Percentage improvement of the QLIKE/Stein statistics, applied to a hold-out sample (see Section 4.3) when using a variable as the only kernel weighting variable, compared to a simple historical average (equal weights). The evaluation is undertaken for the kernel models using RVCM and a recursive sample scheme. This exercise is repeated every 264 trading days. The results for the variables not represented in this Figure are qualitatively comparable to one of the included variables. Eigenvalue Ratios (equation (10)) is similar to the included MVQlike; Sign Differences (equation (12)) is similar to the Temperature in Dubai; Risk Premium and Oil Price are similar to the Yield Spread and VIX and Gold Price are similar to MVQlike.  

Contexts in source publication

Context 1

... described in Section 4.3 this involves an in-sample comparison of forecasts based on a comparison of using the jth variable to an historical average forecast. Figure 4 illustrates the percentage improvement in fit (as measured by the QLIKE/Stein measure) of a selection of variables 39 . ...

Context 2

... the numerical optimisation of the multivariate bandwidths can run for long periods without converging on an optimal solution as any uninformative variables have bandwidths large enough to make all densities for that variable equal and hence have discriminatory power. Figure 4 is based on forecasts of the RVCM with a recursive sampling scheme. The results remain qualitatively very similar when using the RMVK rather than the RVCM as a proxy for the variance covariance matrix. ...

Context 3

... using the rolling sampling scheme the results are again qualitatively fairly similar. What the analysis of univariate improvements in Figure 4 illustrates is how important the respective weighting variables are when considered in isolation. ...

Context 4

... one stark difference between the use of RMVK and RVCM in the univariate kernels occurs in the variable selection exercise for the 2007-10 sample period. The rolling sample exhibits sig- nificantly reduced improvements of the univariate kernels relative to the historical rolling average, during the 2007-10 period, thus all of the lines in fig 4 dip towards the x-axis in the 2007-2010 period when using RVCM rather than RMVK. Otherwise the results illustrated in Figure 4 can be thought of as being a good representation of univariate kernel behaviour across sampling schemes and methods of obtaining observations of the VCM. ...

Context 5

... rolling sample exhibits sig- nificantly reduced improvements of the univariate kernels relative to the historical rolling average, during the 2007-10 period, thus all of the lines in fig 4 dip towards the x-axis in the 2007-2010 period when using RVCM rather than RMVK. Otherwise the results illustrated in Figure 4 can be thought of as being a good representation of univariate kernel behaviour across sampling schemes and methods of obtaining observations of the VCM. ...

Context 6

... particular the Yield Spread, MVQLIKE and the VIX are influential during the 2002-2003 period and the MVQLIKE, the element wise difference (ElemDiff ), the VIX and the Gold Price are important between 2009 and 2012. These findings are consistent with those obtained from evaluating the importance of individual weighting variables in Figure 4. ...