs the Survey of Professional Forecasters median (ex-2020)
| h | n | GDPcast RMSE | SPF RMSE | Ratio |
|---|---|---|---|---|
| h0 | 34 | 1.65 | 1.80 | 0.92× |
| h1 | 33 | 1.70 | 2.10 | 0.81× |
| h2 | 32 | 1.79 | 2.08 | 0.86× |
| h3 | 31 | 1.86 | 1.84 | 1.01× |
| h4 | 30 | 1.74 | 1.59 | 1.09× |
Ratio = GDPcast RMSE ÷ SPF RMSE; below 1.00 means GDPcast beat the survey median. The bar set before launch was ≤1.10 at h=1–4 — met at every horizon.
Live scorecard
Every forecast GDPcast has published is archived the day it's made. This table scores the model's final call for each quarter — the last current-quarter nowcast made before that quarter closed — against the actual growth rate once BEA publishes its advance estimate. It fills in on its own; nothing here is hand-entered.
| Quarter | GDPcast final call | BEA advance | Error | As of |
|---|---|---|---|---|
| 2026Q3 | +3.5% | Pending — expected Oct 28, 2026 | — | Oct 2, 2026, 2:25 PM ET |
“GDPcast final call” is the last current-quarter nowcast made for that quarter — earlier, more-uncertain predictions of the same quarter made a horizon or two out aren't what's scored here. “Error” is actual minus forecast; a positive number means GDPcast came in low.