Time-Series Adaptive Transformation Optimization

• TATO is a data-centric framework that adapts time-series models by optimizing data transformation pipelines rather than adjusting model weights.
• It achieves adaptation by searching over a compact space of transformation operators on a small subset of target historical data, keeping the large time series models frozen.
• Empirical evaluations show significant MSE and MAE reductions across various datasets, highlighting improved robustness and deployment efficiency.

Searching arXiv for relevant papers on Time-Series Adaptive Transformation Optimization and closely related adaptive transformation methods. Time-Series Adaptive Transformation Optimization (TATO) is a data-centric framework for adapting time-series models under domain diversity and nonstationarity by optimizing transformations of the data pipeline rather than modifying model weights. In the specific formulation introduced for domain-shared forecasting with frozen large time series models (LTMs), TATO keeps a pre-trained model fixed and searches for an optimally configured transformation pipeline over historical data from the target domain, then reuses that pipeline at inference time (Qiu et al., 28 Feb 2026). The method is motivated by the observation that LTMs such as Timer, Moirai, and Chronos remain sensitive to scale, periodicity, noise level, trend structure, outliers, and usable context length, and that these factors can often be addressed more efficiently through preprocessing and postprocessing than through domain-specific finetuning (Qiu et al., 28 Feb 2026).

1. Definition, scope, and terminological boundaries

In its strict usage, TATO refers to the framework in "Adapt Data to Model: Adaptive Transformation Optimization for Domain-shared Time Series Foundation Models" (Qiu et al., 28 Feb 2026). That paper frames the task as Frozen Foundation Model-based Domain-shared Forecasting ("FrozenForecasting"), in which a single frozen pre-trained LTM is deployed across domains and adapted through lightweight, domain-specific preprocessing rather than through repeated finetuning.

A recurrent misconception is to treat any adaptive time-series transform as "TATO." The literature supplied here does not support that equivalence. "Adaptive Law-Based Transformation (ALT): A Lightweight Feature Representation for Time Series Classification" introduces Adaptive Law-Based Transformation (ALT) as an extension of Linear Law-Based Transformation (LLT) for time-series classification, but the paper explicitly does not use the term TATO; it can only be viewed, in a broader interpretive sense, as a related adaptive transformation method (Kurbucz et al., 16 Jan 2025). Likewise, other papers describe adaptive wavelet transforms, optimizer adaptation, test-time calibration, output-space correction, or sequential segment selection, but these are separate methods with their own names and objectives (Yu et al., 2024, Dong et al., 10 Mar 2026, Medeiros et al., 29 Jun 2025, Liu et al., 8 May 2026, Messica et al., 20 Feb 2026).

Another misconception is that TATO is a finetuning method. In the core formulation, the model $M$ remains frozen, and adaptation occurs through the transformation pipeline $h$, not through updates to backbone parameters. The optimization target is written as

$h^* = \min \, \mathcal{L}(M, D_{history}, h),$

where $M$ is the frozen LTM, $D_{history}$ is historical data from the target domain, and $h$ is a candidate transformation pipeline (Qiu et al., 28 Feb 2026). This places TATO within a data-centric rather than model-centric adaptation paradigm.

2. Problem formulation in frozen foundation-model forecasting

The central problem addressed by TATO is the degradation of zero-shot or frozen LTM forecasting under domain diversity and nonstationarity. The paper argues that domains differ in scale, periodicity, noise level, trend structure, outliers, and usable context length, so a transformation beneficial in one domain may be harmful in another. The usual response—domain-specific finetuning—creates two drawbacks: cost and scalability, and a generalization trade-off in which domain-specific adaptation can reduce the generality of the base foundation model (Qiu et al., 28 Feb 2026).

TATO addresses this by searching over transformation pipelines on target-domain history and then reusing the selected pipeline for future unseen data. The historical search is intentionally lightweight: the reported setup uses only a small subset of training data, typically 100 or 500 samples, amounting to less than 2% of total training samples, and the test set is never used during pipeline search (Qiu et al., 28 Feb 2026). This design is important because it separates adaptation from retraining while still grounding the selected transformation in domain-specific evidence.

The experimental formulation evaluates TATO on the ETT subsets ETTh1, ETTh2, ETTm1, and ETTm2, together with Electricity, Exchange, Traffic, and Weather, using prediction horizons 24, 48, 96, and 192. The base LTMs are Timer-UTSD, Timer-LOTSA, Moirai-small, Moirai-base, Moirai-large, and Chronos-tiny (Qiu et al., 28 Feb 2026). The primary reported metrics are MSE and MAE, while additional metrics such as RMSE, MAPE, and MSPE are used during ranking.

3. Transformation families and operator design

TATO organizes its search space into three transformation families containing nine operators. The search space is described as intentionally compact but expressive (Qiu et al., 28 Feb 2026).

Family	Operators	Role
Context slicing	Trimmer, Sampler, Aligner	Modify context length, sampling frequency, and patch alignment
Scale normalization	Scaler, Differencer, Warper	Stabilize scale, induce stationarity, compress extremes
Outlier correction	Denoiser, Imputator, Clipper	Smooth noise, replace anomalies, constrain extreme values

Context slicing changes what portion and resolution of the input history the LTM sees. The Trimmer keeps only the most recent $P_1$ steps,

$$T_{\text{trimmer}}(x, P_1) = x[\text{len}(x)-P_1 : \text{len}(x)],$$

reflecting the claim that more history is not always better (Qiu et al., 28 Feb 2026). The Sampler changes sampling frequency, for example via downsampling,

$$T_{\text{sampler}(\text{Downsampling})}(x, P_s) = x[::P_s],$$

which can suppress high-frequency noise and highlight longer-range periodicity (Qiu et al., 28 Feb 2026). The Aligner pads a sequence so its length matches patch-size requirements; one example pads with the mean,

$$T_{\text{aligner}(\text{Mean})}(x) = \text{pad}(x, \text{mean}(x)).$$

Scale normalization addresses magnitude mismatch and nonstationarity. The standard Scaler is

$h$0

while the Differencer induces stationarity through adjacent differences,

$h$1

and the Warper applies a log-style compression,

$h$2

These operators are presented as ways to make the frozen model less sensitive to changing variance, trends, and heavy-tailed magnitudes (Qiu et al., 28 Feb 2026).

Outlier correction targets extreme values that an LTM might otherwise interpret as structural patterns. The Denoiser is given in moving-average form,

$h$3

the Imputator replaces anomalies under an IQR-based rule,

$h$4

and the Clipper bounds values by IQR thresholds,

$h$5

The combined effect is to stabilize the input-output distribution seen by the frozen LTM (Qiu et al., 28 Feb 2026).

4. Pipeline search, augmentation, and ranking

TATO treats transformation discovery as a hyperparameter optimization problem over pipelines. Candidate pipelines are defined by selecting operators, their hyperparameters, and—subject to heuristics—their order (Qiu et al., 28 Feb 2026). The reported search space includes, for example, trimmer_length from $h$6 to $h$7, sampler_factor in $h$8, aligner in $h$9, scaler in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$0, differencer_level in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$1, warper in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$2, outlier_detect in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$3, denoiser in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$4, and clipper_factor in $h^* = \min \, \mathcal{L}(M, D_{history}, h),$5 (Qiu et al., 28 Feb 2026).

The operator order is not fully permuted. Instead, TATO imposes heuristics: Trimmer first, outlier correction before normalization, and Aligner near the end. This restriction is explicitly intended to avoid wasting search on meaningless reorderings and to keep the search space manageable (Qiu et al., 28 Feb 2026). Pipeline search is then carried out with Tree-structured Parzen Estimator (TPE), i.e., Bayesian optimization over black-box pipeline evaluations.

A distinctive feature is the use of augmentations only during optimization, not at final inference time. The reported augmentations include magnitude flip, time flip, magnitude warp, time warp, EWMA smoothing, jitter, horizontal or vertical shifts, and slope injection (Qiu et al., 28 Feb 2026). The stated rationale is robustness: optimization on augmented samples prevents the chosen pipeline from overfitting a narrow historical subset and exposes it to conditions resembling future test distributions.

Pipeline selection uses a two-stage Pareto-based ranking. In Stage 1, pipelines are filtered on augmented samples using metric combinations such as $h^* = \min \, \mathcal{L}(M, D_{history}, h),$6 and $h^* = \min \, \mathcal{L}(M, D_{history}, h),$7; the worst pipelines are excluded, and the framework retains 16 candidates (Qiu et al., 28 Feb 2026). In Stage 2, the survivors are ranked on the original samples using a weighted sum of metric ranks with MSE weight = 3, MAE weight = 2, and other metrics = 1. This structure reflects an explicit trade-off between robustness across perturbations and final accuracy on the domain’s original data (Qiu et al., 28 Feb 2026).

5. Empirical behavior, efficiency, and limitations

The headline empirical claim is that TATO achieves a maximum MSE reduction of 65.4% and an average MSE reduction of 13.6%, and that it matches or outperforms vanilla baselines in 84.3% of the 96 evaluated cases (Qiu et al., 28 Feb 2026). The paper further reports strong gains on Exchange, with up to 35.1% improvement, and on ETTm2, with improvements also above 20%. Electricity and Traffic show consistent moderate gains, while Weather is explicitly identified as weaker, with gains that are small and sometimes slightly negative (Qiu et al., 28 Feb 2026).

At the model level, the average MSE reductions are reported as about 6.0% for Timer-UTSD, about 24.8% for Timer-LOTSA, about 9.2% for Moirai-small, about 12.2% for Moirai-base, about 14.5% for Moirai-large, and about 10.3% for Chronos-tiny (Qiu et al., 28 Feb 2026). The paper also notes that the MAE distribution shifts downward after applying TATO and exhibits a narrower spread, which it interprets as improved robustness.

The reported percentage improvement metric is

$h^* = \min \, \mathcal{L}(M, D_{history}, h),$8

where $h^* = \min \, \mathcal{L}(M, D_{history}, h),$9 is the vanilla baseline error and $M$0 is the TATO error (Qiu et al., 28 Feb 2026). Positive values therefore indicate a reduction in error. This metric is used to summarize how much the transformation pipeline improves a frozen model without weight updates.

Efficiency is a central part of the method’s positioning. The optimization phase is reported as typically completing in under 2 minutes, with most experiments finishing within 120 seconds, and the additional deployment-time overhead for batch size 1 is reported as under 3 milliseconds (Qiu et al., 28 Feb 2026). This profile is presented as a major practical advantage over finetuning, especially when many domains must share one foundation model.

The limitations are also explicit. Some model-dataset pairs show limited improvement or slight degradation, especially on Weather. The paper names Timer-UTSD on Weather, Moirai-large on Exchange, and Chronos-tiny on ETTh2 as edge cases that remain challenging (Qiu et al., 28 Feb 2026). In ablations, the full configuration performs best overall; removing Trimmer or Scaler causes a substantial drop, indicating that these operators are especially important. Removing Denoiser or TwoStageRank can increase mean improvement in some cases but hurts variance and median performance, so the full design is presented as more robust across domains (Qiu et al., 28 Feb 2026).

6. Relation to adjacent adaptive transformation and optimization paradigms

The strict TATO formulation is one member of a broader research landscape in which adaptation is shifted away from full backbone retraining and toward transformations, calibration, or low-capacity corrections.

In feature-space adaptation for classification, ALT extends LLT by replacing fixed-length windows with variable-length shifted time windows, searching over $M$1, where $M$2 is subsequence length, $M$3 is embedding dimension, and $M$4 is shift/stride. ALT extracts subsequences, forms shapelet vectors through time-delay embedding and spectral decomposition, constructs a class-partitioned matrix $M$5, and transforms each instance via $M$6, after which squared responses are aggregated statistically for KNN or SVM classification (Kurbucz et al., 16 Jan 2025). This is closely related in spirit to adaptive transformation, but the paper explicitly states that ALT is not the same as “TATO”.

In learnable transform architectures, AdaWaveNet introduces Adaptive Wavelet Transformation for non-stationary time series through a lifting scheme whose prediction and update operators are parameterized by learnable 1D convolutions. The model decomposes $M$7, applies stacked AdaWave blocks to the seasonal component, and reconstructs through inverse lifting blocks, with the transform trained end-to-end for forecasting, imputation, and super-resolution (Yu et al., 2024). This suggests a distinct line of work in which the transform basis itself is optimized jointly with the task loss.

In parameter-efficient forecasting TTA, PETSA wraps a frozen pretrained forecaster with an input calibration module and an output calibration module, updating only low-rank factors $M$8 and $M$9, gating vector $D_{history}$0, and bias $D_{history}$1 at test time. The calibrated signals are

$D_{history}$2

with $D_{history}$3, and the test-time objective combines a Huber loss, a frequency-domain loss, and a patch-wise structural loss (Medeiros et al., 29 Jun 2025). The related post-training framework "Human in the Loop Adaptive Optimization for Improved Time Series Forecasting" instead applies interpretable output transformations

$D_{history}$4

selected by Random Search, SH-HPO, PPO, or GA, and it proves that the optimal affine correction $D_{history}$5 cannot increase MSE under the stated validation-distribution assumption (Tiomoko et al., 21 May 2025). Those papers broaden the notion of adaptive transformation from preprocessing to lightweight, post hoc forecast correction.

In forecasting adaptation under distribution shift, AdaNODEs and STEPS move the adaptation locus still further. AdaNODEs keeps all pre-trained encoder, latent NODE, and decoder weights frozen and updates only two scalar test-time parameters, $D_{history}$6 and $D_{history}$7, inside the latent ODE,

$D_{history}$8

using a loss that combines negative log-likelihood and KL divergence (Dang et al., 19 Jan 2026). STEPS, by contrast, does not adapt model parameters at all; it treats the revealed prefix residual $D_{history}$9 as a boundary condition on a temporal manifold, solves a Dirichlet Boundary Value Problem, and applies a bounded correction field

$h$0

through a Local Solver, Global Solver, and Spatiotemporal Manifold Fusion (SMF) (Liu et al., 8 May 2026). This is adaptive transformation in prediction space rather than in input space or parameter space.

Other papers push the adaptive idea beyond transformations in the narrow sense. TS_Adam modifies Adam for non-stationary forecasting by removing the second-order bias correction and setting $h$1, thereby improving responsiveness to drifting loss landscapes without introducing new hyperparameters (Dong et al., 10 Mar 2026). The adaptive Student’s $h$2-distribution work replaces static fitting by a moving objective,

$h$3

and tracks evolving $h$4, $h$5, and $h$6 by EMA-based moment updates (Duda, 2023). ARTIST formulates time-series reasoning as a sequential decision problem in which a controller selects temporal segments $h$7 and a reasoner answers from the accumulated segment set, using supervised fine-tuning and reinforcement learning to optimize question-adaptive segment selection (Messica et al., 20 Feb 2026). A plausible implication is that the broader TATO vocabulary is now spanning at least three distinct adaptation loci: data transformation before inference, forecast correction after inference, and optimization or selection mechanisms that adapt to temporal nonstationarity during inference or training.