Shape-Aware Temporal Loss (SATL)

• SATL is a family of loss functions that integrate shape and temporal distortion measures to overcome the limitations of point-wise losses in forecasting.
• The DILATE formulation of SATL employs a smooth dynamic programming approach to combine shape matching and temporal alignment for improved spike detection and timing.
• Variants of SATL extend the concept by incorporating first-order differences, frequency-domain analysis, and perceptual feature losses to capture geometric cues in time-series data.

Shape-Aware Temporal Loss (SATL) denotes a class of objective functions for deep time-series forecasting that are designed to preserve salient temporal structure rather than relying only on point-wise discrepancies such as MSE or MAE. In the supplied literature, the term has two closely related but non-identical uses. One use refers to the DILATE objective, which combines a smooth Dynamic Time Warping (DTW)-based shape term with a differentiable temporal-distortion term for multi-step forecasting of non-stationary signals with sudden changes (Guen et al., 2019, Guen et al., 2021). A later use refers to a three-component loss that approximates the geometric cues of the Time Series Geometric Structure Index (TGSI) by combining first-order difference, frequency-domain, and perceptual feature losses in the time-series modality (Yu et al., 31 Jul 2025). Across these formulations, the common premise is that purely point-wise losses are often insufficient for sharp forecasts, timely change localization, and faithful preservation of rises, drops, peaks, plateaus, and periodic cycles.

1. Terminological scope and research trajectory

The earliest formulation in the supplied material is DILATE, introduced for training deep neural networks on non-stationary time-series forecasting tasks with multiple future steps (Guen et al., 2019). DILATE explicitly combines two terms: one for shape distortion and one for temporal distortion. The later paper "Deep Time Series Forecasting with Shape and Temporal Criteria" extends this perspective by defining shape and temporal similarities and dissimilarities via smooth relaxations of DTW and the Temporal Distortion Index (TDI), enabling differentiable loss functions and positive semi-definite kernels; in that paper, DILATE is the deterministic forecasting objective, while STRIPE++ is introduced for probabilistic forecasting with structured shape and time diversity enforced through a determinantal point process diversity loss (Guen et al., 2021).

A separate 2025 line of work reintroduces the name SATL for a different construction motivated by the non-differentiability of an image-based metric, TGSI. There, SATL is defined as a multi-component loss operating directly on the time-series modality and is intended to bridge a geometric-structure evaluation criterion into end-to-end training (Yu et al., 31 Jul 2025).

This suggests that SATL is not a single canonical formula across the literature. Rather, the label is attached to a family of objectives that share a common ambition: to couple numerical fidelity with explicit modeling of structural or temporal characteristics that are poorly captured by plain $L_p$-type losses.

2. DILATE as a shape-and-time SATL

In the DILATE formulation, the forecast $\hat y$ and target $y$ are compared through a pairwise cost matrix $\Delta$ and a set of admissible warping paths $\mathcal A$. The shape term is a smooth relaxation of DTW:

$$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$

with smoothing parameter $\gamma>0$. The supplied descriptions state that $\Delta$ may be instantiated from squared-Euclidean pointwise costs such as $\| \hat y_i-y_j\|_2^2$, and that as $\gamma\to 0$, $\hat y$0 (Guen et al., 2021).

The temporal term is a differentiable analogue of TDI. Let $\hat y$1 denote the soft alignment induced by the Gibbs distribution over warping paths. Then the temporal distortion loss is

$\hat y$2

where $\hat y$3 is a nonnegative off-diagonal penalty, with examples including $\hat y$4 and $\hat y$5 (Guen et al., 2021).

The combined objective is

$\hat y$6

or equivalently,

$\hat y$7

with $\hat y$8. The supplied descriptions state that $\hat y$9 recovers pure soft-DTW, while $y$0 uses only the TDI term and is not generally recommended alone (Guen et al., 2019, Guen et al., 2021).

Within this construction, the shape term encourages global shape matching under elastic alignment, and the temporal term penalizes delay or advance of aligned features. The stated motivation is that MSE cannot distinguish a shifted spike from a correctly timed spike, whereas the combined objective encourages forecasts that are both sharp and timely (Guen et al., 2021).

3. Differentiability, dynamic programming, and hyperparameters

Both DILATE components are differentiable with respect to the prediction because they are built from smooth operations, notably log-sum-exp and matrix multiplications. For the shape term, the gradient can be written as a sum over soft-alignment weights multiplied by derivatives of the pointwise cost matrix. For squared-Euclidean costs, the supplied exposition gives $y$1; for the temporal term, differentiation proceeds through $y$2, which follows from the Hessian of soft-DTW and can be implemented by back-propagation through the dynamic-programming routine (Guen et al., 2021).

The forward and backward passes are computed by dynamic programming in $y$3 time and memory for length-$y$4 sequences, or $y$5 for general lengths $y$6. The forward recurrence uses a numerically stabilized log-sum-exp form, and the backward pass yields the soft alignment $y$7. One supplied implementation strategy is to write a custom autograd.Function that caches the forward matrix and implements both $y$8 and the Hessian-vector products needed for the temporal term; the supplied material reports that this custom backward is approximately $y$9 faster than naive autograd for $\Delta$0 (Guen et al., 2019).

The principal hyperparameters are the smoothing temperature $\Delta$1, the shape-time tradeoff $\Delta$2, and the choice of temporal penalty matrix $\Delta$3. The supplied descriptions give $\Delta$4 as a typical or suggested default, note that smaller $\Delta$5 is closer to exact DTW but may make training harder or produce vanishing and exploding gradients, and identify $\Delta$6 as a typical grid, with $\Delta$7 a good starting point for non-stationary data with sharp steps. Reported task-specific settings include $\Delta$8 for synthetic data and ECG5000, and $\Delta$9 for traffic occupancy (Guen et al., 2019, Guen et al., 2021).

These implementation details matter because the appeal of DILATE-style SATL is inseparable from its trainability. The formulation is more expressive than point-wise losses, but it also introduces dynamic-programming cost, temperature sensitivity, and endpoint degeneracies when $\mathcal A$0 approaches $\mathcal A$1 or $\mathcal A$2.

4. TGSI-bridging SATL in the time-series modality

The 2025 SATL is motivated by TGSI, an evaluation metric that transforms a time series into an image and measures geometric alignment through luminance and covariance terms. TGSI is defined after a non-differentiable image conversion in which each time step lights up a vertical column of pixels around the normalized value with Gaussian-like falloff over $\mathcal A$3 pixels; because this mapping discretizes and expands the series, the supplied material states that $\mathcal A$4, so TGSI cannot be used directly as a training loss (Yu et al., 31 Jul 2025).

To bridge this gap, the 2025 SATL operates entirely in the time-series modality and combines three terms. The first-order difference loss is

$\mathcal A$5

where $\mathcal A$6. This term measures structural consistency through local slopes (Yu et al., 31 Jul 2025).

The frequency-domain term uses the discrete Fourier transform along the time axis. Let $\mathcal A$7 denote the dominant frequencies, defined as the top-$\mathcal A$8 magnitudes of $\mathcal A$9, and let the complement act as a noise band. The loss is decomposed into a dominant-frequency matching term and a non-dominant suppression term:

$$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$0

and then normalized as

$$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$1

The supplied explanation states that $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$2 reproduces major periodic components and $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$3 suppresses spurious oscillations (Yu et al., 31 Jul 2025).

The third component is a perceptual feature loss. It is constructed in two stages. First, a time-series image autoencoder is trained after converting each $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$4 series into an image $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$5 by a vertical-expansion procedure with height $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$6, width $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$7, and expansion half-width $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$8. The encoder is a 4-layer convolutional network and the decoder is FC plus transposed-CNN, trained by $$L_{\mathrm{shape}}(\hat y,y) = DTW_\gamma(\hat y,y) = -\gamma \log \sum_{A\in\mathcal A_{k,k}} \exp\!\bigl(-\langle A,\Delta(\hat y,y)\rangle/\gamma\bigr),$$9. Second, a temporal feature extractor $\gamma>0$0, consisting of Transformer blocks and a 2-layer MLP, is trained to match the frozen image encoder by minimizing $\gamma>0$1. The forecasting loss then uses

$\gamma>0$2

The total SATL is

$\gamma>0$3

with reported robust settings $\gamma>0$4, $\gamma>0$5, $\gamma>0$6. The supplied implementation notes that one often adds a vanilla MSE term, yielding

$\gamma>0$7

Pretraining of the perceptual modules is reported as 30 epochs for the autoencoder and 10 epochs for the temporal feature extractor, with learning rate $\gamma>0$8 using Adam and batch size $\gamma>0$9. The added training-time cost is reported as approximately $\Delta$0 on an RTX4090, while inference incurs no additional cost because the perceptual modules are used only during training (Yu et al., 31 Jul 2025).

5. Reported empirical behavior

The DILATE papers report experiments on synthetic step functions, ECG5000 waveform forecasting, and traffic occupancy forecasting, using both a fully connected MLP and a Seq2Seq-GRU, and comparing against MSE, soft-DTW, LSTNet, and TT-RNN. The supplied summary states that DILATE dramatically improves DTW and TDI relative to MSE while matching or slightly trading off MSE, and that it matches soft-DTW on shape while substantially improving temporal alignment and MSE. For the synthetic $\Delta$1 task under a Seq2Seq model, the reported MSE values are $\Delta$2 for the MSE-loss model, $\Delta$3 for the soft-DTW-loss model, and $\Delta$4 for DILATE; the broader conclusion given in the supplied text is that DILATE is advantageous on shape and time metrics rather than pure MSE (Guen et al., 2019).

The 2025 SATL is evaluated on eight standard long-range forecasting benchmarks—ETTh1, ETTh2, ETTm1, ETTm2, Weather, Exchange, Electricity, and ILI—using TimeMixer, PatchTST, TimesNet, and Autoformer as backbones. The supplied report states that SATL-trained models uniformly improve both MSE and TGSI. On ETTh2 with TimeMixer, the reported comparison is MSE $\Delta$5 and TGSI $\Delta$6. Averaged over all settings, the reported gains are MSE $\Delta$7 and TGSI $\Delta$8 (Yu et al., 31 Jul 2025).

An ablation study on ETTh1 with a PatchTST backbone and horizon $\Delta$9 is reported as follows (Yu et al., 31 Jul 2025):

Configuration	MSE	TGSI
All SATL comps + MSE	0.469	0.5944
$\| \hat y_i-y_j\|_2^2$0MSE-term (only SATL)	0.515	0.5695
$\| \hat y_i-y_j\|_2^2$1Perceptual	0.478	0.5880
$\| \hat y_i-y_j\|_2^2$2Freq.	0.557	0.5338
$\| \hat y_i-y_j\|_2^2$3Diff.	0.506	0.5790
Pure MSE baseline	0.470	0.5906

These results are accompanied by the claim that removing any component degrades performance, especially the frequency or the MSE term for long horizons. A plausible implication is that the three SATL components in the 2025 formulation play materially different roles: local slope preservation, dominant periodic structure recovery, and feature-space approximation of geometric similarity.

6. Relation to other shape-aware objectives and recurrent points of confusion

The supplied literature makes clear that not all shape-aware forecasting losses encode the same invariances. TILDE-Q defines shape-awareness through three explicitly transformation-invariant components: amplitude shifting $\| \hat y_i-y_j\|_2^2$4, phase shifting $\| \hat y_i-y_j\|_2^2$5, and uniform amplitude scaling. Its loss combines a softmax-based amplitude-shift term, a Fourier-magnitude phase term, and a normalized-autocorrelation amplitude term, with the stated goal that two signals differing only by constant vertical or temporal shift should be treated as very similar (Lee et al., 2022).

This differs materially from DILATE-style SATL. DILATE does not make phase shifts cost-free; rather, it introduces $\| \hat y_i-y_j\|_2^2$6 specifically to penalize off-diagonal matches in the alignment path, so temporal displacement is treated as error even when the global shape is preserved (Guen et al., 2021). The 2025 SATL differs again: it does not use DTW at all, but instead approximates image-based geometric structure with first-order differences, spectral matching, and an aligned perceptual feature space (Yu et al., 31 Jul 2025).

The comparative evidence in the supplied material reflects these differences. TILDE-Q reports consistent average reductions of $\| \hat y_i-y_j\|_2^2$7 in MSE and $\| \hat y_i-y_j\|_2^2$8 in MAE across eight modern architectures and seven real datasets when replacing MSE, and reports that on synthetic abrupt-change data it outperforms both MSE and DILATE in MSE, DTW, TDI, and LCSS metrics. The same source states that qualitative plots show DILATE can be misled by DTW’s noise-sensitivity, whereas TILDE-Q’s combined invariances yield visually crisper forecasts (Lee et al., 2022).

Several practical limitations also recur. DILATE-style SATL has $\| \hat y_i-y_j\|_2^2$9 dynamic-programming cost, requires care with numerical stability, and can exhibit degenerate behavior when $\gamma\to 0$0 is pushed toward the endpoints; too large a $\gamma\to 0$1 makes the shape term nearly $\gamma\to 0$2-like, while too small a $\gamma\to 0$3 risks underflow or overflow (Guen et al., 2019). The 2025 SATL avoids DTW but introduces pretraining stages, a handcrafted vertical-expansion image mapping, and about $\gamma\to 0$4 additional training time; its own discussion identifies extension to multivariate cross-variable structure and anomaly-detection settings as future directions (Yu et al., 31 Jul 2025).

Taken together, these formulations indicate that SATL is best understood as a design family for forecasting losses that privilege structural fidelity. This suggests that the decisive question is not whether a method is “shape-aware” in the abstract, but which distortions it is intended to ignore, which it is intended to penalize, and which representation—warping path, Fourier spectrum, autocorrelation profile, or learned perceptual embedding—it uses to encode temporal structure.