Applications of state space models in finance
Applications of state space models in finance
Applications of state space models in finance
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
7.4 Empirical results 147<br />
Table 7.7: Average out-<strong>of</strong>-sample errors across sectors for multiple factor <strong>models</strong>.<br />
Mean absolute error (×10 2 ) Mean squared error (×10 4 )<br />
KF RLS RR5 RR1 KF RLS RR5 RR1<br />
Average error 1.270 1.401 1.371 1.345 3.143 4.051 3.767 3.505<br />
Average rank 1.00 3.67 3.00 2.33 1.00 3.89 3.06 2.06<br />
for t = 164, . . . , 683 with T = 520 where ˆ Ri,t represents the series <strong>of</strong> return forecasts for<br />
sector i def<strong>in</strong>ed as<br />
ˆRi,t = ˆ β ′<br />
i,t|t−1f t , (7.23)<br />
The betas employed to generate the return forecasts for date t are based only on <strong>in</strong>formation<br />
available at time t − 1. For all specifications, betas are predicted naïvely with<br />
the respective latest estimate <strong>of</strong> beta be<strong>in</strong>g used as the correspond<strong>in</strong>g one-step ahead<br />
prediction. The series <strong>of</strong> factor realizations, f t , are assumed to be known with perfect<br />
foresight, which represents a common modus operandi <strong>in</strong> empirical research to isolate<br />
the impact <strong>of</strong> conditionality.<br />
The result<strong>in</strong>g mean error measures are summarized <strong>in</strong> Table 7.7 which reports the<br />
respective averages together with the correspond<strong>in</strong>g average ranks each across all sectors<br />
for all four employed specifications; for a detailed sectoral breakdown, see Table C.8 <strong>in</strong><br />
the appendix. For the sample under consideration, the Kalman filter based betas yield<br />
the relatively best out-<strong>of</strong>-sample forecasts. For both error measures, the Kalman filter<br />
based specification ranks first for all eighteen sectors. Compared to the least squares<br />
alternatives, the mean absolute error is between 5.9% and 10.4% lower. When us<strong>in</strong>g the<br />
mean squared error as evaluation criterion, the superiority <strong>of</strong> the Kalman filter based<br />
betas becomes even more obvious: compared to the RLS, RR5 and RR1 alternatives,<br />
the average MSE is 28.9%, 19.9% and 11.5% lower, respectively.<br />
7.4.3 Practical relevance <strong>of</strong> time-variation <strong>in</strong> factor load<strong>in</strong>gs<br />
The results above <strong>in</strong>dicate that the proposed dynamic specification, where conditional<br />
betas are modeled as <strong>in</strong>dividual random walks estimated via the Kalman filter, produces<br />
the relatively most accurate estimates <strong>of</strong> time-vary<strong>in</strong>g factor load<strong>in</strong>gs. This subsection<br />
exam<strong>in</strong>es whether the statistical superiority <strong>of</strong> the proposed methodology can<br />
be exploited <strong>in</strong> practice, either <strong>in</strong> the pric<strong>in</strong>g <strong>of</strong> risk or from a portfolio management<br />
perspective.<br />
7.4.3.1 Risk pric<strong>in</strong>g<br />
As outl<strong>in</strong>ed <strong>in</strong> Subsection 7.2.2, a common procedure to analyze how well the chosen<br />
set <strong>of</strong> risk factors expla<strong>in</strong>s the cross-section <strong>of</strong> assets is to employ the Fama-MacBeth<br />
cross-sectional regression approach. Based on the proposed set <strong>of</strong> explanatory variables,<br />
(7.11) implies that the cross-section <strong>of</strong> sector returns follows a l<strong>in</strong>ear factor model <strong>of</strong> the