How Does the Nowcaster Work?
This is a purely statistical nowcast and does not represent the views of any individual or organization.
It uses a wide range of monthly data sources to predict current-quarter GDP by predicting each of its individual expenditure components and then aggregating them. It is updated twice each morning, and checks the St. Louis Federal Reserve Bank's FRED database for updates to any of the series. The first update takes place just before 9:00am, capturing the 8:30am release of many BEA, Census, and BLS series. There is a second update just before 11am capturing later releases and those that were not made available on FRED in time for the nowcaster to incorporate them.
MIDAS Regression Framework
The nowcaster is based heavily on the MIDAS (Mixed Data Sampling) regressions approach. Here, quarterly variables are predicted using monthly frequency data releases. The nowcaster is optimized by giving more weight to the predictions of monthly variables with better predictive power.
As an example, quarterly variable \(i\) is predicted in the month of the quarter \(j\) by optimally-lagged monthly variable \(k\). The lags, \(l\), available for the monthly indicator will refer to different time periods depending on availability at the stage of the data cycle (\(j\)).
MIDAS Base Specification:
$$ \Delta Q^{i}_{t,j} = c +D^{0,1}\beta^{Q}\Delta Q^{i}_{t-1} + \Sigma^{lim}_{l=1}\beta_{l}M^{k}_{l,j}+\epsilon_{t} $$
For each expenditure component, a separate specification for each monthly indicator is estimated to predict \(\Delta Q^{i}\). The chosen specification minimizes the out-of-sample root mean squared error (RMSE) of the forecast, and can differ in each month (\(j\)) of the quarter.
A routine chooses whether to include an AR term in the specification by switching \(D\) to either 1 or 0, and chooses the limit of the number of lags of the monthly variable to include. For example, the best forecasting specification may choose to include 6 lags of the monthly variable in the final month of the quarter, but just 3 in the start of the quarter based on forecasting performance. Forecasting performance may or may not be improved by the inclusion of an AR term in either case.
Multi-Model Specification
The above example allows for just one monthly variable to be used to predict the quarterly variable of interest. The same procedure is used for each monthly variable that is likely to contain useful information about developments in the quarterly variable. In addition, a "multi-model" is produced for each quarterly variable. This allows multiple monthly variables to be included in the same equation. The lags of each monthly variable are chosen to minimize the RMSE of the specification, as in the case of the single variable model. The multimodel is tested for out of sample accuracy for the specific data availability in the quarter at the current moment in time (i.e. I have month 1 consumption and month 2 control retail sales, whats the optimal lag combination to predict quarterly consumption given this point in the datacycle).
MIDAS Multi-Model:
$$ Q^{i}_{t,j} = c +D^{0,1}\beta^{Q}Q^{i}_{t-1,j} + \Sigma^{n}_{k = 1}\Sigma^{lim_{j}}_{l=1}\beta^{k}_{l}M^{k}_{l,j}+\epsilon_{t} $$
Optimal Nowcast Selection
Now, with predictions for the quarterly expenditure component available for multiple variables and the multimodel, a procedure selects the optimal nowcast for the given point in the data cycle. First, each nowcast available in the month of the quarter where data is available is weighted together by the inverse of their squared out-of-sample forecast error. This effectively gives more weight to predictions that tend to have smaller errors.
Then the following procedure selects the nowcast for each expenditure component:
- Check which month of the quarter is the latest with available data:
- Compare the RMSE of the weighted combined model to each individual model, including the multimodel. : Select the nowcast that has the lowest RMSE.
The procedure is performed for each expenditure component and updated when new data is made available.
GDP Aggregation via Chain-Linking
The expenditure components are combined to form the GDP nowcast using the standard chain-linking formula (using quantity indices \(Q\), updated with the nowcast, and price index \(P\)):
Chain-Linking Formula:
$$ \Delta GDP_{t} = \sqrt{\frac{\Sigma^{i=7}_{i=0}P^{i}_{t-1}Q_{t}^{i}}{\Sigma^{i=7}_{i=0}P^{i}_{t-1}Q_{t-1}^{i}}\frac{\Sigma^{i=7}_{i=0}P^{i}_{t}Q_{t}^{i}}{\Sigma^{i=7}_{i=0}P^{i}_{t}Q_{t-1}^{i}}}-1 $$
To simplify matters, we avoid predicting the current quarter price indices, and use the known previous period relative prices alone to construct GDP in the nowcast. Cumulative price level deviations can have an important impact on the outcome of the chain-linked growth rate, but quarter-to-quarter effects tend to be minuscule.
Model Performance
The following tables show forecast errors from the model. These are quasi-out-of-sample, produced using the final vintage of data, but using the model estimated only on data ahead of the GDP release (and one and two months before the GDP release). Month 3 errors refer to all data released ahead of the advanced GDP publication being available. Month 2 and month 1 estimates refer to data availability one and two months before this date.
| Month 1 | Month 2 | Month 3 | |
|---|---|---|---|
| Mean Absolute Error | 1.55 | 1.31 | 1.02 |
| Mean Squared Error | 4.97 | 4.41 | 1.93 |
| Root Mean Squared Error | 2.23 | 2.11 | 1.39 |
| Month 1 | Month 2 | Month 3 | |
|---|---|---|---|
| Mean Absolute Error | 1.31 | 1.09 | 0.93 |
| Mean Squared Error | 2.95 | 2.37 | 1.58 |
| Root Mean Squared Error | 1.72 | 1.54 | 1.26 |
Data Sources & Updates
The nowcast incorporates a wide range of timely economic indicators from official government sources:
Primary Data Sources:
- Bureau of Economic Analysis (BEA)
- U.S. Census Bureau
- Bureau of Labor Statistics (BLS)
- Federal Reserve
- Institute for Supply Management
Update Schedule:
- First update: ~9:00 AM ET (captures 8:30 AM releases)
- Second update: ~11:00 AM ET (captures later releases)
- Automated checks for new FRED data
- Real-time incorporation of latest indicators
Open Source & Transparency
This nowcasting system is designed to provide transparent, timely economic insights based on publicly available data and well-established statistical methods. All data sources are publicly accessible, and the methodology is fully documented above.
Disclaimer: This is a purely statistical nowcast and does not represent the views of any individual or organization. Forecasts are subject to uncertainty and should be interpreted alongside other economic indicators and official statistics.