Skip to content
Methodology 5 of 8

GDPNow Probabilistic Nowcasting

Turning 28 FRED series into a current-quarter GDP distribution.

The Atlanta Fed's GDPNow is one of the most-watched macroeconomic indicators in the world. It nowcasts quarterly US real GDP growth by aggregating dozens of incoming data releases into a single running estimate. A typical release week sees the nowcast shift by 0.1–0.4 percentage points as new data lands; in a surprise release, shifts of 0.5pp or more happen.

Compass's probabilistic GDPNow replica does the same aggregation but produces a distribution, not a point. Where the Atlanta Fed shows you “+2.15%,” Compass shows you “+2.15% with 90% confidence between +1.60% and +2.65%, driven most by industrial production and retail sales.” The distribution is the information.

The 28-input catalog

Compass tracks 28 FRED series covering the six BEA expenditure components plus cross-cutting indicators. Each series maps to a specific component (e.g., RSAFS → PCE Goods) with its natural release frequency (daily, monthly, quarterly). The catalog is deliberately broad — more inputs mean more information but also more places where thin data can degrade the fit. Coverage is the primary lever: on a given day, the catalog typically has data for ~20 of 28 inputs.

Source fusion: four priors per input

Each input can draw its distribution from up to four sources:

  • Revision — historical BEA revision statistics tell us how much advance-estimate prints typically drift from their final values.
  • Forecaster — consensus economic forecasts provide a crowd-sourced anchor.
  • Market — event-contract prices embed trader probabilities for specific thresholds.
  • AI — an LLM-generated estimate grounded in current context and recent history.

Present sources are fused via weighted-quantile averaging per input. Absent sources are simply dropped and remaining weights renormalize — the fusion degrades gracefully as data availability shifts through the release cycle.

Bridge-equation Monte Carlo

Once each input has a distribution, Compass propagates uncertainty through the bridge equations that connect monthly indicators to quarterly GDP components. The bridge equations are BEA-style share-weighted aggregations: retail sales + durable goods + services → PCE Goods growth → PCE Goods contribution to GDP. Monte Carlo over 10,000 trials turns each input's distribution into a distribution of component contributions and ultimately the headline.

The output is a metalog-encoded headline distribution plus per-component summary stats, per-input variance shares (which drive Sobol indices), and a tornado chart showing each input's p10→p90 swing.

The volatility arbitrage signal

One of the most useful downstream products of the probabilistic replica is the vol spread, Δσ = σF − σM:

  • σF (fundamental vol) — the standard deviation of our headline distribution. How uncertain we are about GDP.
  • σM (market-implied vol) — the volatility implied by Treasury options / VIX term structure / Fed funds futures. How uncertain markets are about GDP.

When Δσ > 0, we're more uncertain than markets — markets may be under-pricing GDP vol. When Δσ < 0, markets are more uncertain than us — they may see risks our model isn't capturing. The rolling z-score of Δσ surfaces regimes where the spread is extreme relative to its own history. This is a tradeable signal that depends entirely on the probabilistic framing — a point estimate of GDP can't produce it.

What drove this nowcast?

Two complementary attributions:

  • Sobol indices — first-order Sobol decomposes headline variance into per-input shares. S1 = 0.30 for Industrial Production means 30% of the nowcast uncertainty is attributable to IP alone.
  • Tornado chart — per-input p10→p90 swing on the headline. Directional (positive-weighted inputs push headline up at their high end; net-exports inputs push it down).

Together these answer “which inputs matter most, and which direction do they push?” — the two questions analysts ask every release day.

From the book: Beyond Synthesis, Chapter 23–25: “The Volatility Arbitrage Signal.”