Table 3 Comparison news vs. financial data. Root-mean-squared errors (RMSE) for forecasts on days with a new quarterly GDP release. A lower RMSE implies higher predictive accuracy. h=0 (h=1) denotes the forecast for the current (next) quarter. Panel (a) shows the evaluation for an indicator based only on financial market data. Panel (b) shows the evaluation for an indicator based only on news data. As benchmark, we use the first quarterly GDP release for the corresponding quarter. The Diebold-Mariano-West (DMW) test provides a p value for the null hypothesis of equal predictive accuracy against the alternative written in the column header (Diebold and Mariano 2002; West 1996). We assume a quadratic loss function
(a) Only financial market data | ||||
|---|---|---|---|---|
RMSE | RMSE | Relative RMSE | DMW test (p value) | |
First release | f-curve | First release/f-curve | First release <f-curve | |
h=0 | 0.45 | 0.57 | 0.79 | 0.19 |
h=1 | 0.45 | 0.71 | 0.63 | 0.042 |
(b) Only news data | ||||
RMSE | RMSE | Relative RMSE | DMW test (p value) | |
First release | f-curve | First release/f-curve | First release <f-curve | |
h=0 | 0.45 | 0.64 | 0.71 | 0.047 |
h=1 | 0.45 | 0.7 | 0.63 | 0.019 |