betaFeedback

WorksComputer Science

MORA: Correcting Time Series Forecasts with Coarse Future Signals

Abstract

Recent architectural advances have improved time-series forecasting. However, our analysis of forecasters with diverse architectures reveals a common pattern: a substantial share of residual energy is concentrated in a small number of low-frequency components. Simple linear probes further show that the forecasters' historical inputs still contain predictive information about errors in the forecast mean. Building on these findings, we propose Model-Output Residual Adaptation (MORA), a model-agnostic residual adapter that refines the outputs of frozen forecasters without requiring additional inputs. MORA estimates segment-wise means of the target sequence through structural extrapolation and learned estimation, rather than directly predicting backbone-specific residuals. It then converts discrepancies between these estimates and the corresponding statistics of the backbone forecast into stagewise corrections, scaling each update according to estimation reliability. Extensive experiments on twelve benchmarks with six forecasting backbones show average MSE reductions of 7.2% for long-term forecasting and 28.1% for short-term forecasting, with lower MSE in 96.2% of the 288 evaluated cases. The benefits also extend to three time-series foundation models evaluated at different context lengths. Together, these results demonstrate that predictable coarse-scale residuals can be effectively exploited to improve existing forecasts. Code repository: https://anonymous.4open.science/r/MORA-B0B0/.

Code repository: https://anonymous.4open.science/r/MORA-B0B.

introduction

Time-series forecasting aims to predict how a system will evolve over time based on historical observations and plays an important role in applications such as energy management, weather forecasting, and traffic flow prediction [29, 22]. Advances in deep learning have driven rapid progress in this area. Early approaches relied mainly on recurrent neural networks and temporal convolutional networks [13, 4], followed by a growing variety of Transformer- and MLP-based models [21, 17, 26]. These models capture temporal patterns and dependencies in different ways and have continued to improve forecasting performance.

However, recent work often pursues further gains through new architectural designs or more complex modules [9, 15], while the improvements are typically incremental [20, 6]. This leads us to ask a different question: rather than continuing to refine the model architecture, do the residuals of existing forecasters exhibit systematic patterns? We analyze the residuals of forecasters with diverse architectures and find that residual energy is concentrated in a small number of low-frequency components. As shown in Figure 1(a), on Weather, the first eight DCT components account for roughly 80% of the residual energy for each of the three forecasters, despite comprising only 1/12 of the frequency dimensions. Since these components capture overall level and slow variation, this result indicates that coarse-scale deviations constitute a substantial part of the remaining forecasting error.

Figure 1. Residual analysis of three forecasters on Weather. (a) Residual energy in the first eight DCT components and the remaining components. (b) Test samples are grouped by the output of a history-based linear probe. Probe estimates and observed future-mean residuals are shown as group averages. Here, the future-mean residual refers to the true future mean minus the predicted future mean.

We next examine whether the same observed history contains clues to these coarse-scale errors. Keeping each forecaster fixed, we fit a simple linear probe using the first eight DCT coefficients of its input history to predict the difference between the true and predicted means over the forecast horizon. Figure 1(b) further shows that, after grouping test samples by the probe output, all three forecasters exhibit the same overall pattern: larger group-average estimates generally correspond to larger observed group-mean residuals. Thus, a simple linear probe using only the history already seen by the forecaster can still distinguish groups that are over- or underpredicted on average. Similar patterns are observed on other datasets, with additional results provided in Appendix § 13.

This association suggests scope for history-dependent correction beyond a constant offset. Assuming finite second moments, a nonzero population covariance between the residual along a fixed coarse-scale direction and a scalar function of history implies the existence of an affine correction with strictly lower expected MSE than the best constant correction along that direction (Proposition 1). In practice, however, estimation error can offset this potential gain, motivating corrections that account for the reliability of the estimates.

Building on this, we propose Model-Output Residual Adaptation (MORA), a model-agnostic residual adapter that operates directly on the output of a frozen backbone, without modifying the backbone or requiring additional inputs. MORA consists of two components: Coarse Future Estimation and Stagewise Residual Adaptation. The former represents the future at a coarse temporal resolution using segment-wise means over contiguous intervals, and estimates these statistics from the same input history through structural extrapolation and learned estimation. Because these estimates target coarse-scale statistics of the future itself rather than backbone-specific residuals, they decouple future-state estimation from backbone-specific correction and become residual signals only when compared with the backbone forecast. The latter compares these estimates with the corresponding coarse statistics of the backbone forecast and, guided by estimation reliability, progressively converts their discrepancies into structured trajectory corrections. Rather than relearning the entire future trajectory, MORA uses the backbone forecast as its starting point and corrects only those coarse-scale deviations for which the observed history still contains predictive information.

We evaluate MORA on twelve benchmarks spanning long- and short-term forecasting, using six representative forecasting backbones. Across 288 settings, MORA reduces prediction error in 96.2% of cases, as measured by MSE, with average MSE reductions of 7.2% on the long-term benchmarks and 28.1% on the short-term benchmarks. We further evaluate MORA on three time-series foundation models with frozen parameters, yielding average MSE reductions of 24.0% and 6.9% at context lengths of 96 and 512, respectively. These gains are also accompanied by lower MAE on average. Taken together, these results show that MORA is broadly effective across model architectures, forecast horizons, and context lengths.

In summary, our main contributions are as follows:

  • By analyzing residuals from forecasters with diverse architectures, we show that the remaining errors exhibit pronounced coarse-scale structure and that the same input history used for forecasting still contains predictive signals about these errors.

  • We propose MORA, a model-agnostic residual adapter that separates coarse future estimation from backbone-specific correction. MORA refines the outputs of frozen forecasters in stages, scaling each correction according to the reliability of the underlying estimate.

  • Extensive experiments on twelve forecasting benchmarks show that MORA reduces MSE in 96.2% of 288 long- and short-term forecasting settings across six representative backbones. Results on three time-series foundation models further demonstrate its broad effectiveness.

methodology

Problem Setup

Given a historical window X∈RL×C\mathbf{X} \in\mathbb{R}^{L \times C} and its corresponding future sequence Y∈RH×C\mathbf{Y} \in\mathbb{R}^{H \times C}, where LL, HH, and CC denote the lookback length, forecast horizon, and number of variables, respectively, a pretrained forecaster fθf_{\theta} produces an HH-step prediction

Y^(0)=fθ(X)∈RH×C.(1)\widehat{\mathbf{Y}}^{(0)} = f_{\theta}(\mathbf{X}) \in\mathbb{R}^{H \times C}. \tag*{(1)}

We freeze all parameters of fθf_{\theta} and learn an output-space residual adapter that refines the backbone prediction through a structured correction,

Y^=Y^(0)+Δϕ(X,Y^(0)),(2)\widehat{\mathbf{Y}} = \widehat{\mathbf{Y}}^{(0)} + \Delta_{\phi}\left(\mathbf{X}, \widehat{\mathbf{Y}}^{(0)}\right), \tag*{(2)}

where Δϕ(X,Y^(0))∈RH×C\Delta_{\phi}\left(\mathbf{X}, \widehat{\mathbf{Y}}^{(0)}\right) \in\mathbb{R}^{H \times C} estimates the predictable component of the residual Y−Y^(0)\mathbf{Y} - \widehat{\mathbf{Y}}^{(0)}. The adapter is fitted offline with the backbone frozen. At inference, its parameters remain fixed, and it operates on the standard forecasting query and the backbone prediction without modifying the backbone or extending its input window.

Model-Output Residual Adaptation

We introduce Model-Output Residual Adaptation (MORA), illustrated in Figure 2. MORA comprises two components: Coarse Future Estimation and Stagewise Residual Adaptation. The first estimates a coarse representation of the future rather than backbone-specific residuals. The second compares these estimates with the backbone forecast at the same coarse scale and translates their discrepancies into structured residual updates. Estimates based on the current level and recurring patterns guide an initial update; a learned estimate is then compared with the updated forecast to refine the remaining residual. The resulting correction is calibrated, reconstructed over the forecast horizon, and added to the backbone prediction.

Figure 2. Overview of MORA. (i) A frozen backbone produces the initial forecast. (ii) Temporal extrapolation and learned coarse estimation provide complementary estimates of future segment means. (iii) These estimates guide two sequential residual updates, with the second stage addressing the discrepancy remaining after the first-stage correction. The two updates are combined, calibrated, and reconstructed into a correction that is added to the initial forecast. Curves and matrix entries are schematic; channel indices are omitted for clarity.

Coarse Future Estimation

MORA first estimates a coarse representation of the future rather than a residual tied to a particular backbone. For channel cc, let xc=X:,c∈RL\mathbf{x}_c = \mathbf{X}_{:,c} \in\mathbb{R}^{L} and yc=Y:,c∈RH\mathbf{y}_c = \mathbf{Y}_{:,c} \in\mathbb{R}^{H} denote its historical and future trajectories, respectively. We partition the forecast horizon into QQ contiguous, non-overlapping intervals {Bj}j=1Q\{\mathcal{B}_j\}_{j=1}^{Q} of approximately equal length and define a fixed linear operator A∈RQ×H\mathbf{A} \in\mathbb{R}^{Q \times H} such that

qc=Ayc,qj,c=1∣Bj∣∑t∈Bjyt,c.(3)\mathbf{q}_c = \mathbf{A}\mathbf{y}_c,\qquad q_{j,c} = \frac{1}{\lvert\mathcal{B}_j \rvert}\sum_{t \in\mathcal{B}_j} y_{t,c}. \tag*{(3)}

Here qc∈RQ\mathbf{q}_c \in\mathbb{R}^{Q} captures the evolution of the future at a coarser temporal scale. We choose Q=min⁡{Qmax⁡,⌈H/bmin⁡⌉}Q = \min\{Q_{\max}, \lceil H/b_{\min} \rceil\}, where Qmax⁡Q_{\max} caps the representation dimension and bmin⁡b_{\min} sets the minimum temporal span of each interval. In our experiments, Qmax⁡=64Q_{\max} = 64 and bmin⁡=4b_{\min} = 4.

We estimate the same coarse representation in two ways. First, we extrapolate simple temporal structure from the observed history. Let τ\boldsymbol{\tau} denote the temporal coordinates of the forecast timestamps, and let Cj(xc,τ)∈RH\mathcal{C}_j(\mathbf{x}_c,\boldsymbol{\tau}) \in\mathbb{R}^{H} denote the jj-th extrapolated trajectory. These extrapolations use the current level, recent-pattern continuation, and, when enabled, a periodic template fitted on the training data. Each extrapolation yields

mj,ce=ACj(xc,τ)∈RQ,j=1,…,J.(4)\mathbf{m}_{j,c}^{e} = \mathbf{A}\mathcal{C}_j(\mathbf{x}_c,\boldsymbol{\tau}) \in\mathbb{R}^{Q},\qquad j = 1,\ldots,J. \tag*{(4)}

We retain the JJ estimates separately and stack them as mce∈RJQ\mathbf{m}^{\mathrm{e}}_c \in\mathbb{R}^{JQ}. The exact extrapolation rules and periodic-template construction are given in Appendix § 9.§ 1.

These temporal extrapolations provide low-capacity coarse future estimates under explicit temporal assumptions. We further introduce a learned predictor for the same coarse future target qcq_c. For target channel cc, we construct a deterministic feature vector ϕc(X,τ)∈RD\phi_c(\mathbf{X},\tau) \in\mathbb{R}^{D} from the multivariate history X\mathbf{X} and known temporal context τ\tau, where DD denotes the feature dimension. The feature vector summarizes low-order DCT components [1] of the history, whole-window and block-level statistics, multi-scale recent dynamics, and periodic time encodings. For L=96L = 96, the compact representation has D=78D = 78 dimensions. We fit ridge regressors [8] separately for each target channel to map these historical features to the QQ-dimensional coarse future quantity. To capture both the absolute level and changes relative to the current state, we use two coordinate systems: one retains the original coordinates, while the other is centered at the latest observation. With xL,cx_{L,c} denoting the latest observation of channel cc, the centered history and target are

x~c=xc−xL,c1L,q~c=qc−xL,c1Q,(5)\widetilde{\mathbf{x}}_c = \mathbf{x}_c - x_{L,c}\mathbf{1}_L,\qquad\widetilde{\mathbf{q}}_c = \mathbf{q}_c - x_{L,c}\mathbf{1}_Q, \tag*{(5)}

where 1n\mathbf{1}_n denotes the nn-dimensional all-ones vector. Let mcabs∈RQ\mathbf{m}^{\mathrm{abs}}_c \in\mathbb{R}^{Q} denote the prediction in the original coordinates and m~crel∈RQ\widetilde{\mathbf{m}}^{\mathrm{rel}}_c \in\mathbb{R}^{Q} the prediction in the centered coordinates. After restoring the latter to the original level, the learned coarse estimate is

mcℓ=12mcabs+12(m~crel+xL,c1Q).(6)\mathbf{m}^{\ell}_c = \frac{1}{2}\mathbf{m}^{\mathrm{abs}}_c + \frac{1}{2}\left(\widetilde{\mathbf{m}}^{\mathrm{rel}}_c + x_{L,c}\mathbf{1}_Q\right). \tag*{(6)}

Thus, the learned predictor estimates the future coarse representation qcq_c, rather than a backbone-specific residual. The exact feature construction, feature and target standardization, and ridge-fitting details are provided in Appendix § 9.§ 1.

For series with many variables, fitting a separate learned estimator for every channel makes the parameter count and fitting cost scale linearly with CC. We therefore replace the channel-wise estimator with a factorized coarse estimator based on a shared low-dimensional channel-factor model. Specifically, we first estimate a channel-wise periodic baseline from the training data by averaging each channel at every phase of a configured period. After subtracting this baseline and standardizing the remaining multivariate deviations, we project them into the shared factor space. A shared ridge model predicts the coarse future states of these factors, which are then mapped back to the original channels with the channel-wise mean, scale, and periodic baseline restored. The resulting channel-level estimates retain the form mcℓ∈RQm_c^\ell\in\mathbb{R}^Q, leaving the subsequent residual adaptation unchanged. Details are provided in Appendix C.2.

Backbone-relative residual signal. The coarse future estimates become backbone-relative by comparing them with the backbone forecast at the same temporal scale. For any individual coarse estimate mc∈RQm_c \in\mathbb{R}^Q of qcq_c,

dc=mc−Ay^c(0)=A(yc−y^c(0))+(mc−Ayc).(7)d_c = m_c - A\hat{y}_c^{(0)} = A\left(y_c - \hat{y}_c^{(0)}\right) + (m_c - Ay_c). \tag*{(7)}

The first term is the backbone residual at the same coarse scale, while the second is the error of the coarse future estimate. Thus, a coarse future estimate can be converted into a backbone-relative residual signal by subtracting the corresponding coarse prediction of the backbone.

Stagewise Residual Adaptation

Correction subspace. The coarse representation does not uniquely determine a pointwise correction. We therefore restrict the correction to a predefined low-dimensional subspace over the forecast horizon, reducing its degrees of freedom while retaining the backbone forecast as the reference prediction. Let U∈RH×kU \in\mathbb{R}^{H \times k} contain an orthonormal basis for this correction subspace, where kk is the correction rank and UTU=IkU^\mathsf{T}U = I_k. For the backbone residual rc=yc−y^c(0)r_c = y_c - \hat{y}_c^{(0)}, we write

rc=Uzc+r⊥,c,zc=UTrc,(8)r_c = Uz_c + r_{\perp,c}, \qquad z_c = U^\mathsf{T}r_c, \tag*{(8)}

where zc∈Rkz_c \in\mathbb{R}^k contains the coordinates of the residual within the correction subspace, and r⊥,c∈RHr_{\perp,c} \in\mathbb{R}^H is the orthogonal remainder.

MORA estimates these residual coordinates as z^c\hat{z}_c and reconstructs the correction within the span of UU. We construct UU by tapering low-order inverse discrete cosine transform (IDCT) modes near the forecast boundary and orthonormalizing the resulting columns. The first row of UU is zero, so the correction leaves the backbone’s first predicted value unchanged. For long-term forecasting, we use k=64k = 64; construction details are provided in Appendix C.3.

For any individual coarse estimate mc∈RQm_c \in\mathbb{R}^Q, the backbone-relative discrepancy satisfies

dc=AUzc+ηc,ηc=(mc−qc)+Ar⊥,c.(9)d_c = AUz_c + \eta_c, \qquad\eta_c = (m_c - q_c) + Ar_{\perp,c}. \tag*{(9)}

Here ηc∈RQ\eta_c \in\mathbb{R}^Q contains the coarse-estimation error and the contribution of residuals outside the correction space. Thus, the discrepancy dcd_c provides a noisy signal of the residual coordinates zcz_c.

Residual refinement. Rather than combining the extrapolation-based and learned estimates in a single update, MORA performs residual adaptation in two stages. The extrapolation-based estimates first produce an initial correction, after which the learned estimate is evaluated against the updated forecast to refine the residual discrepancy.

Let Ae∈RJQ×HA^e \in\mathbb{R}^{JQ \times H} denote JJ vertically stacked copies of AA. The channel-specific matrices Ke∈Rk×JQK^e \in\mathbb{R}^{k \times JQ} and Kℓ∈Rk×QK^\ell\in\mathbb{R}^{k \times Q} map coarse discrepancies to residual coordinates. The diagonal matrices Ge,Gℓ∈Rk×kG^e, G^\ell\in\mathbb{R}^{k \times k} provide coordinate-wise reliability scaling shared across channels. The ordered updates are

dce=mce−Aey^c(0),z^ce=GeKedce,dcℓ=mcℓ−A(y^c(0)+Uz^ce),z^cℓ=GℓKℓdcℓ.(10)\begin{aligned} d_c^e &= m_c^e - A^e\hat{y}_c^{(0)}, & \hat{z}_c^e &= G^eK^e d_c^e, \\ d_c^\ell&= m_c^\ell- A\left(\hat{y}_c^{(0)} + U\hat{z}_c^e\right), & \hat{z}_c^\ell&= G^\ell K^\ell d_c^\ell. \tag*{(10)} \end{aligned}

Here, z^ce\hat{z}_c^e and z^cℓ\hat{z}_c^\ell denote the coefficient-space residual updates from the first and second corrections, respectively. The learned discrepancy is then computed against the updated forecast y^c(0)+Uz^ce\hat{y}_c^{(0)} + U\hat{z}_c^e, so the second update targets the residual remaining after the extrapolation-based correction.

Final calibration and prediction. The two ordered updates produce a combined correction z^ce+z^cℓ\hat{z}_c^e + \hat{z}_c^\ell in the kk-dimensional coefficient space. Before reconstructing this correction over the forecast horizon, we calibrate its magnitude to account for differences across channels and basis-coordinate ranges. For channel cc, the kk basis coordinates are partitioned into four consecutive groups, with one validation-fitted scale shared within each group. We collect these scales in a diagonal matrix Sc∈Rk×k\mathbf{S}_{c} \in\mathbb{R}^{k \times k}. Thus, although Sc\mathbf{S}_{c} is written as a k×kk \times k matrix, it contains only four free scalar parameters.

The final forecast is

y^c=y^c(0)+USc(z^ce+z^cℓ)(11)\hat{\mathbf{y}}_{c} = \hat{\mathbf{y}}_{c}^{(0)} + \mathbf{U}\mathbf{S}_{c}\left(\hat{\mathbf{z}}_{c}^{e} + \hat{\mathbf{z}}_{c}^{\ell}\right) \tag*{(11)}

The matrix Sc\mathbf{S}_{c} calibrates the combined residual coefficients before U\mathbf{U} maps them back to the forecast horizon. This calibration is applied only after both residual updates and does not affect the intermediate forecast used by the learned stage. Stacking the channel-wise forecasts in Eq. equation 11 yields Y^\hat{\mathbf{Y}}, while the corresponding corrections form the residual adapter Δϕ(X,Y^(0))\Delta_{\phi}(\mathbf{X}, \hat{\mathbf{Y}}^{(0)}) introduced in Eq. (2).

Fitting the Residual Adapter

For the coarse estimates from temporal extrapolation, the residual mapping is estimated using pairs (dce,zc)(\mathbf{d}_{c}^{e}, \mathbf{z}_{c}). For the learned coarse estimate, we define the remaining target as zcℓ=zc−z^ce\mathbf{z}_{c}^{\ell} = \mathbf{z}_{c} - \hat{\mathbf{z}}_{c}^{e}. The corresponding discrepancy dcℓ\mathbf{d}_{c}^{\ell} is computed against the forecast corrected by z^ce\hat{\mathbf{z}}_{c}^{e}, following Equation equation 10. The learned residual mapping is then estimated using pairs (dcℓ,zcℓ)(\mathbf{d}_{c}^{\ell}, \mathbf{z}_{c}^{\ell}).

The channel-specific matrices Kce\mathbf{K}_{c}^{e} and Kcℓ\mathbf{K}_{c}^{\ell} are computed from regularized second-order statistics of the corresponding discrepancy–target pairs. The gains Ge\mathbf{G}^{e} and Gℓ\mathbf{G}^{\ell} are estimated separately using a coordinate-wise reliability rule shared across channels. The complete estimators, regularization, and reliability-scaling rules are provided in Appendix § 9.§ 4.

The learned coarse predictor is fitted on the training split. For estimating the residual mappings, we use blocked out-of-fold (OOF) predictions on the same split to construct the required statistics and the corresponding z^ce\hat{\mathbf{z}}_{c}^{e} used to form zcℓ\mathbf{z}_{c}^{\ell}. Once Kce\mathbf{K}_{c}^{e}, Kcℓ\mathbf{K}_{c}^{\ell}, Ge\mathbf{G}^{e}, and Gℓ\mathbf{G}^{\ell} are fixed, the final calibration Sc\mathbf{S}_{c} is fitted on validation data using the calibration objective defined above. The complete fitting protocol is deferred to Appendix § 9.5.

Experiments

Experimental Setup

Datasets and backbones. We evaluate MORA on eight long-horizon forecasting datasets, ETTh1, ETTh2, ETTm1, ETTm2, Weather, ECL, Traffic, and Solar, and four short-horizon traffic datasets, PEMS03, PEMS04, PEMS07, and PEMS08. All datasets are split chronologically into training, validation, and test sets. ETT follows the standard 12/4/4-month split; the other long-horizon datasets and PEMS use ratios of 7:1:2 and 6:2:2, respectively. We select six representative backbones spanning classical architectures and recent state-of-the-art methods: DLinear (Zeng et al. [28]), PatchTST (Nie et al. [21]), iTransformer (Liu et al. [17]), TimeFIlter (Hu et al. [9]), CGTFra (Yu et al. [27]), and VPNet (Liu et al. [15]). For long-horizon forecasting, we additionally evaluate three pretrained time-series foundation models: Chronos-2 (Ansari et al. [3]), Sundial (Liu et al. [18]), and TimesFM 2.5 (Das et al. [5]; Google Research [7]). More details on the datasets and backbones are provided in Appendices § 10.§ 1 and § 10.§ 2.

Implementation details. Following common forecasting protocols (Liu et al. [17]), we use a lookback length of L=96L = 96 for task-specific models, with H∈{96,192,336,720}H \in\{96, 192, 336, 720\} for long-horizon forecasting and H∈{12,24,48,96}H \in\{12, 24, 48, 96\} for short-horizon forecasting. To examine adaptation with longer histories, foundation models are evaluated at both L=96L = 96 and L=512L = 512, with an adapter fitted separately for each context length. All backbone parameters remain frozen during adaptation. MORA’s coarse estimators and residual maps are fitted on training data, followed by final calibration on validation forecasts; neither the backbone nor the adapter is updated at test time. For ECL, Traffic, and PEMS, we use the factorized coarse estimator, while the remaining datasets use the channel-wise estimator. We report MSE and MAE, and use Imp. to denote the percentage error reduction relative to the corresponding baseline. More details on hyperparameters, fitting, and computing resources are given in Appendix § 10.§ 3; results for individual horizons and seeds are provided in Appendix § 12.

DatasetTimeFilter MSETimeFilter MAEiTransformer MSEiTransformer MAECGTFra MSECGTFra MAEPatchTST MSEPatchTST MAEDLinear MSEDLinear MAEVPNet MSEVPNet MAE
ETTh1Base0.4180.4270.4550.4470.4350.4280.4480.4490.4620.4590.4360.429
+MORA0.4100.4230.4480.4420.4380.4300.4280.4360.4390.4450.4330.427
Imp. (%)1.9850.9781.5101.196-0.546-0.3344.6732.7615.0482.9260.7370.314
ETTh2Base0.3720.4030.3830.4070.3760.3970.3900.4130.5620.5190.3590.385
+MORA0.3700.3990.3720.3980.3690.3930.3670.3970.4870.4660.3630.389
Imp. (%)0.5080.9252.8952.0781.7341.0695.7523.99713.41010.098-1.044-1.106
ETTm1Base0.3770.3940.4070.4110.3930.3900.3870.4020.4030.4080.3760.382
+MORA0.3570.3850.3690.3940.3600.3800.3620.3900.3620.3860.3550.378
Imp. (%)5.1842.0889.3174.2798.2952.5446.6162.93610.2615.2095.6951.111
ETTm2Base0.2710.3210.2900.3350.2780.3180.2930.3360.3520.4000.2740.315
+MORA0.2640.3170.2700.3220.2650.3130.2750.3270.3030.3540.2650.311
Imp. (%)2.6371.0597.0733.8534.7241.7036.2652.64213.82511.5013.4231.205
WeatherBase0.2400.2700.2590.2790.2400.2620.2580.2800.2670.3190.2380.262
+MORA0.2270.2650.2310.2750.2220.2570.2310.2770.2360.2940.2280.259
Imp. (%)5.5221.68310.9701.6117.3561.75810.6241.17611.6487.5684.0211.188
ECLBase0.1640.2610.1750.2670.1670.2550.2040.2940.2090.2950.1640.252
+MORA0.1560.2560.1600.2590.1590.2520.1740.2750.1770.2770.1560.250
Imp. (%)4.8271.9438.5712.9914.6391.02614.6266.63015.3376.1324.7841.158
TrafficBase0.4080.2690.4230.2820.4300.2570.4810.3080.6250.3850.4210.274
+MORA0.4000.2660.4110.2770.4240.2560.4570.2970.5370.3430.4130.271
Imp. (%)1.9311.0512.7671.8371.5990.0855.1463.60414.12310.6961.8860.883
SolarBase0.2270.2580.2350.2620.2260.2310.2590.3000.3300.4010.2040.211
+MORA0.1980.2500.1980.2510.2100.2300.2050.2700.2120.2780.1990.212
Imp. (%)12.9103.04315.8294.2826.9170.25320.63810.04435.82730.6772.557-0.369

Table 1. Baseline and MORA results for six backbones on eight long-term forecasting datasets. MSE and MAE are averaged over H∈{96,192,336,720}H \in\{96, 192, 336, 720\} at L=96L = 96. In Tables 1–3, bold marks the lower displayed error within each baseline–MORA comparison.

DatasetTimeFilter MSETimeFilter MAEiTransformer MSEiTransformer MAECGTFra MSECGTFra MAEPatchTST MSEPatchTST MAEDLinear MSEDLinear MAEVPNet MSEVPNet MAE
PEMS03Base0.1010.2100.1150.2240.1090.2130.2060.2950.2450.3410.1170.227
+MORA0.0980.2060.1140.2160.0990.2030.1270.2320.1370.2410.1070.215
Imp. (%)3.0192.0081.1783.4949.2464.92838.29121.59244.10329.3978.2015.479
PEMS04Base0.0950.2010.1130.2230.1070.2100.2560.3330.2460.3500.1110.224
+MORA0.0790.1800.0900.1930.0820.1830.0970.1960.1050.2010.0840.187
Imp. (%)16.70610.46720.27713.70023.06612.86262.32141.04157.48242.59324.27916.433
PEMS07Base0.0840.1840.1010.2040.0960.1890.2200.3010.2880.3810.1010.211
+MORA0.0710.1670.0910.1780.0780.1660.0860.1850.0980.1970.0760.176
Imp. (%)15.4599.52010.03012.59318.25012.04460.71438.64065.85148.31724.44916.480
PEMS08Base0.0980.2030.1520.2270.1450.2260.2100.2960.2390.3460.1060.215
+MORA0.0850.1870.1290.2100.1100.2010.1050.2090.1140.2210.0880.193
Imp. (%)13.2347.75414.7977.28123.79611.09250.22429.30052.49936.13016.81110.205

Table 2. Baseline and MORA results for six backbones on the four PEMS datasets. MSE and MAE are averaged over H∈{12,24,48,96}H \in\{12, 24, 48, 96\} at L=96L = 96.

Forecasting Results

Main Results. Tables 1 and 2 report horizon-averaged results on long- and short-term forecasting, respectively. Across the 192 dataset–backbone–horizon settings on the long-term benchmarks, MORA reduces MSE in 187 cases (97.4%). Averaged across the eight datasets, MORA lowers the MSE of DLinear and PatchTST by 14.9% and 9.3%, respectively. The gains remain evident for more competitive recent forecasters, with average MSE reductions of 4.4% for TimeFilter and 4.3% for CGTFra. Overall, across all six backbones, MORA reduces MSE by 7.2% on average. Similar improvements are observed on the four short-term PEMS benchmarks, where MORA improves both MSE and MAE for all 24 dataset–backbone combinations in the horizon-averaged results. Across the six backbones, average MSE reductions range from 11.6% to 55.0%. Notably, the benefit persists for TimeFilter, a strong recent backbone: MORA further reduces its MSE by 3.0%–16.7% across the four datasets, averaging 12.1% relative to the same frozen checkpoints. Together, these results show that MORA complements diverse forecasting architectures in both long- and short-term settings, including strong backbones, without updating their weights.

Time-Series Foundation Models. Table 3 reports the results of MORA on three time-series foundation models using their officially released pretrained checkpoints. Since all three models support substantially longer context windows at inference, we evaluate L=512L = 512 in addition to L=96L = 96 to examine whether MORA remains effective when the backbone is provided with more historical information. Each adapted forecast is compared against its unadapted counterpart using the same context length. At L=96L = 96, MORA reduces MSE by 22.3%, 20.8%, and 28.8% for Chronos-2, Sundial, and TimesFM 2.5, respectively, averaged equally across the eight datasets. Even at L=512L = 512, where the foundation models can draw on substantially more historical context, MORA yields further average MSE reductions of 6.3%, 6.0%, and 8.5% for the same three models. MORA also reduces MAE on average at both context lengths. Together with the results on task-specific forecasters, these findings show that MORA provides a broadly applicable output-space refinement mechanism whose benefits persist across model families and context lengths.

DatasetChronos-2SundialTimesFM 2.5
L=96L = 96L=512L = 512L=96L = 96L=512L = 512L=96L = 96L=512L = 512
MSEMAEMSEMAEMSEMAEMSEMAEMSEMAE
ETTh1Base0.4920.4390.4180.4120.4730.4350.3970.4120.4840.435
+MORA0.4560.4310.4020.4150.4360.4280.3870.4090.4350.429
Imp. (%)7.3701.8803.900-0.7537.7391.6102.4630.68210.1571.439
ETTh2Base0.4060.4050.3520.3770.3990.4100.3440.3830.4130.406
+MORA0.3840.3930.3450.3770.3890.4050.3400.3780.3970.404
Imp. (%)5.4133.0382.0310.0052.5981.0681.0011.1703.8950.512
ETTm1Base0.6160.4770.3870.3720.5370.4570.3660.3700.8750.584
+MORA0.3660.3910.3380.3600.3690.3900.3380.3650.3700.390
Imp. (%)40.49618.12712.7873.16331.32914.6327.5211.38957.68233.247
ETTm2Base0.3180.3380.2800.3110.3040.3400.2750.3200.3720.378
+MORA0.2750.3150.2570.2990.2710.3200.2520.3080.2770.321
Imp. (%)13.3766.9468.1823.57410.9165.8588.1623.75425.50014.896
WeatherBase0.3070.2910.2580.2520.2850.3020.2440.2750.3140.299
+MORA0.2430.2900.2130.2500.2330.2910.2130.2620.2300.283
Imp. (%)20.8070.33017.4670.76118.4073.80312.8904.85326.7045.584
ECLBase0.2340.2940.1630.2440.2040.2840.1480.2430.2390.308
+MORA0.1840.2740.1540.2410.1730.2700.1430.2400.1870.283
Imp. (%)21.5896.8815.0261.02214.9104.9712.8990.97922.0018.226
TrafficBase0.5740.3130.3920.2380.6040.3550.4580.2900.5990.334
+MORA0.5180.3010.3890.2390.5310.3270.4410.2830.5440.316
Imp. (%)9.7963.6870.846-0.17112.1437.9293.7342.1319.1195.503
SolarBase0.5280.3270.1870.1930.6660.5300.2290.2570.9130.520
+MORA0.2150.2860.1870.1940.2100.2780.2080.2560.2220.279
Imp. (%)59.28112.5210.174-0.19668.42147.4919.2120.17075.70746.260

Table 3. Baseline and MORA results for three frozen time-series foundation models at L=96L = 96 and L=512L = 512. MSE and MAE are averaged over H∈{96,192,336,720}H \in\{96, 192, 336, 720\} separately for each context length.

Ablation Study

We first conduct component ablations to examine the roles of MORA’s coarse future estimates and stagewise residual updates. We then compare alternative adaptation strategies to assess the benefit of deriving residual corrections from coarse future estimates.

StrategyETTm1 MSEETTm1 MAEWeather MSEWeather MAESolar MSESolar MAE
TimeFilter0.3130.3540.1530.1990.1990.239
Affine calibration0.3140.3570.1530.2280.2000.239
Direct residual0.3000.3570.1500.2110.1810.239
δ\delta-Adapter (O)0.3700.4150.1560.2240.1920.243
δ\delta-Adapter (I+O)0.3120.3560.1530.2010.1710.234
MORA0.3040.3500.1480.1980.1730.231

Table 5. Alternative strategies on TimeFilter.

O: output-only; I+O: input–output.

Component Ablations. Table 4 examines the contributions of the two sources of coarse future estimates and the design of the second-stage correction. Removing either structural extrapolation or the learned estimator increases both MSE and MAE on all three datasets. Although both sources target the same coarse future quantities, neither alone matches the performance of the full adapter, indicating that they provide complementary correction signals. With both sources retained, we compare the stagewise design against an independent-update variant in which both corrections are fitted to the original backbone residual. The stagewise design achieves slightly lower MSE and MAE on all three datasets. This result suggests that the learned correction benefits from accounting for the first-stage correction, allowing it to target the remaining residual rather than fitting another correction to the original backbone residual.

VariantETTm1 MSEETTm1 MAEETTm2 MSEETTm2 MAEECL MSEECL MAE
TimeFilter0.4520.4370.3900.3930.2060.302
w/o learned0.4270.4310.3890.3900.1940.291
w/o extrapolation0.4360.4270.3800.3910.1950.294
Independent0.4270.4290.3790.3890.1910.290
MORA0.4220.4260.3760.3880.1900.288

Table 4. Component ablations on TimeFilter.

Comparison with Alternative Strategies. Table 5 compares MORA with affine calibration, direct residual regression, and two variants of δ-Adapter [14]. The direct regressor uses the same historical features and correction basis as MORA, with a comparable parameter count, but predicts the backbone residual directly. Affine calibration improves neither MSE nor MAE. The learned alternatives reduce MSE in several settings, but their gains do not consistently extend to MAE. Direct residual regression achieves the lowest MSE on ETTm1, while δ-Adapter (I+O) does so on Solar. MORA achieves the lowest MAE on all three datasets while reducing MSE relative to the unadapted backbone in every case. These results suggest that deriving corrections from coarse future estimates provides an effective adaptation strategy, yielding consistent improvements in both metrics across the evaluated settings. Additional ablations and experimental details are provided in Appendix E.

Low-Frequency Residual Analysis

To examine whether MORA effectively corrects low-frequency forecasting errors, we visualize the low-frequency residuals of DLinear and TimeFilter on Solar before and after adaptation. As shown in Figure 3, MORA reduces low-frequency reconstruction MSE by 41.0% and 13.9%, respectively. The corresponding overall forecasting MSE decreases by 38.9% and 12.8%. The reduction in low-frequency error accounts for nearly all of the overall MSE improvement in these settings, supporting the intended role of MORA’s coarse future estimates in correcting coarse-scale deviations.

Figure 3. Low-frequency residuals before and after applying MORA on Solar. DCT-8 residuals are represented by their signed projection onto a shared principal direction fitted on the validation set, together with the norm of the orthogonal residual. The squared distance from the origin corresponds to the low-frequency MSE.

Conclusion

We presented MORA, a model-agnostic residual adapter for frozen time-series forecasters. Motivated by the coarse-scale structure of forecasting errors, MORA estimates segment-wise future means and uses their discrepancies with the corresponding means of backbone forecasts to construct stagewise corrections, scaled by estimation reliability. Experiments on long- and short-term forecasting benchmarks demonstrate broad improvements across diverse architectures, including time-series foundation models at different context lengths. These findings highlight coarse future estimation as an effective complement to existing forecasters.

AI Use Statement

We used generative AI tools to assist with dataset preparation and reformatting, manuscript organization, drafting, translation, language editing, and LaTeX typesetting. We carefully reviewed all AI-assisted content and take full responsibility for the final manuscript.

Reproducibility Statement

We describe MORA’s coarse future estimation and stagewise residual adaptation in the main paper. The estimator definitions, correction subspace, and fitting and calibration procedures are detailed in Appendix § 9. Dataset splits, preprocessing, backbone configurations, hyperparameters, and computing resources are documented in Appendix § 10. Additional ablations and horizon-wise results are provided in Appendices § 11 and § 12, respectively. The anonymous code repository is linked in the abstract.

Related Work

Time-Series Forecasting. Early neural forecasting models used recurrent and convolutional networks to capture temporal patterns, as exemplified by LSTNet [13]. Subsequent Transformer-based approaches extended this line of research to long-horizon prediction: Informer [29] improved attention efficiency, while Autoformer [24] incorporated progressive trend–seasonal decomposition. Spectral structure became another useful modeling tool, with FEDformer combining decomposition and frequency-domain representations and FiLM using Fourier filtering alongside Legendre memory [30 b]. Later work explored simpler predictors and alternative representations. DLinear [28] revisited decomposition-based linear forecasting, and PatchTST [21] combined temporal patching with channel-independent modeling. FreTS [26] and FITS [25] further developed spectral forecasting through frequency-domain MLPs and complex-valued interpolation, respectively, while iTransformer [17] used variate tokens to model cross-variate dependencies. More recent approaches have refined multivariate dependency modeling through patch-specific graph filtration in TimeFilter [9], local cross-variate interactions in VPNet [15], and information-theoretic alignment in CGTFra [27]. Across these designs, temporal and cross-variate structure is incorporated into the predictor through its architecture or representation.

Time-Series Foundation Models. Large-scale pretraining has broadened forecasting beyond dataset-specific model development, allowing a single model to learn from heterogeneous time-series collections and transfer to new tasks. Chronos [2] formulates forecasting as language modeling over quantized time-series values, whereas TimesFM [5] adopts a patched decoder-only architecture. Moirai [23] further supports heterogeneous sampling frequencies and variable dimensions within a unified forecasting model. Subsequent developments have expanded both the modeling approach and the range of supported tasks: Sundial [18] uses flow matching for continuous-valued generative forecasting, while Chronos-2 [3] supports univariate, multivariate, and covariate-informed forecasting through group attention. These models provide reusable forecasting backbones that can be applied without task-specific fine-tuning, creating a natural setting for studying how their predictions can be refined without modifying pretrained parameters.

Forecast Adaptation and Post-Processing. Beyond backbone design and pretraining, another line of research improves forecasts by addressing non-stationarity or adapting existing predictions. RevIN [12] removes and restores instance-level statistics, while Non-stationary Transformers [16] combine series stationarization with de-stationary attention to preserve informative non-stationary structure. These mechanisms are incorporated into the forecasting pipeline rather than fitted as standalone residual corrections. Test-time adaptation instead adjusts a deployed forecasting system as new observations become available. TAFAS [11] uses partially observed targets for online calibration, and COSA [10] updates a gated output-space residual adapter using recently observed ground truth. Closer to our setting, δ\delta-Adapter [14] learns lightweight input perturbations and output residual corrections around a frozen forecaster. MORA also keeps the backbone frozen, but separates coarse future estimation from backbone-specific residual adaptation. Its coarse estimators predict segment-wise means of the future from the same historical window. Backbone-specific residual mappings then translate discrepancies between these estimates and the corresponding coarse statistics of the current forecast into reliability-scaled, stagewise corrections. Unlike TAFAS and COSA, MORA is fitted offline, and both the backbone and adapter remain fixed throughout testing.

History-Dependent Residual Correction

This appendix proves Proposition 1 from the introduction. The result connects a history-dependent residual signal to a reduction in expected MSE beyond the best constant correction along the same coarse-scale direction. We first give the population least-squares argument, then relate it to the forecast-mean probe and the residual coordinates used by MORA.

Improvement Beyond a Constant Correction

Following the problem setup, keep the pretrained forecaster fθf_\theta frozen and consider channel cc, with backbone residual rc=yc−y^c(0)∈RH\mathbf{r}_c = \mathbf{y}_c - \hat{\mathbf{y}}_c^{(0)} \in\mathbb{R}^{H}. Let u∈RH\mathbf{u} \in\mathbb{R}^{H} be a fixed unit vector along a coarse-scale direction, and let p=g(X)p = g(\mathbf{X}) be a scalar signal from the same historical window available to the backbone. The function gg need not be linear; one example is the fixed linear probe used in the introduction. All expectations below refer to the joint distribution of (X,Y)(\mathbf{X}, \mathbf{Y}), with fθf_\theta and gg held fixed. Define

z=uTrc,μz=E[z],μp=E[p],σp2=Var⁡(p),γ=Cov⁡(z,p).(12)z = \mathbf{u}^{\mathsf{T}}\mathbf{r}_c,\qquad\mu_z = \mathbb{E}[z],\qquad\mu_p = \mathbb{E}[p],\qquad\sigma_p^2 = \operatorname{Var}(p),\qquad\gamma= \operatorname{Cov}(z,p). \tag*{(12)}

Here zz is a scalar projection of the backbone residual; its relation to the vector zc=UTrc\mathbf{z}_c = \mathbf{U}^{\mathsf{T}}\mathbf{r}_c is given in Appendix B.3. Assume E∥rc∥22<∞\mathbb{E}\lVert\mathbf{r}_c\rVert_2^2 < \infty and E[p2]<∞\mathbb{E}[p^2] < \infty. For coefficients a,b∈Ra,b \in\mathbb{R}, the corrected forecast is y^c(0)+u(a+bp)\hat{\mathbf{y}}_c^{(0)} + \mathbf{u}(a + bp), with expected MSE

Rc(a,b)=1HE[∥rc−u(a+bp)∥22].(13)R_c(a,b) = \frac{1}{H}\mathbb{E}\left[\lVert\mathbf{r}_c - \mathbf{u}(a + bp)\rVert_2^2\right]. \tag*{(13)}

The factor 1/H1/H averages the error over the forecast horizon. A constant correction has b=0b = 0: its coefficient is the same for every query, while its temporal shape is fixed by u\mathbf{u}.

Proposition 1 (History-dependent correction beyond a constant offset).** Under the assumptions above, suppose σp2>0\sigma_p^2 > 0. The best constant correction along u\mathbf{u} has coefficient μz\mu_z. The optimal affine correction has

b⋆=γσp2,a⋆=μz−b⋆μp,a⋆+b⋆p=μz+γσp2(p−μp).(14)b^\star= \frac{\gamma}{\sigma_p^2},\qquad a^\star= \mu_z - b^\star\mu_p,\qquad a^\star+ b^\star p = \mu_z + \frac{\gamma}{\sigma_p^2}(p-\mu_p). \tag*{(14)}

Its improvement over the best constant correction is exactly

min⁡aRc(a,0)⏟Rconst−min⁡a,bRc(a,b)⏟Raff=γ2Hσp2.(15)\underbrace{\min_a R_c(a,0)}_{R_{\mathrm{const}}} - \underbrace{\min_{a,b} R_c(a,b)}_{R_{\mathrm{aff}}} = \frac{\gamma^2}{H\sigma_p^2}. \tag*{(15)}

Consequently, Cov⁡(z,p)≠0\operatorname{Cov}(z,p) \ne0 implies a strict reduction in expected MSE beyond the best constant correction along that direction.

Proof. Since ∥u∥2=1\lVert\mathbf{u}\rVert_2 = 1 and uT(rc−uz)=0\mathbf{u}^{\mathsf{T}}(\mathbf{r}_c-\mathbf{u}z)=0, the squared error decomposes pointwise as

∥rc−u(a+bp)∥22=∥rc−uz∥22+(z−a−bp)2.(16)\lVert\mathbf{r}_c-\mathbf{u}(a+bp)\rVert_2^2 = \lVert\mathbf{r}_c-\mathbf{u}z\rVert_2^2 + (z-a-bp)^2. \tag*{(16)}

The first term is unaffected by either coefficient. Centering zz and pp gives

E[(z−a−bp)2]=Var⁡(z)−2bγ+b2σp2+(a+bμp−μz)2.(17)\mathbb{E}[(z-a-bp)^2] = \operatorname{Var}(z) - 2b\gamma+ b^2\sigma_p^2 + (a+b\mu_p-\mu_z)^2. \tag*{(17)}

For any fixed bb, the last term is minimized at a=μz−bμpa = \mu_z - b\mu_p. Completing the square in the remaining terms yields

Var⁡(z)−2bγ+b2σp2=Var⁡(z)−γ2σp2+σp2(b−γσp2)2.(18)\operatorname{Var}(z) - 2b\gamma+ b^2\sigma_p^2 = \operatorname{Var}(z) - \frac{\gamma^2}{\sigma_p^2} + \sigma_p^2\left(b-\frac{\gamma}{\sigma_p^2}\right)^2. \tag*{(18)}

Because σp2>0\sigma_p^2 > 0, the minimizing slope is b⋆=γ/σp2b^\star= \gamma/\sigma_p^2. Restricting instead to b=0b = 0 gives the best constant coefficient a=μza = \mu_z. Their minimum scalar squared errors differ by γ2/σp2\gamma^2/\sigma_p^2. The common orthogonal term cancels, and division by HH proves Eq. (15). ■

If σp2=0\sigma_p^2 = 0, the affine and constant correction classes coincide; if σp2>0\sigma_p^2 > 0 but γ=0\gamma= 0, the chosen signal yields no additional improvement over the best constant correction within the affine class.

Connection to the Forecast-Mean Probe

For the forecast-mean probe, take u=1HH\mathbf{u} = \frac{\mathbf{1}_{H}}{\sqrt{H}} and define the future-mean residual rˉc=H−11HTrc\bar{r}_{c} = H^{-1}\mathbf{1}_{H}^{\mathsf{T}}r_{c}, namely the true future mean minus the predicted future mean. Then z=Hrˉcz = \sqrt{H}\bar{r}_{c}, and Proposition 1 gives

Rconst−Raff=Cov⁡(rˉc,p)2Var⁡(p).(19)\mathcal{R}_{\mathrm{const}} - \mathcal{R}_{\mathrm{aff}} = \frac{\operatorname{Cov}(\bar{r}_{c},p)^{2}}{\operatorname{Var}(p)}. \tag*{(19)}

The corresponding optimal correction is

u(a⋆+b⋆p)=1H[E[rˉc]+Cov⁡(rˉc,p)Var⁡(p)(p−E[p])].(20)\mathbf{u}(a^{\star} + b^{\star}p) = \mathbf{1}_{H}\left[\mathbb{E}[\bar{r}_{c}] + \frac{\operatorname{Cov}(\bar{r}_{c},p)}{\operatorname{Var}(p)}\left(p-\mathbb{E}[p]\right)\right]. \tag*{(20)}

This history-dependent level correction is added to every forecast step. It can therefore reduce the MSE of the full future sequence without changing the deviations around the predicted future mean.

Because p=g(X)p = g(\mathbf{X}), the law of total expectation also gives

Cov⁡(z,p)=Cov⁡(E[z∣X],p).(21)\operatorname{Cov}(z,p) = \operatorname{Cov}(\mathbb{E}[z \mid\mathbf{X}],p). \tag*{(21)}

Nonzero covariance rules out a conditional mean E[z∣X]\mathbb{E}[z \mid\mathbf{X}] that is constant almost surely. Hence, the historical window contains predictive information about the residual beyond a constant offset. The diagnostic probe differs from MORA’s learned coarse estimator, which predicts the future quantity qc=Aycq_{c} = A y_{c} rather than a backbone-specific residual.

Connection to Stagewise Residual Adaptation

The correction subspace in Section 2.2.2 makes the connection to the residual coordinates explicit. For any fixed v∈Rk\mathbf{v} \in\mathbb{R}^{k} with ∥v∥2=1\lVert\mathbf{v}\rVert_{2} = 1, orthonormality of UU gives

u=Uv,∥u∥2=1,z=vTzc.(22)\mathbf{u} = U\mathbf{v}, \qquad\lVert\mathbf{u}\rVert_{2} = 1, \qquad z = \mathbf{v}^{\mathsf{T}}z_{c}. \tag*{(22)}

Thus Proposition 1 applies to a direction within MORA’s correction subspace whenever its covariance condition holds. The forecast-mean direction illustrates the motivation, whereas the tapered basis used by MORA leaves the first predicted value unchanged. A mean-probe association alone does not establish nonzero covariance for every direction in that subspace.

The remaining issue is estimation reliability. For σp2>0\sigma_{p}^{2} > 0 and any fixed fitted coefficients (a^,b^)(\hat{a},\hat{b}) in the affine family above, the difference from the best constant correction is

Rc(a^,b^)−Rconst=1H[(a^+b^μp−μz)2+σp2(b^−b⋆)2−γ2σp2].(23)\mathcal{R}_{c}(\hat{a},\hat{b}) - \mathcal{R}_{\mathrm{const}} = \frac{1}{H}\left[(\hat{a}+\hat{b}\mu_{p}-\mu_{z})^{2}+\sigma_{p}^{2}(\hat{b}-b^{\star})^{2}-\frac{\gamma^{2}}{\sigma_{p}^{2}}\right]. \tag*{(23)}

The two nonnegative terms quantify coefficient-estimation error, which can offset the population improvement. This motivates accounting for estimation reliability, as in MORA’s coordinate-wise reliability scaling GeG^{e} and GℓG^{\ell}. These gains are estimated from training data; the final calibration ScS_{c} is fitted separately on validation data, following Section 2.2.3.

Proposition 1 establishes the existence of an improving correction under the evaluation distribution. It does not guarantee that finite-sample estimates, or coefficients fitted on a different distribution, achieve that improvement. The grouped probe results provide empirical evidence relevant to the covariance condition, rather than a proof of it. Window overlap leaves the risk identity unchanged but must be considered when estimating uncertainty.

Additional Method Details

Coarse Future Estimation

Temporal extrapolation. The current-level estimate extends the latest observation over the forecast horizon. Recent-pattern continuation instead repeats the last Peff=min⁡{P,L}P_{\mathrm{eff}}=\min\{P,L\} observations, where PP is the configured base period in sampling steps. For h=1,…,Hh=1,\ldots,H,

[C1(xc,τ)]h=xL,c,[C2(xc,τ)]h=xL−Peff+1+((h−1) mod Peff),c.(24)\begin{aligned} \left[\mathcal{C}_1(x_c,\tau)\right]_h &= x_{L,c},\\ \left[\mathcal{C}_2(x_c,\tau)\right]_h &= x_{L-P_{\mathrm{eff}}+1+((h-1)\bmod P_{\mathrm{eff}}),c}. \tag*{(24)} \end{aligned}

Both estimates use only the supplied historical window. When P>LP>L, recent-pattern continuation repeats the available window rather than accessing observations outside it.

Training-fitted periodic template. When enabled, the periodic estimate uses a channel-wise harmonic template fitted on the training data,

Tc(s)=a0,c+∑p∈P[ap,ccos⁡(2πs/p)+bp,csin⁡(2πs/p)].(25)T_c(s)=a_{0,c}+\sum_{p\in\mathcal{P}}\left[a_{p,c}\cos(2\pi s/p)+b_{p,c}\sin(2\pi s/p)\right]. \tag*{(25)}

Here ss denotes the absolute sampling index. The period dictionary P\mathcal{P} combines configured daily and weekly periods with separated spectral peaks identified from detrended training series. The harmonic coefficients are estimated by regularized least squares. For a query ending at index s0s_0, we estimate a local level offset bcloc=∑ℓ=1Lwℓ[xℓ,c−Tc(s0−L+ℓ)]b_c^{\mathrm{loc}}=\sum_{\ell=1}^{L}w_\ell[x_{\ell,c}-T_c(s_0-L+\ell)], where the normalized exponential weights wℓw_\ell give greater weight to recent observations. The extrapolated future is [C3(xc,τ)]h=Tc(s0+h)+bcloc\left[\mathcal{C}_3(x_c,\tau)\right]_h=T_c(s_0+h)+b_c^{\mathrm{loc}}. Each trajectory is mapped through AA; the resulting estimates are retained separately in mce\mathbf{m}_c^e.

The dictionary contains at most eight periods, including PP and 7P7P, with the remaining periods selected from training-only spectral peaks. Harmonic fitting uses ridge regularization of 10−310^{-3} under an average squared-loss convention. The local level offset uses normalized exponential weights with a half-life of 16 sampling steps.

Historical features. The compact feature vector contains 78 entries at L=96L=96: 16 orthonormal DCT coefficients, eight whole-window statistics, 12 block means, 36 recent-dynamics features, and six time encodings. The whole-window statistics comprise the mean, standard deviation, first and last values, endpoint difference, linear slope, and the mean and standard deviation of first differences. Recent dynamics comprise 12 block means of the last 24 observations, 12 block means of their first differences, and the mean, standard deviation, slope, and endpoint difference over the last 8, 16, and 24 observations. Time encodings are sine–cosine pairs evaluated at the forecast origin for periods PP, P/2P/2, and 7P7P.

The channel-wise estimator also uses extended feature representations. At L=96L=96, the full representation has 252 dimensions, additionally including raw history, block standard deviations, recent increments, and cross-channel summaries. Adding interactions between historical state features and temporal encodings gives 348 dimensions. These representations use the same historical window and coarse future target, without changing the stagewise residual updates.

Channel-wise ridge estimation. We fit the absolute and last-centered views separately. For view v∈{abs,rel}v\in\{\mathrm{abs},\mathrm{rel}\}, let Φcv\Phi_c^v and QcvQ_c^v collect the standardized historical features and coarse targets over the fitting samples. The channel-wise coefficient matrix is estimated as

Wcv=arg⁡min⁡W∥Qcv−ΦcvW∥F2+λv∥W∥F2.(26)\mathbf{W}_c^v=\underset{\mathbf{W}}{\arg\min}\left\|\mathbf{Q}_c^v-\Phi_c^v\mathbf{W}\right\|_F^2+\lambda_v\left\|\mathbf{W}\right\|_F^2. \tag*{(26)}

Feature and target statistics are computed on the fitting samples, with standard-deviation floors of 10−510^{-5} and 10−410^{-4}, respectively. Predictions are restored to the target scale, and the centered prediction is additionally shifted by xL,c1Qx_{L,c}\mathbf{1}_Q. The two estimates are then averaged as in Equation (6). The fitting and regularization-selection protocol is specified in Appendix § 9.5.

Factorized Coarse Estimation

Periodic baseline. Let xt∈RC\mathbf{x}_t \in\mathbb{R}^{C} denote the multivariate observation at absolute sampling index tt, and let PfP_f be the configured period. For training indices T\mathcal{T}, define Tt={s∈T:s mod Pf=t mod Pf}\mathcal{T}_t = \{s \in\mathcal{T}: s \bmod P_f = t \bmod P_f\}. The channel-wise periodic baseline is

b(t)=1∣Tt∣∑s∈Ttxs.(27)\mathbf{b}(t) = \frac{1}{|\mathcal{T}_t|} \sum_{s \in\mathcal{T}_t} \mathbf{x}_s. \tag*{(27)}

Unlike the harmonic template used for temporal extrapolation, this baseline directly averages observations sharing the same phase.

Shared factor space. We subtract the periodic baseline and standardize the remaining deviations using their training means μ\boldsymbol{\mu} and standard deviations σ\boldsymbol{\sigma}:

ξt=(xt−b(t)−μ)⊘σ,st=VTξt.(28)\boldsymbol{\xi}_t = (\mathbf{x}_t - \mathbf{b}(t) - \boldsymbol{\mu}) \oslash\boldsymbol{\sigma}, \qquad\mathbf{s}_t = \mathbf{V}^{\mathsf{T}}\boldsymbol{\xi}_t. \tag*{(28)}

Here ⊘\oslash denotes elementwise division, with standard deviations floored at 10−510^{-5}. The columns of V∈RC×df\mathbf{V} \in\mathbb{R}^{C \times d_f} are the leading dfd_f eigenvectors of the training covariance of ξt\boldsymbol{\xi}_t, satisfying VTV=I\mathbf{V}^{\mathsf{T}}\mathbf{V} = \mathbf{I}. The resulting factor states summarize variation shared across channels.

Coarse factor prediction. For a query ending at s0s_0, the factor-space target for interval Bj\mathcal{B}_j is

qjf=1∣Bj∣∑h∈Bjss0+h.(29)\mathbf{q}^{f}_j = \frac{1}{|\mathcal{B}_j|} \sum_{h \in\mathcal{B}_j} \mathbf{s}_{s_0+h}. \tag*{(29)}

The predictor uses the latest factor state; means over the last 6, 12, 24, 48, and 96 steps; differences from the states 6, 12, and 24 steps earlier; slopes over the last 24 steps; and the six time encodings described in Appendix § 9.§ 1. At L=96L = 96, these features contain 10df+610d_f+6 entries. A single multi-output ridge regressor is fitted to the concatenated absolute targets qjf\mathbf{q}^{f}_j and last-relative targets qjf−ss0\mathbf{q}^{f}_j-\mathbf{s}_{s_0} for all factors and intervals. The relative predictions are restored to absolute factor coordinates by adding ss0\mathbf{s}_{s_0}. Training-internal rolling validation selects the ridge strength and one fixed prediction rule: the absolute prediction, the restored relative prediction, or their equal-weight average. The selected rule remains fixed at inference. Feature and target standardization follows the channel-wise ridge estimator. Denote the selected prediction by q^jf\widehat{\mathbf{q}}^{f}_j.

Reconstruction in channel space. The factor predictions are mapped back to the original channels, with the channel-wise mean, scale, and periodic baseline restored:

[mcℓ]j=1∣Bj∣∑h∈Bjbc(s0+h)+μc+σcVc,:q^jf.(30)[\mathbf{m}^{\ell}_{c}]_j = \frac{1}{|\mathcal{B}_j|} \sum_{h \in\mathcal{B}_j} b_c(s_0+h) + \mu_c + \sigma_c V_{c,:}\widehat{\mathbf{q}}^{f}_j. \tag*{(30)}

We use df=32d_f = 32 and Pf=7PP_f = 7P, where P=24P = 24 for ECL/Traffic and P=288P = 288 for PEMS. All train-dependent quantities follow the fitting protocol in Appendix § 9.5. The resulting mcℓ∈RQ\mathbf{m}^{\ell}_c \in\mathbb{R}^{Q} is used in the stagewise updates without changing their formulation.

Correction Subspace

Cosine modes and boundary taper. For a horizon HH and correction rank k<Hk < H, let B∈RH×k\mathbf{B} \in\mathbb{R}^{H \times k} contain the first kk orthonormal IDCT modes. For h=1,…,Hh = 1,\ldots,H and j=0,…,k−1j = 0,\ldots,k-1,

Bh,j+1=ajcos⁡(πj(h−12)H),a0=H−1/2,aj=2H for j≥1.(31)B_{h,j+1} = a_j \cos\left(\frac{\pi j\left(h-\frac{1}{2}\right)}{H}\right), \qquad a_0 = H^{-1/2}, \qquad a_j = \sqrt{\frac{2}{H}}\ \text{for } j \ge1. \tag*{(31)}

We apply a taper over the first nbn_b forecast steps, where 2≤nb≤H2 \le n_b \le H:

whtap={sin⁡2(π2h−12nb−1),1≤h≤nb,1,nb<h≤H.(32)w^{\mathrm{tap}}_h = \begin{cases} \sin^2\left(\frac{\pi}{2}\frac{h-1}{2n_b-1}\right), & 1 \le h \le n_b,\\ 1, & n_b < h \le H. \end{cases} \tag*{(32)}

The taper starts at zero and reaches one at nbn_b. The tapered modes and their reduced QR factorization are

Btap=diag⁡(w1tap,…,wHtap)B=UR,(33)\mathbf{B}_{\mathrm{tap}} = \operatorname{diag}(w^{\mathrm{tap}}_1,\ldots,w^{\mathrm{tap}}_H)\mathbf{B} = \mathbf{U}\mathbf{R}, \tag*{(33)}

where U∈RH×k\mathbf{U} \in\mathbb{R}^{H \times k} has orthonormal columns. This orthonormalization preserves the span of the tapered modes.

Boundary preservation. The first row of Btap\mathbf{B}_{\mathrm{tap}} is zero. Since R\mathbf{R} is nonsingular, the first row of U\mathbf{U} is also zero. Consequently, the final correction satisfies

[USc(z~ce+z~cℓ)]1=0,y^1,c=y^1,c(0).(34)\left[\mathbf{U}\mathbf{S}_{c}\left(\widetilde{\mathbf{z}}^{e}_{c}+\widetilde{\mathbf{z}}^{\ell}_{c}\right)\right]_{1}=0,\qquad\widehat{y}_{1,c}=\widehat{y}^{(0)}_{1,c}. \tag*{(34)}

The constraint therefore preserves the backbone’s first predicted value; it does not force that value to equal the last observation. The basis is computed once for each configuration and shared across channels. Because tapering modifies the cosine modes, its span is not identical to an untapered DCT truncation.

Dimensions. For long-term forecasting, we use k=64k=64 and nb=13n_b=13. The fixed short-horizon settings for PEMS are listed in Table 9. These settings retain the same construction with k<Hk < H.

Forecast horizon HH12244896
Coarse intervals QQ361224
Correction rank kk8163264
Boundary taper23613

Table 9. PEMS dimensions, fixed across datasets and backbones. Boundary taper is measured in forecast steps.

Estimating the Residual Updates

Stage-specific signals. For the extrapolation-based stage, we use (dce,zce)(\mathbf{d}^{e}_{c},\mathbf{z}^{e}_{c}) and Me=AeU\mathbf{M}^{e}=\mathbf{A}^{e}\mathbf{U}. For the learned stage, we use (dcℓ,zcℓ)(\mathbf{d}^{\ell}_{c},\mathbf{z}^{\ell}_{c}) and Mℓ=AU\mathbf{M}^{\ell}=\mathbf{A}\mathbf{U}. We describe the estimator for either stage below, omitting the stage superscript. Let pp be the dimension of dc\mathbf{d}_{c} and define ηi,c=di,c−Mzi,c\eta_{i,c}=d_{i,c}-M\mathbf{z}_{i,c}.

Residual mapping. From NN training pairs, we form the uncentered second-order moments

Cab,c=1N∑i=1Nωi,cai,cbi,cT,a,b∈{z,d,η}.(35)C_{ab,c}=\frac{1}{N}\sum_{i=1}^{N}\omega_{i,c}\mathbf{a}_{i,c}\mathbf{b}_{i,c}^{\mathsf{T}},\qquad a,b\in\{z,d,\eta\}. \tag*{(35)}

No sample mean is subtracted. The weights are ωi,c=si,c−2γ\omega_{i,c}=s_{i,c}^{-2\gamma}, where si,c=max⁡{RMS⁡(di,c),sc}s_{i,c}=\max\{\operatorname{RMS}(d_{i,c}),s_c\}. The floor scs_c is the larger of 10−510^{-5} and the training tenth percentile of the discrepancy RMS. Setting γ=0\gamma=0 gives uniform weights. These weights affect moment estimation; the deployed mapping remains linear in dc\mathbf{d}_{c}.

We apply diagonal shrinkage Dβ(C)=(1−β)C+βDiag⁡(C)\mathcal{D}_{\beta}(\mathbf{C})=(1-\beta)\mathbf{C}+\beta\operatorname{Diag}(\mathbf{C}), where Diag⁡(C)\operatorname{Diag}(\mathbf{C}) retains only the diagonal entries. Writing C‾zz,c=Dβ(Czz,c)\overline{\mathbf{C}}_{zz,c}=\mathcal{D}_{\beta}(\mathbf{C}_{zz,c}) and C‾ηη,c=Dβ(Cηη,c)\overline{\mathbf{C}}_{\eta\eta,c}=\mathcal{D}_{\beta}(\mathbf{C}_{\eta\eta,c}), the regularized mapping is

Σcd=(1−ρ)(C‾zz,cMT+νC‾ηη,c)+ρCdd,c,Kc=[(1−ρ)C‾zz,cMT+ρCzd,c](Σcd+ϵcIp)−1.(36)\begin{aligned} \boldsymbol{\Sigma}^{d}_{c}&=(1-\rho)\left(\overline{\mathbf{C}}_{zz,c}\mathbf{M}^{\mathsf{T}}+\nu\overline{\mathbf{C}}_{\eta\eta,c}\right)+\rho\mathbf{C}_{dd,c},\\ \mathbf{K}_{c}&=\left[(1-\rho)\overline{\mathbf{C}}_{zz,c}\mathbf{M}^{\mathsf{T}}+\rho\mathbf{C}_{zd,c}\right]\left(\boldsymbol{\Sigma}^{d}_{c}+\epsilon_{c}\mathbf{I}_{p}\right)^{-1}. \tag*{(36)} \end{aligned}

Here β,ρ∈[0,1]\beta,\rho\in[0,1], ν>0\nu>0, and ϵc=10−6max⁡{tr⁡(Σcd)/p,10−12}\epsilon_{c}=10^{-6}\max\{\operatorname{tr}(\boldsymbol{\Sigma}^{d}_{c})/p,10^{-12}\}. The parameter ρ\rho interpolates between structured and empirical second-order estimates. The mapping is computed by solving the corresponding linear system, rather than explicitly forming its inverse.

Reliability scaling. Let z~i,c,r\widetilde{z}_{i,c,r} be the unscaled prediction for coordinate rr from a mapping fitted without the temporal block containing sample ii. For a set of training samples I\mathcal{I}, define

αr(I)=∑i∈I∑cz~i,c,rzi,c,r∑i∈I∑cz~i,c,r2.(37)\alpha_{r}(\mathcal{I})=\frac{\sum_{i\in\mathcal{I}}\sum_{c}\widetilde{z}_{i,c,r}z_{i,c,r}}{\sum_{i\in\mathcal{I}}\sum_{c}\widetilde{z}_{i,c,r}^{2}}. \tag*{(37)}

The ratio is set to zero when its denominator does not exceed 10−1210^{-12}. We use the pooled rule for Ge\mathbf{G}^{e} and the blockwise rule for Gℓ\mathbf{G}^{\ell}:

grpool=clip⁡[0,gmax⁡](αr(I)),g_{r}^{\mathrm{pool}}=\operatorname{clip}_{[0,g_{\max}]}\left(\alpha_{r}(\mathcal{I})\right),
grblock=max⁡{0,min⁡[gmax⁡,αr(I),min⁡bαr(Ib)]},(38)g_{r}^{\mathrm{block}}=\max\left\{0,\min\left[g_{\max},\alpha_{r}(\mathcal{I}),\min_{b}\alpha_{r}(\mathcal{I}_{b})\right]\right\}, \tag*{(38)}

where Ib\mathcal{I}_{b} is temporal block bb and gmax⁡=1.5g_{\max}=1.5. Each stage uses its own targets and cross-fitted predictions. Both rules pool channels, producing one gain per residual coordinate. The gains form the diagonal matrices Ge\mathbf{G}^{e} and Gℓ\mathbf{G}^{\ell} and remain fixed at inference. Parameter selection and block construction are described in Appendix C.5.

Fitting and Calibration

Training signals. We construct blocked out-of-fold (OOF) forecasts on the training split. The latter half of the training-window origins is partitioned into four contiguous blocks, and held queries are sampled at stride six. For each block, task-specific backbone replicas exclude its raw-time span and any training windows whose histories or targets overlap that span. Train-dependent coarse estimators follow the same exclusion. The replicas are used only to construct training signals; deployment uses the original frozen backbone. For foundation models, pretrained weights remain fixed throughout, while the local coarse estimators are fitted using the corresponding training partitions. Coarse-predictor ridge strengths are selected from {1,10,100,1000,10000}\{1, 10, 100, 1000, 10000\} using training-internal rolling validation with an L+H−1L + H - 1 separation between fitting and held window origins.

Stagewise fitting. The first-stage mapping is cross-fitted over the training blocks. Its held-block predictions are used to estimate GeG^{\mathrm{e}} according to Appendix § 9.§ 4. After applying this gain, the resulting first-stage coefficients define the remaining target zcℓz_{\mathrm{c}}^{\ell} and the adjusted discrepancy dcℓd_{\mathrm{c}}^{\ell} used to fit the second stage. The second-stage mapping is cross-fitted over the same blocks, and its held-block predictions determine GℓG^{\ell}. Each mapping is then refitted on all of its corresponding training pairs. The learned coarse predictor continues to target the future coarse quantities, not the remaining backbone residual. Train-dependent coarse estimators are fitted on the full training split for validation and test inference.

Regularization settings. The residual mappings use parameter values drawn from β∈{0.5,1}\beta\in\{0.5, 1\}, ν∈{0.5,1,2}\nu\in\{0.5, 1, 2\}, ρ∈{0,0.25,0.5,1}\rho\in\{0, 0.25, 0.5, 1\}, and γ∈{0,0.5}\gamma\in\{0, 0.5\}. Candidates are ranked by their minimum relative reduction in residual-coordinate squared error across training blocks, using the blockwise gain rule. Mean block improvement breaks ties. The factorized realization uses the fixed setting (β,ν,ρ,γ)=(1,1,0,0)(\beta, \nu, \rho, \gamma) = (1, 1, 0, 0) in both stages.

Validation calibration. Once both update mappings and their gains are fixed, we fit ScS_{\mathrm{c}} using validation forecasts. The kk coordinates are partitioned into four consecutive groups, with one scale sc,g∈[0,1.5]s_{\mathrm{c},g} \in[0, 1.5] per channel and group. We divide the validation queries into four contiguous blocks and minimize the equally weighted normalized objective

Lcal=14∑b=14[0.1MSEbMSEb(0)+0.9MAEb,εMAEb(0)].(39)\mathcal{L}_{\mathrm{cal}} = \frac{1}{4}\sum_{b=1}^{4}\left[0.1\frac{\mathrm{MSE}_{b}}{\mathrm{MSE}_{b}^{(0)}} + 0.9\frac{\mathrm{MAE}_{b,\varepsilon}}{\mathrm{MAE}_{b}^{(0)}}\right]. \tag*{(39)}

Each numerator is evaluated from the calibrated forecasts and averaged over queries, forecast steps, and channels within its block. Superscript (0) denotes the corresponding unadapted backbone error; both denominators are floored at 10−1210^{-12}. The smoothed absolute error replaces ∣e∣|e| by e2+ϵ2\sqrt{e^2+\epsilon^2}, with ϵ=max⁡{10−314∑bMAEb(0),10−7}\epsilon= \max\{10^{-3}\frac{1}{4}\sum_b \mathrm{MAE}_b^{(0)}, 10^{-7}\}. Starting from unit scales, we perform at most eight iteratively reweighted least-squares updates, each solving a four-dimensional box-constrained quadratic problem per channel. Iteration stops when the largest scale change is below 10−510^{-5}.

The resulting calibration is applied only to the sum of the two updates; it does not alter the intermediate forecast used by the learned stage. At inference, the backbone, coarse estimators, residual mappings, reliability gains, and calibration remain fixed.

Additional Experimental Details

Datasets and Forecasting Windows

Datasets. Table 6 summarizes the twelve benchmarks, including their dimensions, sampling intervals, forecast horizons, and split sizes. ETT [29] records oil temperature and six load measurements from two electricity transformers, with hourly and 15-minute versions. Weather (Max Planck Institute for Biogeochemistry, n.d.) contains meteorological measurements from the Max Planck Institute for Biogeochemistry in Jena. ECL, Traffic, and Solar use the LSTNet releases [13] and record electricity consumption, road occupancy, and photovoltaic generation, respectively. PEMS03/04/07/08 use the STSGCN traffic-flow releases [22]; we forecast the first feature of each sensor in the supplied arrays. The data sources are the ETT repository, MPI-BGC weather archive, LSTNet data repository, and STSGCN repository.

Chronological splits. ETT uses the first 20 benchmark months, divided into 12/4/4 months for training/validation/test; each month contains 30 days in the released loader. The remaining long-horizon datasets use 7:1:2 splits, and PEMS uses 6:2:2. Boundaries follow the corresponding loaders. Table 6 reports raw observations, not overlapping forecasting windows.

DatasetCCHorizons HHSplit sizes: Train / Val / TestSamplingDomain
ETTh1, ETTh27{96,192,336,720}\{96, 192, 336, 720\}(8640, 2880, 2880)HourlyElectricity
ETTm1, ETTm27{96,192,336,720}\{96, 192, 336, 720\}(34560, 11520, 11520)15 minElectricity
Weather21{96,192,336,720}\{96, 192, 336, 720\}(36887, 5270, 10539)10 minWeather
ECL321{96,192,336,720}\{96, 192, 336, 720\}(18412, 2632, 5260)HourlyElectricity
Traffic862{96,192,336,720}\{96, 192, 336, 720\}(12280, 1756, 3508)HourlyTransportation
Solar137{96,192,336,720}\{96, 192, 336, 720\}(36792, 5256, 10512)10 minEnergy
PEMS03358{12,24,48,96}\{12, 24, 48, 96\}(15724, 5242, 5242)5 minTransportation
PEMS04307{12,24,48,96}\{12, 24, 48, 96\}(10195, 3398, 3399)5 minTransportation
PEMS07883{12,24,48,96}\{12, 24, 48, 96\}(16934, 5645, 5645)5 minTransportation
PEMS08170{12,24,48,96}\{12, 24, 48, 96\}(10713, 3571, 3572)5 minTransportation

Table 6. Forecasting benchmarks. CC is the number of variables and HH is the forecast horizon in time steps. Split sizes count raw observations before windowing.

Long-horizon preprocessing. Each channel is standardized using its training mean and population standard deviation, with unit scale for constant channels. These statistics are reused on validation and test data. Forecast origins advance by one step; the future stays within its evaluation split, while the initial history may extend into the preceding split. Each baseline–MORA pair uses identical forecast origins, as do the foundation-model comparisons between L=96L = 96 and L=512L = 512. Incomplete evaluation batches are retained.

PEMS preprocessing. For a record of length TT, split boundaries are ⌊0.6T⌋\lfloor0.6T \rfloor and ⌊0.8T⌋\lfloor0.8T \rfloor; windows stay within their split without borrowing preceding-split history. iTransformer and CGTFra use training-fitted standardization and test stride one. TimeFilter uses split-specific standardization and test stride 12, also adopted by the PatchTST, DLinear, and VPNet transfer experiments. The latter convention uses test-split normalization statistics, but does not fit MORA parameters on test labels. Every local baseline/MORA pair shares preprocessing and window enumeration, including the endpoint convention. Absolute normalized errors are not directly comparable across the two preprocessing groups.

Models and Inference

Task-specific backbones. We use the released architectures and training objectives, with model- and dataset-specific configurations at L=96L = 96. Optimizers, batch sizes, and training budgets follow these configurations. Checkpoints are selected on validation data before adaptation and then frozen. Each local baseline–MORA pair uses the same checkpoint; backbone hyperparameters are not retuned during adaptation.

PEMS baselines. TimeFilter, iTransformer, and CGTFra reproductions are checked against the corresponding paper versions and horizon-specific references. TimeFilter at H=96H = 96 is an extension without an original-paper score. PatchTST, DLinear, and VPNet have no matching PEMS results in the selected source papers; their transfer recipes retain the original training principles while adapting to L=96L = 96, the PEMS channel count, and TimeFilter’s preprocessing. VPNet uses its Traffic recipe with patch length 12 and stride six.

Table 15 marks six quoted baseline entries with †: five from Table 9 of iTransformer ([17]) and one from the archived CGTFra public submission, rather than a verified final-ICML result table. The remaining 90 baseline entries are local reproductions, and all +MORA forecasts use local checkpoints. These six comparisons use literature references rather than paired local checkpoints. Local paired scores remain in the result records.

DatasetHHForecastTimeFilter MSETimeFilter MAEiTransformer MSEiTransformer MAECGTFra MSECGTFra MAEPatchTST MSEPatchTST MAEDLinear MSEDLinear MAEVPNet MSEVPNet MAE
PEMS0312Base0.0630.1640.0680.1740.0600.1600.0840.1910.1070.2190.0660.171
PEMS0312+MORA0.0620.1640.0690.1720.0590.1590.0750.1820.0830.1930.0660.171
PEMS0312Imp. (%)1.704-0.059-0.1681.2041.7450.57610.5314.78222.06311.975-0.578-0.060
PEMS0324Base0.0790.1860.0950.2040.0790.1840.1300.2390.1840.2900.0900.202
PEMS0324+MORA0.0770.1840.0880.1920.0750.1800.0980.2080.1090.2190.0840.194
PEMS0324Imp. (%)3.0621.0237.2995.7004.6412.12624.08312.86240.67124.4216.5814.129
PEMS0348Base0.1090.2220.125†0.236†0.119†0.228†0.2310.3240.3030.3890.1250.239
PEMS0348+MORA0.1060.2170.1260.2300.1100.2170.1420.2480.1520.2560.1150.225
PEMS0348Imp. (%)3.0461.989-1.1212.5057.3224.70638.68823.65049.88534.1038.3675.605
PEMS0396Base0.1540.2690.1710.2810.1780.2810.3790.4270.3870.4680.1850.297
PEMS0396+MORA0.1480.2580.1710.2690.1510.2550.1930.2890.2040.2960.1630.269
PEMS0396Imp. (%)3.5193.970-0.0134.14815.0859.41049.02332.42847.28636.72312.0049.496
PEMS0412Base0.0680.1670.0810.1890.0700.1690.0980.2070.1110.2260.0730.177
PEMS0412+MORA0.0640.1610.0720.1730.0660.1630.0730.1740.0760.1790.0660.165
PEMS0412Imp. (%)5.4063.68411.1348.4895.8533.90325.55715.75831.53220.9219.1276.719
PEMS0424Base0.0800.1830.0990.2100.0840.1870.1560.2640.1820.2980.0950.210
PEMS0424+MORA0.0720.1710.0820.1850.0740.1740.0860.1890.0930.1950.0780.181
PEMS0424Imp. (%)9.4946.31016.88612.03611.0826.86045.18028.57248.63534.41017.95613.426
PEMS0448Base0.1000.2090.120†0.233†0.1130.2200.2810.3640.3050.4040.1230.241
PEMS0448+MORA0.0840.1870.0960.2000.0870.1900.1050.2050.1180.2100.0900.195
PEMS0448Imp. (%)16.17310.64120.07613.95022.97113.61162.41143.51761.15148.03327.45419.035
PEMS0496Base0.1340.2470.150†0.262†0.1600.2660.4900.4970.3870.4710.1510.268
PEMS0496+MORA0.0970.2020.1090.2130.1010.2070.1220.2170.1310.2190.1010.207
PEMS0496Imp. (%)27.13717.98727.63218.56736.95822.19675.10256.35966.22153.50432.94722.874
PEMS0712Base0.0550.1500.0670.1650.0560.1470.0790.1840.1000.2200.0640.167
PEMS0712+MORA0.0530.1450.0680.1560.0550.1430.0630.1610.0710.1740.0590.155
PEMS0712Imp. (%)3.9182.748-2.2945.3821.2712.57519.73512.43029.35121.1818.1927.121
PEMS0724Base0.0680.1660.0870.1900.0750.1670.1330.2400.1860.3050.0920.204
PEMS0724+MORA0.0630.1590.0840.1710.0670.1580.0800.1810.0890.1920.0710.176
PEMS0724Imp. (%)7.4504.7203.4599.63310.0565.68840.00624.76652.01436.99422.47614.099
PEMS0748Base0.0900.1930.1100.2140.1040.1980.2520.3360.3500.4360.1070.220
PEMS0748+MORA0.0760.1750.0980.1870.0850.1750.0960.1950.1070.2030.0800.182
PEMS0748Imp. (%)15.3069.40510.54913.01418.84811.95262.02441.92269.53453.55454.57116.933
PEMS0796Base0.1220.2290.1410.2470.1480.2430.4160.4440.5170.5640.1390.252
PEMS0796+MORA0.0910.1880.1130.1990.1060.1890.1070.2020.1270.2200.0930.191
PEMS0796Imp. (%)25.25017.54519.49319.32428.42522.19174.29254.52775.43660.98533.11824.199
PEMS0812Base0.0640.1620.0810.1840.0710.1680.0840.1910.1020.2180.0670.174
PEMS0812+MORA0.0600.1570.0700.1680.0670.1610.0690.1700.0740.1810.0620.163
PEMS0812Imp. (%)5.2943.57013.6088.6186.0074.13918.50110.92227.13517.2117.3356.417
PEMS0824Base0.0790.1820.1180.2210.0960.1940.1310.2390.1690.2870.0860.193
PEMS0824+MORA0.0730.1730.0910.1910.0860.1820.0880.1940.0960.2070.0760.179
PEMS0824Imp. (%)7.6114.86522.63513.5759.9656.11832.50118.71543.04327.90910.9987.132
PEMS0848Base0.1040.2120.186†0.235†0.1530.2440.2350.3260.2950.3980.1130.224
PEMS0848+MORA0.0910.1960.1340.2270.1210.2150.1180.2260.1290.2390.0960.203
PEMS0848Imp. (%)12.0027.33228.1313.23920.98411.76949.94130.52956.18340.02715.2259.331
PEMS0896Base0.1450.2570.221†0.267†0.2590.2960.3920.4290.3900.4790.1590.270
PEMS0896+MORA0.1150.2240.2210.2540.1670.2440.1440.2470.1540.2570.1190.228
PEMS0896Imp. (%)20.67912.789-0.1804.70835.49317.73963.13042.41960.45546.44125.04515.562

Table 15. Complete short-horizon results on PEMS at L=96L = 96.

Foundation models. Table 7 lists the frozen checkpoints. Chronos-2 uses its multivariate interface and released median forecast; TimesFM 2.5 and Sundial use their univariate interfaces, returning the released point forecast and the mean of 20 sampled trajectories, respectively. These choices are fixed within each baseline/MORA pair. The evaluation does not establish whether the benchmark datasets were excluded from the external models’ pretraining corpora.

ModelCheckpoint identifier
Chronos-2amazon/chronos-2
Sundialthuml/sundial-base-128m
TimesFM 2.5google/timesfm-2.5-200m-pytorch

Table 7. Pretrained checkpoints used at both context lengths. All foundation-model weights remain frozen.

Adapter Configuration and Fitting

Tables 8 and 9 collect the shared configuration and fixed short-horizon dimensions. The correction basis, residual estimators, and fitting procedure are specified in Appendices § 9.§ 3, § 9.§ 4, and § 9.5. The feature and periodic-template configurations are kept fixed within each dataset–backbone group across horizons, without horizon-wise test selection or per-query fallback.

SettingValue or selection rule
Long-horizon correction rankk=64k = 64
Coarse intervalsQ=min⁡{64,⌊H/4⌋}Q = \min\{64, \lfloor H/4 \rfloor\}
Coarse-predictor ridge candidates{1,10,100,1000,10000}\{1, 10, 100, 1000, 10000\}
Channel-factor dimension32, for the factorized estimator
Factor-predictor training strideOne time step
Residual-fitting OOF blocksFour; held origins in the later half of training
Residual-fitting origin strideSix time steps
Final calibrationFour coefficient groups per channel
Calibration scale range[0,1.5][0, 1.5]; fitted on validation data
Test-time parameter updatesNone

Table 8. Shared MORA configuration. PEMS-specific dimensions are given in Table 9.

Dataset-specific fitting. ECL, Traffic, and PEMS use the factorized estimator described in Appendix § 9.§ 2; other datasets use the channel-wise estimator. This dataset-level choice was fixed during development on the inspected benchmarks. For PEMS OOF replicas, checkpoint selection uses a 6% training-internal validation interval; the official validation and test splits are not used for replica fitting or checkpoint selection.

Repeated runs reuse the residual mappings fitted for the initial run. For Sundial, runs with the channel-wise estimator refit validation calibration; runs with the factorized estimator reuse the complete adapter. These repetitions vary inference or adaptation, not the pretrained foundation-model weights.

Computing environment. Independent experiment jobs run across eight NVIDIA GeForce RTX 4090 GPUs with 24 GB each. Model-specific environments follow the released implementations: VPNet uses Python 3.8.20, PyTorch 1.12.1, and CUDA 11.3; CGTFra uses Python 3.9.24, PyTorch 2.4.1, and CUDA 12.1. The foundation-model environments use PyTorch 2.7.1. Chronos-2 and TimesFM 2.5 use Transformers 4.57.6, and Sundial uses Transformers 4.40.1. The corresponding packages are chronos-forecasting 2.3.2 and timesfm 2.0.2.

Repeated Runs, Metrics, and Aggregation

The long-horizon task-specific experiments use one run per horizon on ETTh1, ETTm1, ETTm2, and Weather, and three backbone seeds on ETTh2, ECL, Traffic, and Solar. DLinear, PatchTST, TimeFilter, and VPNet use seed 2021 for the single-run datasets and seeds 2021–2023 for repeated datasets. CGTFra uses seed 2023 and seeds 2023–2025, respectively. iTransformer uses seed 2021 for single-run settings except ETTm2 at H=96H = 96, which uses seed 2023; repeated datasets use seeds 2023–2025. PEMS retains one local seed per backbone: 2021 for TimeFilter/PatchTST/DLinear/VPNet and 2023 for iTransformer/CGTFra.

Chronos-2 and TimesFM 2.5 each use one fixed pretrained checkpoint. Sundial uses sampling seed 2023 on ETTh1, ETTm1, ETTm2, and Weather, and seeds 2023–2025 on ETTh2, ECL, Traffic, and Solar. The result tables report mean errors over the recorded runs for each setting; a single-run entry retains the error from that run. Available standard deviations are retained in the result records and measure training or sampling variability, not uncertainty over independent test episodes.

For NN queries, MSE and MAE average over queries, forecast steps, and variables:

MSE⁡=1NHC∑i,t,c(yi,t,c−y^i,t,c)2,\operatorname{MSE} = \frac{1}{NHC}\sum_{i,t,c}(y_{i,t,c}-\hat{y}_{i,t,c})^2,
MAE⁡=1NHC∑i,t,c∣yi,t,c−y^i,t,c∣.(40)\operatorname{MAE} = \frac{1}{NHC}\sum_{i,t,c}\lvert y_{i,t,c}-\hat{y}_{i,t,c}\rvert. \tag*{(40)}

Local baseline–MORA pairs use identical queries, targets, and normalization. Let Eh,s(0)E_{h,s}^{(0)} and Eh,sE_{h,s} denote the baseline and adapted errors at horizon hh for run ss. We first average the ShS_h runs within a horizon, then give each horizon equal weight:

E‾(0)=1∣H∣∑h∈H1Sh∑s=1ShEh,s(0),E‾=1∣H∣∑h∈H1Sh∑s=1ShEh,s.(41)\overline{E}^{(0)} = \frac{1}{\lvert\mathcal{H}\rvert}\sum_{h\in\mathcal{H}}\frac{1}{S_h}\sum_{s=1}^{S_h}E_{h,s}^{(0)},\qquad \overline{E} = \frac{1}{\lvert\mathcal{H}\rvert}\sum_{h\in\mathcal{H}}\frac{1}{S_h}\sum_{s=1}^{S_h}E_{h,s}. \tag*{(41)}

For either metric, the reported improvement is

Imp.⁡(E)=100E‾(0)−E‾E‾(0).(42)\operatorname{Imp.}(E)=100\frac{\overline{E}^{(0)}-\overline{E}}{\overline{E}^{(0)}}. \tag*{(42)}

Imp. is computed from unrounded mean errors, not by averaging seed- or horizon-specific percentages. Dataset-level Imp. values are then averaged equally for model-level summaries, so additional runs do not increase a dataset’s weight.

Long-horizon and foundation-model tables use local paired baselines. For PEMS, Imp. uses the displayed reference score for the six quoted entries and the local baseline otherwise. Thus, the literature-reference comparison is distinct from the local paired comparison. Win counts are descriptive because overlapping windows and shared checkpoints make the evaluation settings dependent.

Variation Across Datasets and Metrics

Dataset-dependent gains. The benefit of residual adaptation depends on the dataset–backbone pair. For TimeFilter, the horizon-averaged MSE reduction is 12.910% on Solar, 1.931% on Traffic, and 0.508% on ETTh2. CGTFra on ETTh1 and VPNet on ETTh2 increase both errors, while VPNet on Solar improves MSE but not MAE. The full results retain these cases, so average gains should not be read as improvements in every setting.

Metric-dependent gains. At L=512L = 512, all 24 foundation-model–dataset pairs improve in MSE, but only 19 improve in MAE. For example, TimesFM 2.5 on Solar reduces MSE by 21.222% while increasing MAE by 4.503%. Squared loss gives larger errors greater weight than absolute loss. The two metrics therefore capture different aspects of a correction, and a reduction in MSE need not produce a reduction in MAE.

Reference-dependent gains. On PEMS, the gains for DLinear and PatchTST are measured against their local transfer baselines. Across 96 settings, local paired comparisons improve MSE and MAE in 92 and 94 cases; using the displayed reference baselines gives 90 and 94, respectively. For PEMS08/iTransformer at H=96H = 96, the adapted MSE matches the quoted 0.221 at three decimals despite a larger gain over the local baseline. These comparisons distinguish adaptation gains from margins over published reference values.

Additional Ablations

Tables 4 and 5 use TimeFilter (seed 2021) with L=96L = 96, at H=720H = 720 and H=96H = 96, respectively. The component examples are a retrospectively selected subset of the full study; ECL uses the channel-wise estimator in that comparison. Methods share the same frozen checkpoint and evaluation windows within each setting.

Individual extrapolation sources. We remove each extrapolation source in turn, retain the learned estimate, and refit the residual adapter (Table 10). The study covers eight datasets, TimeFilter, iTransformer, PatchTST, and DLinear, and H∈{96,720}H \in\{96, 720\} at L=96L = 96, using the channel-wise estimator. For either error metric EE, entries average the paired change 100(Eremoved/Efull−1)100(E_{\mathrm{removed}}/E_{\mathrm{full}} - 1) over matched settings; positive values indicate higher error after removal. Removing recent-pattern continuation or the static periodic template increases both errors on average over their matched settings.

Estimate removedSettingsMSE changeMAE change
Current level64+0.038-0.173
Recent-pattern continuation40+0.906+0.796
Static periodic template40+1.603+0.380

Table 10. Contribution of individual extrapolation sources. Entries are mean relative error changes (%) after removal and refitting. The last two rows use the same 40 template-enabled settings.

Additional comparisons. Table 11 extends the input–output δ\delta-Adapter comparison to five further settings at L=H=96L = H = 96. MORA achieves lower MSE in all five and lower MAE in three, including both metrics on ECL/TimeFilter. These results complement the three-dataset comparison in the main text.

DatasetBackboneMSEMAE
Baseδ\delta (I+O)MORABaseδ\delta (I+O)MORA
ECLPatchTST0.1800.1610.1500.2730.2520.253
TrafficPatchTST0.4590.4320.4310.2980.2780.285
ETTm2PatchTST0.1830.1820.1710.2680.2690.258
ETTh2PatchTST0.2990.2980.2850.3490.3480.337
ECLTimeFilter0.1330.1320.1290.2300.2300.228

Table 11. Additional comparisons with δ\delta-Adapter (input + output) at L=H=96L = H = 96. MORA uses the factorized estimator on ECL and Traffic and the channel-wise estimator elsewhere.

Comparison details. Affine calibration fits a channel-wise scale and offset on validation data. Direct residual regression uses the original OOF training pairs and reduced-rank ridge, matching MORA’s predictive coefficient count to within approximately 2%. The δ\delta-Adapter comparisons use the authors’ released implementations [14] with the backbone frozen and online test-time updates disabled. Neural adapters use seed 2025; input–output candidates (additive budgets 0.01 and 0.1, or a multiplicative budget of 0.2) are selected by validation MSE.

Complete Forecasting Results

Tables 12–15 report the horizon-level results for the main comparisons, with runs aggregated as described in Appendix § 10.§ 4. Bold marks the lower displayed error within each baseline–MORA pair; ties at the displayed precision are unbolded. Shaded rows give Imp. (%).

Task-Specific Forecasters

DatasetHHForecastTimeMixer MSETimeMixer MAEiTransformer MSEiTransformer MAECGTFra MSECGTFra MAEPatchTST MSEPatchTST MAEDLinear MSEDLinear MAEVPNet MSEVPNet MAE
ETTh196Base0.3690.3940.3860.4040.3730.3880.3790.3990.3960.4110.3750.387
ETTh196+MORA0.3580.3880.3820.4020.3710.3890.3650.3920.3780.4010.3680.385
ETTh196Imp. (%)2.8901.4991.0460.5290.442-0.2343.6521.7004.4862.3761.9450.647
ETTh1192Base0.4120.4210.4420.4360.4290.4220.4250.4310.4470.4420.4280.419
ETTh1192+MORA0.4030.4150.4310.4280.4240.4180.4110.4220.4240.4300.4200.415
ETTh1192Imp. (%)2.3041.4422.5651.9661.0120.8153.4182.1785.1172.8401.8430.940
ETTh1336Base0.4470.4390.4910.4610.4700.4420.4710.4580.4910.4700.4660.441
ETTh1336+MORA0.4380.4350.4840.4550.4700.4420.4520.4480.4590.4520.4620.438
ETTh1336Imp. (%)2.0220.7911.4621.1780.0980.0784.1672.2826.4833.8600.9600.535
ETTh1720Base0.4450.4560.4990.4870.4700.4630.5180.5050.5150.5130.4760.468
ETTh1720+MORA0.4410.4540.4940.4820.4850.4710.4820.4820.4940.4990.4830.471
ETTh1720Imp. (%)0.9000.2800.9841.078-3.398-1.8576.9124.5324.0522.584-1.427-0.731
ETTh296Base0.2880.3400.3010.3500.2960.3390.3020.3530.3490.4010.2900.335
ETTh296+MORA0.2830.3360.2920.3440.2880.3350.2850.3380.3000.3550.2880.334
ETTh296Imp. (%)1.7681.0243.0911.8432.5381.2695.7484.23313.99111.5260.7160.187
ETTh2192Base0.3810.4060.3800.3990.3680.3860.3830.4030.4750.4750.3580.377
ETTh2192+MORA0.3690.3930.3690.3910.3620.3810.3640.3880.4040.4210.3560.376
ETTh2192Imp. (%)3.0603.1082.8471.9501.7551.1934.9423.80315.01811.3440.5920.175
ETTh2336Base0.4080.4310.4230.4320.4180.4250.4290.4380.5990.5440.3970.409
ETTh2336+MORA0.4030.4210.4130.4240.4130.4210.4100.4240.5410.5070.4090.417
ETTh2336Imp. (%)1.3232.2242.4381.8911.1300.7774.4873.2069.5986.871-3.032-1.933
ETTh2720Base0.4100.4360.4290.4470.4220.4390.4440.4600.8250.6550.3930.419
ETTh2720+MORA0.4240.4460.4150.4350.4150.4340.4100.4380.7010.5830.4000.430
ETTh2720Imp. (%)-3.559-2.4693.2492.5581.7511.0887.6784.74115.00410.999-1.824-2.483
ETTm196Base0.3130.3540.3410.3750.3210.3490.3230.3640.3440.3730.3120.341
ETTm196+MORA0.3040.3500.3130.3570.2960.3400.3070.3530.3100.3510.2970.340
ETTm196Imp. (%)3.0561.2758.0574.8007.7052.5375.1833.11210.0515.9624.6630.352
ETTm1192Base0.3550.3800.3800.3940.3690.3730.3690.3880.3800.3910.3620.372
ETTm1192+MORA0.3380.3730.3500.3800.3460.3670.3460.3790.3430.3760.3400.367
ETTm1192Imp. (%)4.9191.6847.7913.5496.3271.6016.1902.5049.8073.9296.0131.229
ETTm1336Base0.3860.4030.4200.4190.3970.3940.3990.4090.4140.4150.3860.390
ETTm1336+MORA0.3650.3930.3790.4030.3720.3900.3690.3990.3680.3940.3630.385
ETTm1336Imp. (%)5.4172.5979.6393.8556.3181.1287.5862.50511.0714.9866.0131.117
ETTm1720Base0.4520.4370.4860.4560.4850.4420.4580.4450.4750.4530.4450.425
ETTm1720+MORA0.4220.4260.4320.4340.4280.4220.4260.4290.4270.4260.4180.418
ETTm1720Imp. (%)6.8622.62811.1144.87311.8004.6097.1277.56710.0725.8985.9681.611
ETTm296Base0.1690.2550.1850.2720.1720.2490.1830.2680.1950.2950.1730.247
ETTm296+MORA0.1650.2520.1690.2540.1650.2470.1710.2580.1750.2650.1670.244
ETTm296Imp. (%)1.8871.1268.8654.6874.0801.0566.5396.62310.28810.0813.4570.988
ETTm2192Base0.2350.2980.2530.3110.2390.2940.2470.3090.2850.3630.2360.291
ETTm2192+MORA0.2290.2960.2330.3000.2290.2900.2330.3020.2510.3230.2290.288
ETTm2192Imp. (%)2.3550.9717.9563.3904.5051.2695.7202.29011.95511.0123.0331.022
ETTm2336Base0.2930.3360.3160.3530.3040.3370.3180.3520.3790.4240.3010.333
ETTm2336+MORA0.2860.3330.2930.3410.2890.3300.2990.3450.3360.3800.2900.329
ETTm2336Imp. (%)2.2050.7837.2193.4815.0112.0205.8061.90411.22610.3453.6881.153
ETTm2720Base0.3900.3930.4060.4030.3970.3930.4260.4140.5480.5200.3870.388
ETTm2720+MORA0.3760.3880.3840.3920.3770.3850.3970.4020.4510.4490.3730.382
ETTm2720Imp. (%)3.4541.3185.5922.7574.9152.1666.8052.89617.85013.5903.4401.525
DatasetHHForecastTimeFilter MSETimeFilter MAEiTransformer MSEiTransformer MAECGTFra MSECGTFra MAEPatchTST MSEPatchTST MAEDLinear MSEDLinear MAEVPNet MSEVPNet MAE
Weather96Base0.1530.1990.1740.2130.1530.1900.1760.2170.1950.2550.1540.192
Weather96+MORA0.1480.1980.1510.2100.1440.1910.1510.2120.1510.2190.1490.191
Weather96Imp. (%)3.5050.52313.1531.6715.384-0.09113.9432.32222.54513.8793.4170.872
Weather192Base0.2020.2460.2230.2550.2050.2400.2220.2570.2400.2990.2020.240
Weather192+MORA0.1940.2430.1940.2510.1900.2380.1940.2530.2040.2720.1950.236
Weather192Imp. (%)4.1441.45012.9431.7357.6121.00312.3541.58014.8179.1693.3911.568
Weather336Base0.2590.2890.2810.2980.2600.2820.2800.2990.2840.3350.2590.281
Weather336+MORA0.2470.2850.2470.2930.2410.2750.2470.2950.2580.3130.2470.278
Weather336Imp. (%)4.8391.46212.2701.7797.6122.39411.9301.2019.3586.6154.5570.999
Weather720Base0.3450.3440.3580.3500.3420.3360.3540.3480.3480.3850.3370.336
Weather720+MORA0.3190.3350.3310.3450.3150.3260.3300.3470.3300.3740.3220.332
Weather720Imp. (%)7.7362.7087.6631.3417.8882.8156.8600.1415.2202.9834.2631.257
ECL96Base0.1350.2330.1480.2400.1390.2280.1800.2730.1950.2770.1350.224
ECL96+MORA0.1300.2290.1360.2340.1320.2260.1500.2530.1560.2580.1300.223
ECL96Imp. (%)3.6181.4098.4542.3684.6231.08316.7887.13619.6796.9463.8730.783
ECL192Base0.1580.2530.1660.2570.1550.2430.1880.2800.1940.2800.1520.240
ECL192+MORA0.1500.2470.1520.2490.1480.2400.1600.2610.1620.2620.1460.238
ECL192Imp. (%)5.5802.4118.1712.8424.4411.10714.5286.69716.3596.4974.1641.025
ECL336Base0.1660.2640.1790.2710.1700.2600.2040.2960.2070.2960.1690.258
ECL336+MORA0.1580.2590.1640.2630.1620.2560.1750.2760.1770.2780.1600.254
ECL336Imp. (%)4.5881.9338.1152.9674.8411.39414.1476.58414.5626.0615.3881.466
ECL720Base0.1960.2940.2090.2990.2040.2880.2450.3280.2420.3290.2000.287
ECL720+MORA0.1860.2880.1900.2880.1940.2860.2120.3070.2140.3120.1890.283
ECL720Imp. (%)5.2551.9739.3583.6394.6330.57913.5106.19411.6965.2015.3621.285
Traffic96Base0.3750.2510.3920.2670.3900.2400.4600.2970.6500.3970.3840.257
Traffic96+MORA0.3680.2490.3800.2620.3840.2390.4320.2850.5300.3400.3760.255
Traffic96Imp. (%)1.8901.0502.9312.0751.4960.2555.9694.23418.45314.3032.0780.980
Traffic192Base0.3960.2620.4150.2780.4220.2490.4670.3010.5990.3710.4080.267
Traffic192+MORA0.3870.2590.4010.2710.4150.2480.4390.2880.5010.3240.4010.264
Traffic192Imp. (%)2.2241.3463.3562.5171.6660.2076.0014.41016.33812.6981.8791.107
Traffic336Base0.4150.2710.4270.2840.4380.2610.4820.3070.6060.3740.4270.275
Traffic336+MORA0.4060.2680.4150.2780.4300.2610.4570.2960.5220.3360.4190.273
Traffic336Imp. (%)2.0811.1942.8971.9491.8410.1045.3023.46313.81310.2322.0460.879
Traffic720Base0.4470.2910.4570.3000.4710.2770.5160.3250.6460.3960.4650.296
Traffic720+MORA0.4400.2890.4480.2970.4650.2780.4980.3170.5940.3730.4580.294
Traffic720Imp. (%)1.5690.6531.9690.8901.398-0.1903.4944.2168.0065.6311.5870.599
Solar96Base0.1970.2320.2030.2370.1920.2070.2230.2780.2900.3770.1760.190
Solar96+MORA0.1710.2250.1710.2280.1750.2090.1740.2480.1770.2530.1700.192
Solar96Imp. (%)13.1302.87115.6313.6708.900-0.76222.03510.79638.90632.9803.163-0.814
Solar192Base0.2280.2570.2380.2630.2210.2270.2600.3030.3200.3980.1950.206
Solar192+MORA0.2010.2510.1940.2490.2040.2280.1980.2660.2000.2710.1890.207
Solar192Imp. (%)12.0562.46518.3015.4467.695-0.31423.85112.27737.28731.7723.342-0.476
Solar336Base0.2390.2690.2500.2740.2410.2420.2770.3080.3520.4150.2170.221
Solar336+MORA0.2040.2580.2070.2590.2210.2390.2170.2790.2210.2810.2110.221
Solar336Imp. (%)14.6714.22417.2915.5228.0191.46421.6169.45437.23132.3373.0180.071
Solar720Base0.2440.2740.2500.2760.2480.2450.2740.3090.3570.4130.2290.227
Solar720+MORA0.2150.2670.2200.2690.2390.2440.2320.2850.2470.3060.2270.228
Solar720Imp. (%)11.8082.57212.1842.4673.6220.44115.4717.76430.62725.8520.986-0.326

Table 12. Complete long-horizon results for six task-specific backbones at L=96L = 96.

Foundation Models at L=96L = 96

DatasetHHForecastChronos-2 MSEChronos-2 MAESundial MSESundial MAETimesFM 2.5 MSETimesFM 2.5 MAE
ETTh196Base0.4240.3970.4060.3960.4290.402
ETTh196+MORA0.3940.3930.3760.3920.3790.394
ETTh196Imp. (%)6.9671.1117.2401.17611.6622.190
ETTh1192Base0.4840.4310.4690.4300.4850.432
ETTh1192+MORA0.4410.4190.4230.4190.4240.420
ETTh1192Imp. (%)8.8802.7959.8542.63612.6822.680
ETTh1336Base0.5360.4560.5160.4510.5260.450
ETTh1336+MORA0.4920.4440.4790.4420.4660.441
ETTh1336Imp. (%)8.2272.6077.0681.88411.3582.106
ETTh1720Base0.5260.4710.5000.4610.4980.457
ETTh1720+MORA0.4970.4670.4660.4570.4720.462
ETTh1720Imp. (%)5.4310.9886.8520.7585.130-1.050
ETTh296Base0.3170.3440.3170.3530.3300.345
ETTh296+MORA0.2950.3310.2880.3370.2980.335
ETTh296Imp. (%)7.0213.9779.0784.6229.7582.901
ETTh2192Base0.4130.4000.4050.4050.4200.402
ETTh2192+MORA0.3790.3820.3750.3880.3960.391
ETTh2192Imp. (%)8.1744.4177.5214.0315.8062.620
ETTh2336Base0.4490.4330.4380.4360.4530.434
ETTh2336+MORA0.4190.4170.4180.4240.4340.425
ETTh2336Imp. (%)6.6013.8124.4792.6204.2852.192
ETTh2720Base0.4430.4420.4370.4450.4500.443
ETTh2720+MORA0.4410.4410.4740.4720.4620.465
ETTh2720Imp. (%)0.4870.303-8.556-5.959-2.575-4.912
ETTm196Base0.5920.4520.4780.4200.8410.560
ETTm196+MORA0.3120.3560.3170.3560.3110.354
ETTm196Imp. (%)47.22121.32933.78515.21963.05136.912
ETTm1192Base0.6070.4670.5270.4490.8810.582
ETTm1192+MORA0.3520.3790.3530.3790.3580.380
ETTm1192Imp. (%)42.09218.86432.87115.73159.36334.684
ETTm1336Base0.6180.4820.5480.4670.8720.588
ETTm1336+MORA0.3760.3990.3780.3980.3810.397
ETTm1336Imp. (%)39.08617.13331.10314.68156.32832.427
ETTm1720Base0.6460.5080.5960.4940.9050.607
ETTm1720+MORA0.4250.4290.4280.4290.4310.429
ETTm1720Imp. (%)34.18815.54428.20513.08652.35729.275
ETTm296Base0.2100.2730.2010.2760.2570.312
ETTm296+MORA0.1740.2500.1690.2540.1740.255
ETTm296Imp. (%)16.9658.32115.6378.13432.14618.140
ETTm2192Base0.2740.3140.2660.3180.3330.357
ETTm2192+MORA0.2360.2910.2330.2990.2380.299
ETTm2192Imp. (%)13.7577.33912.4346.18328.51016.374
ETTm2336Base0.3400.3550.3260.3550.3950.393
ETTm2336+MORA0.2980.3320.2930.3370.2980.338
ETTm2336Imp. (%)12.3246.50810.2265.10324.50214.135
ETTm2720Base0.4480.4120.4240.4090.5010.448
ETTm2720+MORA0.3930.3860.3890.3900.3970.394
ETTm2720Imp. (%)12.2626.1118.2634.72320.87812.128

Table 13. Complete long-horizon results for frozen foundation models at L=96L = 96.

DatasetHHForecastChronos-2 MSEChronos-2 MAESundial MSESundial MAETimesFM 2.5 MSETimesFM 2.5 MAE
Weather96Base0.2440.2300.1950.2350.2070.229
Weather96+MORA0.1710.2240.1510.2180.1520.210
Weather96Imp. (%)29.8722.52922.5427.02926.7178.175
Weather192Base0.2730.2700.2540.2830.2640.277
Weather192+MORA0.2060.2670.1960.2660.1940.258
Weather192Imp. (%)24.5011.25822.6746.14026.3766.830
Weather336Base0.3190.3050.3090.3220.3360.318
Weather336+MORA0.2560.3060.2510.3090.2480.302
Weather336Imp. (%)19.551-0.13418.6853.81126.1094.913
Weather720Base0.3920.3570.3820.3690.4500.374
Weather720+MORA0.3390.3620.3310.3700.3270.360
Weather720Imp. (%)13.635-1.39113.228-0.04627.3353.647
ECL96Base0.2140.2740.1800.2610.2230.293
ECL96+MORA0.1590.2530.1490.2470.1640.264
ECL96Imp. (%)25.8867.62217.1975.21026.4599.869
ECL192Base0.2130.2790.1870.2690.2200.292
ECL192+MORA0.1670.2600.1590.2560.1710.268
ECL192Imp. (%)21.6106.87914.6504.84622.5928.431
ECL336Base0.2300.2950.2040.2860.2360.307
ECL336+MORA0.1840.2750.1740.2720.1860.283
ECL336Imp. (%)20.1556.71914.3344.92020.9697.820
ECL720Base0.2790.3290.2450.3190.2780.341
ECL720+MORA0.2240.3070.2110.3040.2260.317
ECL720Imp. (%)19.4586.41213.9084.92618.8307.000
Traffic96Base0.5740.3100.6050.3530.6020.338
Traffic96+MORA0.4970.2950.5130.3170.5300.311
Traffic96Imp. (%)13.2805.01815.18910.14811.9708.010
Traffic192Base0.5500.3010.5820.3440.5750.324
Traffic192+MORA0.4910.2890.5020.3130.5150.303
Traffic192Imp. (%)10.7814.18613.6529.22310.3286.483
Traffic336Base0.5660.3090.5960.3510.5920.329
Traffic336+MORA0.5130.2980.5240.3240.5380.313
Traffic336Imp. (%)9.3133.51912.1737.7499.0235.011
Traffic720Base0.6060.3310.6350.3720.6280.346
Traffic720+MORA0.5690.3240.5850.3540.5940.337
Traffic720Imp. (%)6.0572.1427.8324.8015.3722.600
Solar96Base0.6400.3430.6250.5020.8750.503
Solar96+MORA0.1860.2650.1770.2540.1840.254
Solar96Imp. (%)70.98722.87871.72249.45078.94149.516
Solar192Base0.5270.3240.6800.5340.9370.528
Solar192+MORA0.2080.2850.2010.2720.2110.274
Solar192Imp. (%)60.56912.31870.49549.01477.47348.213
Solar336Base0.4940.3230.6840.5370.9090.520
Solar336+MORA0.2210.2890.2220.2850.2340.287
Solar336Imp. (%)55.34410.34267.60946.91674.25844.733
Solar720Base0.4510.3190.6760.5450.9300.528
Solar720+MORA0.2460.3070.2430.3010.2580.303
Solar720Imp. (%)45.4933.78664.10944.76372.30142.712

Table 13 (continued). , part 2 of 2). Complete long-horizon results for frozen foundation models at L=96L = 96.

Foundation Models at L=512L = 512

DatasetHHForecastChronos-2 MSEChronos-2 MAESundial MSESundial MAETimesFM 2.5 MSETimesFM 2.5 MAE
ETTh196Base0.3710.3750.3560.3830.3840.382
ETTh196+MORA0.3500.3760.3430.3780.3540.380
ETTh196Imp. (%)5.652-0.2873.5931.3187.7810.545
ETTh1192Base0.4140.4060.3900.4060.4280.408
ETTh1192+MORA0.3920.4060.3780.4010.4010.409
ETTh1192Imp. (%)5.260-0.0933.0251.1066.117-0.190
ETTh1336Base0.4390.4230.4030.4150.4460.419
ETTh1336+MORA0.4210.4270.3960.4140.4270.424
ETTh1336Imp. (%)4.179-0.9061.8570.2664.340-1.193
ETTh1720Base0.4490.4450.4380.4440.4280.427
ETTh1720+MORA0.4450.4520.4310.4430.4280.437
ETTh1720Imp. (%)0.925-1.6021.6030.134-0.148-2.455
ETTh296Base0.2800.3200.2790.3340.3060.338
ETTh296+MORA0.2730.3160.2710.3250.2950.332
ETTh296Imp. (%)2.5351.0682.9382.5503.7301.796
ETTh2192Base0.3510.3690.3420.3760.3760.384
ETTh2192+MORA0.3420.3640.3380.3710.3650.379
ETTh2192Imp. (%)2.7171.3161.2111.4682.8581.444
ETTh2336Base0.3770.3970.3610.3950.4010.409
ETTh2336+MORA0.3710.3950.3670.3970.3940.405
ETTh2336Imp. (%)1.7260.402-1.778-0.4821.7130.875
ETTh2720Base0.3980.4240.3920.4260.4050.422
ETTh2720+MORA0.3930.4340.3840.4200.4060.418
ETTh2720Imp. (%)1.362-2.3101.9971.358-0.1950.756
ETTm196Base0.3080.3230.2920.3250.3130.329
ETTm196+MORA0.2770.3150.2750.3230.2840.326
ETTm196Imp. (%)10.0242.4635.9200.6769.1691.008
ETTm1192Base0.3620.3550.3420.3570.3610.359
ETTm1192+MORA0.3180.3450.3180.3530.3210.353
ETTm1192Imp. (%)12.1702.8737.0181.17610.8971.813
ETTm1336Base0.3950.3810.3780.3800.3940.382
ETTm1336+MORA0.3450.3700.3500.3750.3510.374
ETTm1336Imp. (%)12.8032.7347.5571.43410.8202.128
ETTm1720Base0.4840.4300.4510.4180.4560.418
ETTm1720+MORA0.4110.4120.4110.4090.4090.411
ETTm1720Imp. (%)14.9974.3068.9102.08610.3371.501
ETTm296Base0.1700.2370.1760.2550.1790.246
ETTm296+MORA0.1610.2330.1580.2440.1680.241
ETTm296Imp. (%)5.5041.84110.2004.2446.4482.062
ETTm2192Base0.2410.2860.2390.2980.2490.293
ETTm2192+MORA0.2240.2770.2160.2850.2310.285
ETTm2192Imp. (%)6.8352.8239.7764.4027.0062.958
ETTm2336Base0.3060.3290.2970.3360.3140.336
ETTm2336+MORA0.2770.3150.2720.3230.2940.326
ETTm2336Imp. (%)9.3274.4038.2943.6236.4552.956
ETTm2720Base0.4040.3900.3870.3930.4200.396
ETTm2720+MORA0.3660.3730.3640.3810.3990.387
ETTm2720Imp. (%)9.2494.4786.1383.0565.0892.445

Table 14. Complete long-horizon results for frozen foundation models at L=512L = 512.

DatasetHHForecastChronos-2 MSEChronos-2 MAESundial MSESundial MAETimesFM 2.5 MSETimesFM 2.5 MAE
Weather96Base0.1560.1760.1600.2070.1630.194
Weather96+MORA0.1390.1780.1390.1950.1420.186
Weather96Imp. (%)10.563-1.28113.1875.58812.7164.399
Weather192Base0.2110.2260.2130.2550.2160.242
Weather192+MORA0.1810.2250.1820.2400.1840.227
Weather192Imp. (%)14.2080.50514.8736.00214.8316.405
Weather336Base0.2810.2750.2670.2950.2890.289
Weather336+MORA0.2290.2690.2310.2800.2320.267
Weather336Imp. (%)18.6992.06913.5805.08219.7447.461
Weather720Base0.3830.3320.3360.3440.3850.345
Weather720+MORA0.3020.3290.3000.3320.3020.318
Weather720Imp. (%)21.1680.93710.9453.35921.6317.811
ECL96Base0.1280.2110.1180.2130.1330.220
ECL96+MORA0.1230.2100.1160.2120.1270.218
ECL96Imp. (%)3.4830.3941.6570.5884.3200.810
ECL192Base0.1460.2290.1350.2310.1500.236
ECL192+MORA0.1410.2280.1320.2290.1430.234
ECL192Imp. (%)3.7320.5382.1260.7544.4801.014
ECL336Base0.1670.2490.1500.2470.1670.253
ECL336+MORA0.1580.2470.1460.2450.1590.251
ECL336Imp. (%)5.0130.9732.8650.9795.0961.169
ECL720Base0.2100.2870.1880.2800.2060.287
ECL720+MORA0.1960.2810.1800.2760.1950.283
ECL720Imp. (%)6.8691.9124.2571.4615.6961.545
Traffic96Base0.3530.2190.4250.2700.3550.231
Traffic96+MORA0.3490.2190.4080.2650.3520.230
Traffic96Imp. (%)0.965-0.2123.9201.9230.8730.149
Traffic192Base0.3790.2310.4480.2830.3710.238
Traffic192+MORA0.3750.2310.4280.2760.3680.238
Traffic192Imp. (%)1.106-0.1064.3412.4710.9030.155
Traffic336Base0.3960.2400.4620.2920.3820.244
Traffic336+MORA0.3920.2400.4440.2850.3790.244
Traffic336Imp. (%)1.006-0.1583.9442.3500.730-0.022
Traffic720Base0.4420.2640.4990.3140.4180.264
Traffic720+MORA0.4400.2650.4850.3080.4160.264
Traffic720Imp. (%)0.384-0.2052.8371.7990.5190.052
Solar96Base0.1570.1720.2120.2340.2360.221
Solar96+MORA0.1570.1720.1900.2340.1860.231
Solar96Imp. (%)0.011-0.03210.075-0.26620.996-4.608
Solar192Base0.1800.1900.2290.2530.2700.242
Solar192+MORA0.1800.1900.2090.2530.2170.253
Solar192Imp. (%)0.069-0.1198.832-0.02419.555-4.666
Solar336Base0.1970.2020.2340.2640.3030.262
Solar336+MORA0.1970.2020.2140.2640.2410.274
Solar336Imp. (%)0.148-0.2148.5620.16120.378-4.620
Solar720Base0.2140.2100.2430.2770.3170.274
Solar720+MORA0.2130.2110.2200.2750.2420.285
Solar720Imp. (%)0.405-0.3849.4430.72223.618-4.162

Table 14 (continued). , part 2 of 2). Complete long-horizon results for frozen foundation models at L=512L = 512.

Short-Horizon Forecasting on PEMS

Additional Residual Analysis

We extend Figure 1 to the eight long-term datasets using frozen DLinear, iTransformer, and TimeFilter forecasts at L=H=96L = H = 96. The analysis retains the original chronological splits and training-set normalization. DLinear and TimeFilter use seed 2021. iTransformer uses seed 2021 on ETTh1, ETTm1, and Weather, and seed 2023 on the other five datasets. The linear probes are used only for this analysis; MORA is fitted as described in Appendix § 9.5.

Frequency Decomposition of Forecast Residuals

For test window ii and channel cc, let ri,c=yi,c−y^i,c(0)\mathbf{r}_{i,c} = \mathbf{y}_{i,c} - \widehat{\mathbf{y}}_{i,c}^{(0)} be the backbone residual. Let DH\mathbf{D}_{H} denote the orthonormal DCT-II matrix over the forecast horizon. With NN test windows, the contribution of frequency index kk to the overall forecasting MSE is

Ek=1NHC∑i=1N∑c=1C[(DHri,c)k]2,k=0,…,H−1.(43)E_k = \frac{1}{NHC} \sum_{i=1}^{N} \sum_{c=1}^{C} \left[(\mathbf{D}_{H}\mathbf{r}_{i,c})_k\right]^2,\qquad k = 0,\ldots,H-1. \tag*{(43)}

Orthonormality gives

MSE=∑k=0H−1Ek,F8=∑k=07Ek∑k=0H−1Ek.(44)\mathrm{MSE} = \sum_{k=0}^{H-1} E_k,\qquad F_8 = \frac{\sum_{k=0}^{7} E_k}{\sum_{k=0}^{H-1} E_k}. \tag*{(44)}

Thus, F8F_8 measures the share of residual energy in the first eight modes, including the constant component. Display ranks 1–8 in Figure 1(a) correspond to DCT indices 0–7. The reported fractions use the full residual spectrum and all test windows and channels; they are ratios of summed energies, not averages of per-window energy fractions.

Table 16 reports F8F_8 alongside the history-based probe statistic defined below. On Weather, the energy shares are 84.0%, 81.8%, and 79.5% for DLinear, iTransformer, and TimeFilter, respectively, matching Figure 1(a). The first eight modes also account for substantial error on the other datasets, although the fraction varies: it exceeds 92% on Solar but is around 29%–33% on Traffic. The diagnostic band is fixed and is distinct from MORA’s tapered correction subspace.

DatasetDLineariTransformerTimeFilter
F8F_8 (%)rmedr_{\mathrm{med}}F8F_8 (%)rmedr_{\mathrm{med}}F8F_8 (%)rmedr_{\mathrm{med}}
ETTh148.1950.11550.0600.26450.5580.219
ETTh281.7340.16177.1320.06878.7490.277
ETTm188.0410.00886.8250.12887.1960.122
ETTm284.8580.20682.7510.31882.2260.210
Weather83.9840.47181.8470.30279.5370.206
ECL61.5410.17657.4380.22855.5670.225
Traffic32.9930.39329.3190.28529.5980.240
Solar94.5300.05792.7860.15892.4900.101

Table 16. Residual diagnostics at L=H=96L = H = 96. F8F_8 is the DCT-8 share of squared residual energy; rmedr_{\mathrm{med}} is the median channel-wise test correlation of the history-based mean-residual probe.

History-Based Forecast-Mean Probes

Following Appendix § 8.§ 2, the probe targets the signed future-mean residual:

rˉi,c=1H1Tri,c,hi,c=[(DLxi,c)0,…,(DLxi,c)7]T.(45)\bar{r}_{i,c} = \frac{1}{H}\mathbf{1}^{\mathsf{T}}\mathbf{r}_{i,c},\qquad\mathbf{h}_{i,c} = \left[(\mathbf{D}_{L}\mathbf{x}_{i,c})_0,\ldots,(\mathbf{D}_{L}\mathbf{x}_{i,c})_7\right]^{\mathsf{T}}. \tag*{(45)}

Here hi,c\mathbf{h}_{i,c} contains eight coefficients of the complete same-channel historical window, not eight selected time points. A positive rˉi,c\bar{r}_{i,c} indicates that the backbone underpredicts the future mean; a negative value indicates overprediction.

For each backbone and channel, we fit a linear ridge probe with an intercept on that backbone’s validation windows. Let h~i,c\widetilde{\mathbf{h}}_{i,c} denote the features standardized by their validation means and standard deviations, with a scale floor of 10−810^{-8}. The probe predicts pi,c=ac+wcTh~i,cp_{i,c} = a_c + \mathbf{w}_c^{\mathsf{T}}\widetilde{\mathbf{h}}_{i,c} by minimizing

1Nval∑i∈Ival(rˉi,c−pi,c)2+0.01∥wc∥2,(46)\frac{1}{N_{\mathrm{val}}}\sum_{i\in\mathcal{I}_{\mathrm{val}}}\left(\bar{r}_{i,c} - p_{i,c}\right)^2 + 0.01\lVert\mathbf{w}_c\rVert^2, \tag*{(46)}

where Ival\mathcal{I}_{\mathrm{val}} contains the NvalN_{\mathrm{val}} validation windows. Predictions remain in the benchmark’s standardized units. The probe uses neither future observations nor calendar covariates; its weights, normalization statistics, and group thresholds remain fixed on test data. Test targets are used only to evaluate the associations.

For Figure 1(b), each channel’s validation probe scores define five groups through their 20th, 40th, 60th, and 80th percentiles. We apply these same thresholds to test scores computed from history. Within each group, we first average the estimated and observed mean errors over test windows separately for each channel, then average channels equally. To keep the channel composition the same across the five groups, this visualization uses channels with at least one test window in every group. The resulting coverage on Weather is 19 of 21 channels for DLinear and iTransformer and 18 of 21 for TimeFilter. Coverage depends on the history-based group assignments, not on the observed errors. The plotted points are group means, not individual forecasts or a temporal trajectory.

For the cross-dataset summary, we compute the Pearson correlation between the frozen probe output and rˉi,c\bar{r}_{i,c} over test windows within each channel, and report its median over channels with a defined correlation. This uses all such channels, independently of the five-group visualization’s coverage criterion. Correlations are defined for 320 of 321 channels on ECL and for all channels in the other settings reported here.

Interpretation. Spectral concentration and predictability answer different questions. DLinear has substantial low-frequency residual energy on both Weather and ETTm1, yet the corresponding probe correlations are 0.471 and 0.008. Low-frequency energy alone therefore does not establish that an error can be inferred from history. The probe provides separate evidence of history-dependent variation in the future-mean residual, without implying that every low-frequency component is predictable or that all backbones share a common training failure. Since test windows overlap, we report correlations and group means as descriptive statistics, not independent-sample significance tests.

References

  1. [1]N. Ahmed, T. Natarajan, and K. R. Rao. Discrete cosine transform. IEEE Transactions on Computers, C-23(1):90–93, 1974. doi: 10.1109/T-C.1974.223784. URL https://doi.org/10.1109/T-C.1974.223784.
  2. [2]Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Schhur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, Jasper Zschiegner, Danielle C. Maddix, Hao Wang, Michael W. Mahoney, Kari Torkkola, Andrew Gordon Wilson, Michael Bohlke-Schneider, and Yuyang Wang. Chronos: The language of time series. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URL https://openreview.net/forum?id=gerNCVqqtR.arxiv.org/abs/2403.07815
  3. [3]Abdul Fatir Ansari, Oleksandr Schhur, Jaris Kük̈en, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider. Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.15821, 2025. URL https://arxiv.org/abs/2510.15821.
  4. [4]Shaojie Bai, J. Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018. URL https://arxiv.org/abs/1803.01271.
  5. [5]Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp. 10148–10167, 2024. URL https://proceedings.mlr.press/v235/das24c.html.arxiv.org/abs/2310.10688
  6. [6]Wanjin Feng, Yuan Yuan, Jingtao Ding, and Yong Li. Beyond model ranking: Predictability-aligned evaluation for time series forecasting. In International Conference on Machine Learning, 2026. URL https://arxiv.org/abs/2509.23074.
  7. [7]Google Research. TimesFM 2.5: 200m pytorch checkpoint. Official model release and model card, 2025. URL https://huggingface.co/google/timesfm-2.5-200m-pytorch. Checkpoint identifier: google/timesfm-2.5-200m-pytorch.
  8. [8]Arthur E. Hoerl and Robert W. Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12(1):55–67, 1970. doi: 10.1080/00401706.1970.10488634. URL https://doi.org/10.1080/00401706.1970.10488634.
  9. [9]Yifan Hu, Guibin Zhang, Peiyuan Liu, Disen Lan, Naiqi Li, Dawei Cheng, Tao Dai, Shu-Tao Xia, and Shirui Pan. TimeFilter: Patch-specific spatial-temporal graph filtration for time series forecasting. In Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, pp. 24893–24911, 2025. URL https://proceedings.mlr.press/v267/hu25ac.html.arxiv.org/abs/2501.13041
  10. [10]Jeonghwan Im and Hyuk-Yoon Kwon. COSA: Context-aware output-space adapter for test-time adaptation in time series forecasting. In International Conference on Learning Representations, 2026. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/2a8ce71baac4c89bf9ff479d8240c7d9-Abstract-Conference.html.
  11. [11]HyunGi Kim, Siwon Kim, Jisoo Mok, and Sungroh Yoon. Battling the non-stationarity in time series forecasting via test-time adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp. 17868–17876, 2025. doi: 10.1609/aaai.v39i17.33965. URL https://ojs.aaai.org/index.php/AAAI/article/view/33965.DOI
  12. [12]Taesung Kim, Jinhee Kim, Yunwon Tae, Cheonbok Park, Jang-Ho Choi, and Jaegul Choo. Reversible instance normalization for accurate time-series forecasting against distribution shift. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=cGDAkQo1COp.
  13. [13]Guokun Lai, Wei-Cheng Chang, Yiming Yang, and Hanxiao Liu. Modeling long- and short-term temporal patterns with deep neural networks. In Proceedings of the 41st International ACM SIGIR Conference on Research and Development in Information Retrieval, 2018. URL https://arxiv.org/abs/1703.07015.
  14. [14]Daojun Liang, Qi Li, Yinglong Wang, Jing Chen, Hu Zhang, Xiaoxiao Cui, Qizheng Wang, and Shuo Li. The forecast after the forecast: A post-processing shift in time series. In The Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum?id=syfWdclGE1.arxiv.org/abs/2601.20280
  15. [15]Kun Liu, Renjun Jia, Ruifeng Yang, Xirui Zeng, Yuqi Liang, and Cen Chen. Are global dependencies necessary? scalable time series forecasting via local cross-variate modeling. In International Conference on Learning Representations, 2026. URL https://proceedings.iclr.cc/paper_files/paper/2026/hash/a832b65da26c4867a8581cb13d46dd-Abstract-Conference.html.
  16. [16]Yong Liu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Non-stationary Transformers: Exploring the stationarity in time series forecasting. In Advances in Neural Information Processing Systems, volume 35, 2022. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/4054556fcaa934b0bf76da52cf4f92cb-Abstract-Conference.html.arxiv.org/abs/2205.14415
  17. [17]Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. iTransformer: Inverted transformers are effective for time series forecasting. In International Conference on Learning Representations, 2024. URL https://arxiv.org/abs/2310.06625.
  18. [18]Yong Liu, Guo Qin, Zhiyuan Shi, Zhi Chen, Caiyin Yang, Xiangdong Huang, Jianmin Wang, and Mingsheng Long. Sundial: A family of highly capable time series foundation models. In International Conference on Machine Learning, 2025. URL https://arxiv.org/abs/2502.00816.
  19. [20]Reza Nematirad, Anil Pahwa, and Balasubramaniam Natarajan. Are data embeddings effective in time series forecasting? Transactions on Machine Learning Research, 2025. URL https://openreview.net/forum?id=yeu44ZRvZZ.arxiv.org/abs/2505.20716
  20. [21]Yuqi Nie, Nam H. Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. In International Conference on Learning Representations, 2023. URL https://arxiv.org/abs/2211.14730.
  21. [22]Chao Song, Youfang Lin, Shengnan Guo, and Huaiyu Wan. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pp. 914–921, 2020. doi: 10.1609/aaai.v34i01.5438. URL https://ojs.aaai.org/index.php/AAAI/article/view/5438.DOI
  22. [23]Gerald Woo, Chenghao Liu, Akshat Kumar, Caiming Xiong, Silvio Savarese, and Doyen Sahoo. Unified training of universal time series forecasting transformers. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pp. 53140–53164, 2024. URL https://proceedings.mlr.press/v235/woo24a.arxiv.org/abs/2402.02592
  23. [24]Haixu Wu, Jiehui Xu, Jianmin Wang, and Mingsheng Long. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. In Advances in Neural Information Processing Systems, volume 34, 2021. URL https://proceedings.neurips.cc/paper/2021/hash/bccd040288793e8bdcd7c19a8ac0c2b-Abstract.html.arxiv.org/abs/2106.13008
  24. [25]Zhijian Xu, Ailing Zeng, and Qiang Xu. FITS: Modeling time series with 10k parameters. In International Conference on Learning Representations, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/hash/701251e1db4a2e4dd2ef23f5265d5936-Abstract-Conference.html.arxiv.org/abs/2307.03756
  25. [26]Kun Yi, Qi Zhang, Wei Fan, Shoujin Wang, Pengyang Wang, Hui He, Ning An, Defu Lian, Longbing Cao, and Zhendong Niu. Frequency-domain MLPs are more effective learners in time series forecasting. In Advances in Neural Information Processing Systems, volume 36, 2023. URL https://proceedings.neurips.cc/paper/2023/hash/f1d16af76939f476b5f040fd1398c0a3-Abstract.html.
  26. [27]Wuqing Yu, Weichen Guo, Jian Zhou, Shuyi Luo, and Jiacai Zhang. Robust inter-series dependency modeling for time series forecasting via information-theoretic alignment. In International Conference on Machine Learning, 2026. URL https://openreview.net/forum?id=YQPQuHVLLQ.
  27. [28]Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp. 11121–11128, 2023. URL https://arxiv.org/abs/2205.13504.
  28. [29]Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pp. 11106–11115, 2021. URL https://arxiv.org/abs/2012.07436.
  29. [30]Tian Zhou, Ziqing Ma, Xue Wang, Qingsong Wen, Liang Sun, Tao Yao, Wotao Yin, and Rong Jin. FiLM: Frequency improved Legendre memory model for long-term time series forecasting. In Advances in Neural Information Processing Systems, volume 35, 2022a. URL https://proceedings.neurips.cc/paper_files/paper/2022/hash/524ef58c2bd075775861234266e5e020-Abstract.html.arxiv.org/abs/2205.08897

Paper details

vlnc:W_wxbuvaup
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

generative AI tools

View statement
Looking for related papers…
Contents