K.E.Y.N.E.S.Knowledge Engine Yielding Nonstop Economic Scenarios

Models › Production › GDP-2

The GDP nowcast · GDP-2

An estimate of the current quarter's real GDP growth, updated every Friday from the economic data published that week. The model is our replica of the New York Fed's Staff Nowcast 2.0, rebuilt from public data.

How it works, in brief. Many statistics rise and fall together because they respond to the same few forces: the overall pace of the economy, business sentiment, prices and the job market. The model treats these forces as hidden variables (factors), estimates them from 35 monthly and quarterly data series, and reads GDP growth off them. Each week it updates the factors with that week's releases, and each data release moves the nowcast in proportion to how much it surprised the model.

Every nowcast on this page uses only the data that had been published by its date, including the unrevised first versions of each figure. That is the honest test of how the model would have done in real time.

Every quarter, week by week

Each quarter's nowcast closes in on GDP as the data arrive

weekly nowcast, one line per quarter true GDP growth, at its first release date
Real GDP growth, % at an annual rate. Each light line runs from about five months before a quarter's first official estimate to the day before it. The axis is cut at ±9%, so 2020's swings (−33% and +33%) run off the chart. Hover over a line or dot for its quarter.

One quarter in detail

What moved the nowcast each week

Bars (left axis): each week's change in the nowcast, in percentage points, split by the kind of data released that week. Line (right axis): the nowcast itself.

Bars above zero pushed the nowcast up, bars below pushed it down; the first week is the starting point, so it has no bar. "Data revisions" are changes to figures already published; "re-estimation" is the effect of re-fitting the model each quarter and of updating its volatility and outlier estimates.

The biggest surprises this quarter

Actual and expected values are in the units the model uses: % changes for most series, changes in thousands for payrolls, index levels for surveys. Each surprise (actual minus expected) moves the nowcast by a weight the model estimates for that series and release; the weight can be positive or negative.

Accuracy

We score the nowcast by its mean squared error (MSE): the average of the squared misses, in percentage points of annualized growth. An MSE of 1.4 means a typical miss of about 1.2 points. Lower is better. The yardstick is a simple autoregressive (AR) benchmark, which forecasts GDP growth from its own past alone.

We replayed every quarter from 2006 to 2019, each week using only the data published by then. On the day before the government's first estimate, the nowcast's mean squared error was 1.37, about the same as the New York Fed reports for its own model (1.3) and 65% lower than the benchmark (3.93).

Mean squared error on the day before BEA's first estimate, 2006–2019, in squared percentage points of annualized growth; lower is better. How we measure accuracy.

The error falls by two-thirds as the quarter's data arrive

Mean squared error against BEA's first estimate, 55 quarters 2006Q2–2019Q4

our 2.0 replica, every week New York Fed 2.0, as published our replica of the older (2016–2021) New York Fed model AR benchmark (dashed)
Comparison points are scored 56, 28 and 1 days before release. The New York Fed figures come from their own backtest (Almuzara et al. 2023, Table 2), not from our sample.

Scores in full

Comparison with the New York Fed's published nowcasts

From 2016 to 2019 the New York Fed published its (older) model's nowcast every week. On the same dates, our replica had smaller errors against BEA's first estimate, with p-values near 0.05. Against the latest revised GDP figures the difference disappears. These were calm years, and the simple AR benchmark did about as well as either model.

The features that matter

The 2.0 model adds four features to the older design. Switching each one off in turn shows which matter: a slowly drifting trend in long-run growth, and outlier scales that stop a single odd month from moving the estimate. With all four off, the model is no better than the older one.

The trend and the outlier scales carry the gain

Mean squared error against BEA's first estimate, 2006–2019, with one feature switched off at a time

The COVID factor exists only in 2020, so switching it off cannot change these quarters. Stochastic volatility mainly changes the error bands, not the nowcast itself.

Series and factors

GDP-2 has 35 series and six hidden components:

  • Trend: the economy's long-run growth rate, which drifts slowly. Only GDP and GDI load on it.
  • Global: the overall pace of real activity. Almost every series loads on it.
  • Soft: business sentiment, from regional Fed manufacturing and services surveys.
  • Nominal: prices and series measured in dollars.
  • Labor: the job market.
  • COVID: a factor that exists only from March to September 2020, to absorb the pandemic's extreme swings.

Bars show the estimated loading of each series on each factor (posterior mean, standardized data). Blue bars move with the factor, red bars against it. An empty cell means the model does not allow that loading.

The model in detail

The model has four main factors (a global factor that every real-activity series loads on, and factors for surveys, nominal series such as prices and dollar values, and labor), a factor that exists only during the 2020 pandemic, and a slowly moving trend in long-run growth. Series and factors shows the loadings.

For economistsEquations

For a standardized monthly series \(i\):

$$y_{it} = \iota_i\, g_t + \lambda_i' f_t + \lambda^c_i\, c_t + e_{it}$$

Quarterly series (GDP, GDI, unit labor cost) are observed every third month as a weighted sum of the unobserved monthly values (Mariano and Murasawa 2003):

$$y^q_{it} = \tfrac{1}{9}\,(1, 2, 3, 2, 1) \cdot \big(\iota_i g + \lambda_i' f + e^q_i\big)_{t,\,t-1,\,\dots,\,t-4}$$

The trend, the factors and the series-specific terms evolve as

$$g_t = g_{t-1} + \gamma_g v_t, \qquad f_t = \sum_{l=1}^{4} \Phi_l f_{t-l} + s^f_t\, \sigma^f_t\, \varepsilon^f_t, \qquad e_{it} = \phi_i e_{i,t-1} + s^e_{it}\, \sigma^e_{it}\, \varepsilon_{it}$$ $$\ln (\sigma_t)^2 = \ln (\sigma_{t-1})^2 + \gamma\, \eta_t, \qquad s_t \in \{1, 2, 3, 4, 5\}, \qquad c_t = s^c_t\, \varepsilon^c_t \text{ in March–September 2020, else } 0$$
  • \(g_t\): trend in long-run growth, loading 1 on GDP and GDI only.
  • \(f_t\): global, soft, nominal and labor factors; every series loads on global except the price indexes, and each also loads on its group's factor.
  • \(\sigma\) and \(s\): each shock has its own random-walk volatility and, each month, an outlier scale of 1 (usual) or 2–5.
For economistsEstimation, data and departures from the paper
  • Sampler (paper, appendix A): the state by the simulation smoother; loadings and AR coefficients by Bayesian regression; the VAR kept stationary; volatilities by Kim, Shephard and Chib (1998) with the 10-component mixture of Omori et al. (2007); outlier scales from their discrete posterior. 1,500 burn-in sweeps, 1,000 kept, thinned to 100 draws; about 11 minutes per fit.
  • State: 45 elements (trend, 4 factors, COVID factor and 3 quarterly series-specific terms, each with 5 lags). Monthly series-specific terms are quasi-differenced out of the state.
  • Parameters: about 270, plus a volatility and an outlier scale for each of 39 shocks in every month.
  • Backtest: re-estimated on the first Wednesday of each quarter on data from January 1985; nowcast every Friday; the nowcast is the posterior mean over the kept draws.
  • Data: the 35 series of the paper's list (the older model's list less capacity utilization and the PPI), with Dallas and Philadelphia Fed survey indexes in place of the ISM indexes, which are not on FRED. ALFRED vintages for every series that has them; values from before a series' archive begins are left out unless the series is barely revised (surveys, price indexes).
  • Departures: the paper does not publish its prior values, so ours are our own (loadings and AR coefficients N(0, 1); trend volatility 0.025 with 1,000 prior observations, which the data barely move; outlier probability about one month in 48). Price indexes do not load on the global factor, and quarterly loadings are drawn with their series-specific terms integrated out; both were needed for the model to fit. Standardization leaves out March–December 2020.
For economistsWhich data move the nowcast

Average share of the absolute weekly news, by data category, compared with the New York Fed's reported shares.