Nowcasting models are used to estimate GDP growth based on indicators that have been published recently and that historically have shown to be highly correlated with the growth of GDP. They are a useful instrument to monitor cyclical developments on a high frequency basis before the publication of quarterly national accounts.
Our own modelling work covers the eurozone and France. All our models are calibrated (Q1 2000 – Q4 2016) and tested (Q1 2017 – Q4 2022) over the same periods. We have worked on two model families: one with non-reduced dimensions and the other with reduced dimensions (factor-augmented mixed-frequency approach). The resulting forecast corresponds to the average forecast of the two approaches.
Models without reduced dimensions
Within this family of models, data are selected by a 'stepwise'[1] regression. This family consists of five distinct models:
1) OLS model containing 12 variables: non-sparse linear regression;
2) OLS model containing 7 variables: sparse linear regression;
3) lasso model containing 10 variables: linear regression with a linear constraint on the regression coefficients in order to avoid the overweighting of certain regressors in relation to others;
4) ridge model containing 10 variables: linear regression with a quadratic constraint on the regression coefficients in order to avoid the overweighting of certain regressors compared to others;
5) random forest model containing all variables.
The GDP nowcast from these different models is then aggregated using weights based on the forecast accuracy (the inverse of root mean squared forecast error, RMSFE) calculated on the test sample.
Reduced dimension models
For the second family of models (factor-augmented mixed-frequency approach), activity and survey data are enriched with financial and international data. These data are selected, as in the first model family, by a stepwise regression that retains the regressors with the greatest predictive power. The selection of the chosen variables is then reduced through a principal component analysis (PCA). This statistical technique makes it possible to estimate, by linear combination of the initial variables, a monthly factor which represents the common dynamic. We then we use this monthly common factor to explain quarterly growth using mixed-frequency linear regression (MIDAS[1]).
A forecast at different times of the quarter
A nowcasting model offers a forecast of GDP growth at different points in the quarter. There are four in a quarter: month 0, month 1, month 2, month 3. These moments are dictated by the pace of publication of indicators. When doing an estimate in month 0, no hard data for the estimated quarter is published yet, but the surveys for the first month of the current quarter are available. Similarly, an estimate made in month 1 means that we have the first month of hard data and the first two months of surveys, and so on. The time when the forecast is made is critical: the quality of the forecast improves significantly as hard data are published and added. An estimate made, for example, in month 1 therefore has a greater degree of uncertainty than the estimate made in month 3, which is the most solid and a priori the best. The prediction error decreases as we move forward in the quarter.
Tarik Rharrab
