Implementation of news analysis in JDemetra+

The weights associated to each piece of news are calculated as follows:

$\begin{equation} \underbrace{\left[ w^{k,h}_{1} \ldots w^{k,h}_{J}\right]}_{w} = \underbrace{E[y_{k,t+h}\mathcal{I}^{'}]}_{A} \underbrace{E[\mathcal{I}\mathcal{I}^{'}]^{-1}}_{(LL')^{-1}} \end{equation}$

In order to calculate the weights, we use the Kalman smoother for the determination of the precision matrices, which are an essential part of the formula:

The covariance between the target variable and the news vector has the following form:

$\begin{equation} E[y_{k,t+h}\mathcal{I}^{'}]=\left[ \begin{array}{c} \Lambda_{k} E \left[ \left( \mathrm{f_{t_{k}}}- E[\mathrm{f_{t_{k}}}|\mathcal{F}^{old}] \right)\left( \mathrm{f_{t_{1}}}- E[\mathrm{f_{t_{1}}}|\mathcal{F}^{old}] \right) \right] \Lambda^{'}_{i_{1}} \\ \Lambda_{k} E \left[ \left( \mathrm{f_{t_{k}}}- E[\mathrm{f_{t_{k}}}|\mathcal{F}^{old}] \right)\left( \mathrm{f_{t_{2}}}- E[\mathrm{f_{t_{2}}}|\mathcal{F}^{old}] \right) \right] \Lambda^{'}_{i_{2}} \\ \vdots \\ \Lambda_{k} E \left[ \left( \mathrm{f_{t_{k}}}- E[\mathrm{f_{t_{k}}}|\mathcal{F}^{old}] \right)\left( \mathrm{f_{t_{J}}}- E[\mathrm{f_{t_{J}}}|\mathcal{F}^{old}] \right) \right] \Lambda^{'}_{i_{J}} \end{array} \right] \end{equation}$

The covariance of the news vector can be written as follows:

$\begin{equation} E[\mathcal{I}\mathcal{I}^{'}]_{j,l}= \Lambda_{i_{j}} E \left[ \left( \mathrm{f_{t_{j}}}- E[\mathrm{f_{t_{j}}}|\mathcal{F}^{old}] \right)\left( \mathrm{f_{t_{l}}}- E[\mathrm{f_{t_{l}}}|\mathcal{F}^{old}] \right) \right] \Lambda^{'}_{i_{l}} + E[\mathrm{e_{i_{j},t_{j}}}\mathrm{e_{i_{l},t_{l}}}] \end{equation}$

Our implementation can be summarized in a few simple steps:

computeNewsCovariance() computes a lower-triangular matrix lcov_, which is the Choleski factor of $E[\mathcal{I}\mathcal{I}^{'}]$
weights(int series, TsPeriod p) first stores $E[y_{k,t+h}\mathcal{I}^{'}]$ in a DataBlock a, and then it solves the system without the need to invert $E[\mathcal{I}\mathcal{I}^{'}]$
The weights are computed by solving two systems, which can be represented in matrix notation:
- First, calculate $B=wL$ by solving $LB’ =A’$ using the rsolve method contained in the class lower triangular
- Second, calculate $w$ by solving $wL=B$ using the lsolve method