Quantile Cost Calculations

The purpose of this vignette is to present the calculations for a peicewise quantile regression where for each time step there are multiple independent observations.

In the follow variables identified by Greek letters are considered unknown.

Quantile regression

Data belongs to group k whose time stamps are the set t ∈ Tk which have common regression parameters θk and residual variance σk At time step t the vector of iid observations yt = {yt, 1, …, tt, nt} is explained by the design matrix Xt.

For a given quantile τ and using the check function ρ(u, τ) = u(τ − I(u < 0)) Koenker and Bassett (1978) show that an estimate of θ in QR model can be obtained by solving the convex optimization problem $$ \min_{\theta} \left( \sum_{i=1}^{n_{t}} \rho\left(\mathbf{y}_{t,i}- \mathbf{X}_{t,i}\left(\mathbf{m}_{t}+\theta_{k}\right),\tau\right) \right) $$

Solving this gives the maximum likelihood estimator of the asymmetric Laplace (AL) distributions (Geraci and Bottai, 2007 and Yu, Lu, and Stander, 2003) which has likelihood $$ L\left(\mathbf{y}_{t} \left| \theta_k\right.\right) = \tau^{n_{t}}\left(1-\tau\right)^{n_{t}}\exp\left(- \sum_{t=1}^{n_{t}} \rho\left(\mathbf{y}_{t,i}- \mathbf{X}_{t,i}\left(\mathbf{m}_{t}+\theta_{k}\right),\tau\right) \right) $$

With $\hat{\mathbf{y}}_{t} = \mathbf{y}_{t} - \mathbf{X}_{t} \mathbf{m}_{t}$ the log likelihood is given by $$ l\left(\mathbf{y}_{t} \left| \theta_k,\sigma_k \right.\right) = n_{j}\log \left(\tau \left(1-\tau\right)\right) - \sum_{i=1}^{n_{t}} \rho\left(\hat{\mathbf{y}}_{t,i} - \mathbf{X}_{t,i}\theta_{k},\tau\right) $$

The log-likelihood of yt ∈ Tk is with $n_{k}=\sum\limits_{t\in T_{k}} n_{t}$ $$ l\left(\mathbf{y}_{t \in T_{k}} \left| \theta_k,\sigma_k,\mathbf{X}_{t}\right.\right) = n_{k}\log\left(\tau \left(1-\tau\right)\right) - \sum_{t \in T_{k}}\sum_{i=1}^{n_{t}} \rho\left(\hat{\mathbf{y}}_{t,i} - \mathbf{X}_{t,i}\theta_{k},\tau\right) $$

with the cost being twice the negative log likelihood plus a penalty β giving

$$ C\left(\mathbf{y}_{t \in T_{k}} \left| \mu_t,m_k,\sigma_k,s_k\right.\right) = \sum_{t \in T_{k}}\sum_{i=1}^{n_{t}} \rho\left(\hat{\mathbf{y}}_{t,i} - \mathbf{X}_{t,i}\theta_{k},\tau\right) - 2n_{k}\log\left(\tau \left(1-\tau\right)\right) + \beta $$

Baseline: No Anomaly

Here θk = 0 and is no penalty so β = 0

Collective anomaly

Estimate θk using ??? and then with penalty β $$ C\left(\mathbf{y}_{t \in T_{k}} \left| \mu_t,m_k,\sigma_k,s_k\right.\right) = \sum_{t \in T_{k}}\sum_{i=1}^{n_{t}} \rho\left(\hat{\mathbf{y}}_{t,i} - \mathbf{X}_{t,i}\hat{\theta}_{k},\tau\right) - 2n_{k}\log\left(\tau \left(1-\tau\right)\right) + \beta $$

Point Anomaly

if nt > 0 then could proceed like a collective anomaly. Otherwise select θ̂ such that $\mathbf{y}}_{t,i} - \mathbf{X}_{t,i}\hat{\theta}_{k}= 0$