CLAIM C3 of entrylib:2026.000011 Multi-step estimators and shrinkage effect in time series models licence CC-BY-4.0; contributions C3 DEFINITIONS D1 [theory] SSOE state space model: An additive-error linear state space model of the form $$\begin{aligned} y_t &= w' v_{t-1} + \epsilon_t \\ v_t &= F v_{t-1} + g \epsilon_t \end{aligned}$$ where $y_t$ is the observed value at time $t$, $v_t$ is the state vector, $F$ is the transition matrix, $g$ is the persistence vector, $w$ is the measurement vector, and $\epsilon_t \sim \text{i.i.d. } \mathcal{N}(0, \sigma_1^2)$, with multi-step coefficients $c_j = w' F^{j-1} g$. [source B1] D2 [method] MSE_1: Conventional one-step-ahead mean squared error estimator minimising $$MSE_1 = \frac{1}{T-1} \sum_{t=1}^{T-1} e_{t+1|t}^2$$ where $e_{t+1|t} = y_{t+1} - \hat{\mu}_{t+1|t}$ is the one-step-ahead in-sample forecast error over $T$ observations. D6 [method] GTMSE: Geometric Trace Mean Squared Error estimator introduced in this paper, minimising the sum of logarithmic multi-step mean squared errors across horizons $1$ to $h$: $$GTMSE = \sum_{j=1}^{h} \log \left( \frac{1}{T-h} \sum_{t=1}^{T-h} e_{t+j|t}^2 \right)$$ D7 [method] General Predictive Likelihood: General Predictive Likelihood for consecutive observations $Y_t = (y_{t+1}, \dots, y_{t+h})'$ under Gaussian errors, whose concentrated log-likelihood is $$l_{GPL_h}(\theta, \hat{\Sigma} \mid Y) = -\frac{T-h}{2} \left( h \log(2\pi e) + \log|\hat{\Sigma}| \right)$$ where $\hat{\Sigma} = \frac{1}{T-h} \sum_{t=1}^{T-h} E_t E_t'$ is the sample multi-step forecast error covariance matrix. D8 [method] Generalised Variance: Determinant of the sample multi-step forecast error covariance matrix, $$GV = |\hat{\Sigma}|$$ whose minimisation is equivalent to maximising the [[D7:General Predictive Likelihood]]. CLAIMS C3 [formal; principal; support: formal_proof E2; assumes C19, C21; uses C7, C13; limitations L3; verification unchecked; verifiability open] Maximising the [[D7:General Predictive Likelihood]] yields parameter estimates identical to those of the conventional one-step-ahead estimator [[D2:MSE_1]] under correct model specification. scope: conditions: correctly specified linear state space models with Gaussian errors | excludes: misspecified models where |\hat{A}| \neq 1 extends B5, at p. 77: Clements and Hendry (1998) formulate the multi-step predictive log-likelihood and note diagonal covariance under exact assumptions. C7 [formal; supporting; support: formal_proof E1; assumes C19; verification unchecked; verifiability open] In an [[D1:SSOE state space model]] with independent homoscedastic errors, the covariance $\sigma_{i,j}$ between $i$-step-ahead and $j$-step-ahead forecast errors is given by $$\sigma_{i,j} = \begin{cases} \sigma_1^2 c_{|i-j|} & \text{when } i=1 \text{ or } j=1 \\ \sigma_1^2 \left(c_{|i-j|} + \sum_{l=1}^{i-1} c_{|l-j|} c_{|i-l|}\right) & \text{otherwise} \end{cases}$$ so multi-step forecast errors have non-zero cross-horizon covariance whenever $c_j \ne 0$. contradicts B8, at p. 11: Martinez (2017) assumes zero correlation between multi-step forecast errors when assumptions hold. C13 [formal; supporting; support: formal_proof E2; assumes C19, C21; verification unchecked; verifiability open] The logarithm of the [[D8:Generalised Variance]] decomposes as $\log GV = \log s_1^2 + \log |\hat{A}|$, where $|\hat{A}| = 1$ under correct model specification. C19 [formal; supporting; assumption (statistical, unchecked); verifiability not_applicable] The one-step-ahead innovation errors $\epsilon_t$ of the state space model are independent and identically distributed with zero mean and constant variance $\sigma_1^2$. C21 [formal; supporting; assumption (statistical, unchecked); verifiability not_applicable] The one-step-ahead innovation errors $\epsilon_t$ follow a Normal distribution $\mathcal{N}(0, \sigma_1^2)$. EVIDENCE E1 [formal; foundation Kolmogorov probability axioms and linear algebra] Analytical derivation in the single source of error (SSOE) state space framework showing that multi-step loss functions minimise multi-step forecast error variances and covariances, shrinking dynamic state-space parameters toward zero (and regression parameters not at all because $g=0$), along with specific derivations for ARIMA(1,1,1) and pure additive ETS models (Sections 2, 3, and Appendices A–B). S1 \eta_{t+h|t} = y_{t+h} - \mu_{t+h|t} = \sum_{j=1}^{h-1} c_j \epsilon_{t+j} + \epsilon_{t+h}, \quad c_j = w' F^{j-1} g (by D1; definition) S2 \sigma_h^2 = \sigma_1^2 \left(1 + \sum_{j=1}^{h-1} c_j^2\right), \quad \sigma_{i,j} = \sigma_1^2 \left(c_{|i-j|} + \sum_{l=1}^{i-1} c_{|l-j|} c_{|i-l|}\right) \text{ for } i,j > 1 (by S1, C19; variance and covariance of linear combinations) S3 \text{Minimising } MSE_h, TMSE, \text{ or } MSCE \text{ minimises } s_1^2 \text{ and } \hat{c}_j^2 \text{ toward } 0 \text{ for } h > 1, \text{ whereas } g=0 \text{ in static regression implies } c_j = 0 \text{ and no shrinkage.} (by S2; sample substitution and minimisation) E2 [formal; foundation Kolmogorov probability axioms and multivariate normal likelihood theory] Analytical derivation of concentrated predictive log-likelihoods for MSE_h, MSCE, and the General Predictive Likelihood (GPL), proving that log GV = log s_1^2 + log |\hat{A}| and that |\hat{A}| = 1 under correct specification, and defining GTMSE via diagonal restriction on \hat{\Sigma} (Section 4 and Appendices C–E). S1 l_{GPL_h}(\theta, \hat{\Sigma} \mid Y) = -\frac{T-h}{2}\left(h \log(2\pi e) + \log|\hat{\Sigma}|\right), \quad GV = |\hat{\Sigma}| (by D7, C21; concentrating multivariate normal log-likelihood) S2 \log GV = \log s_1^2 + \log|\hat{A}|, \quad |\hat{A}| = 1 \text{ under correct model specification} (by S1; matrix determinant reduction) S3 GTMSE \propto \sum_{j=1}^{h} \log(s_j^2) = h \log(s_1^2) + \sum_{j=2}^{h} \log\left(1 + \sum_{i=1}^{j-1} \hat{c}_i^2\right) (by S1, D6; diagonal restriction and logarithmic scaling) LIMITATIONS L3 [model_specification; affects C3; could_reverse; mitigation none] The equivalence between General Predictive Likelihood and one-step-ahead MSE_1 relies on correct model specification so that $|\hat{A}| = 1$; under model misspecification, off-diagonal covariances do not cancel and GPL estimates diverge from MSE_1. REFERENCES B1 10.1111/j.2517-6161.1985.tb01355.x (doi) HOW TO CITE This claim: Svetunkov, I. (guarantor) (2026). Claim C3, in: Multi-step estimators and shrinkage effect in time series models. EntryLib. entrylib:2026.000011@v1#C3. https://entrylib.org/p/2026.000011/C3 The claimset it is from: Svetunkov, I. (guarantor) (2026). Multi-step estimators and shrinkage effect in time series models. EntryLib. entrylib:2026.000011@v1. https://entrylib.org/p/2026.000011@v1