Informational efficiency Properties of good forecasts and measures used

4.1.2 Testing for accuracy

Applying the MAE to check for accuracy the size of the absolute forecast error, we find that GET forecasts were more accurate the shorter the time horizon. In absolute terms, our one, two, three, and four years ahead forecasts were off by 0.5, 0.9, 1.4 and almost 2.4 percentage points respectively see Table 3 and Annex 2, Table B2 for more details. The MedAE was smaller than the mean absolute error indicating that there were some extreme values that are punished to a greater extent in the mean absolute forecast error. Values close to zero indicate accurate predictions; the MedAE shows that 50 per cent of the range of MAE for one and two years ahead was below 0.2 and 0.7 percentage points, respectively. Similarly, for three and four years ahead, 50 per cent of the absolute forecast errors were below 1 percentage point. Table 3. Summary of accuracy statistics for global unemployment rate forecasts Years ahead MAE Mean Absolute Error RMSE Root Mean Squared Error MedSE Median Squared Error MedAE Median Absolute Error MSE Mean Squared Error UB Bias proportion of MSE UV Variance proportion of MSE UC Covariance proportion of MSE R 2 World 1 0.45 0.72 0.06 0.25 0.52 0.02 0.00 0.98 0.98 2 0.94 1.27 0.52 0.72 1.60 0.00 0.02 0.98 0.94 3 1.36 1.98 0.98 0.99 3.92 0.02 0.00 0.98 0.85 4 1.91 3.02 1.01 1.01 9.12 0.11 0.00 0.89 0.73 Source: ILO calculations based on the GET January 2010; January 2011; January 2012; January 2013. According to the MAE and the MedAE, our forecasts for one and two years ahead were relatively accurate on average in comparison to longer term forecasts. The decomposition of the MSE into the bias and the variance proportion UB and UV, which give an indication for systematic forecast errors and are close to zero for accurate forecasts, indicates that our forecasts one, two and three years ahead are indeed close to zero and therefore accurate ranging from 0 to 2 per cent. However, for four years ahead, the bias proportion increases to 11 per cent. The co-variance proportion of the MSE UC also points to the accuracy of our forecasts one to three years ahead with a value close to unity of 98 per cent for one to three years ahead but only 89 per cent for the four years ahead forecasts. Furthermore, based on the R 2 of forecasts which indicates accuracy with values close to unity, 98, 94, 85 and only 73 per cent of the variation of the actual outcomes is captured by the forecasts for one, two, three, and four years ahead, respectively. Overall, the shorter the prediction period, the more accurate our forecasts; this result is also confirmed by the RMSE which increases largely with the time horizon of the forecasts. In every GET report, the global and regional forecasts are accompanied with a confidence interval to acknowledge uncertainty around the baseline forecast, particularly during the economic crisis. Therefore, we also compare our errors with these confidence intervals i.e. at the country level. We found that at the global level, one year ahead forecasts, 87 per cent of the errors were within the confidence interval 80 out of 92; for two years ahead forecast 84 per cent fall into the confidence interval 72 out of 86; for three years ahead 65 per cent also were not larger than the confidence interval 53 out of 82; and for four years ahead only 52 per cent of the errors lied within the confidence interval 32 out of 61. These results do not come as a surprise; the effects of the crisis, which had particularly severe consequences on unemployment, enter the model as a huge exogenous shock, which pushes