betaFeedback

WorksComputer Science

SWIFT: Sample-Wise Instance Fine-Tuning of Time-Series Foundation Models for Forecasting

Abstract

While current few-shot adaptation strategies for time series forecasting foundation models are predominantly dataset-centric, the necessity of obtaining an additional dataset for the finetuning process poses a significant hurdle for adoption. To address this constraint, we propose Sample-Wise Instance Fine-Tuning (SWIFT), a novel forecasting adaptation framework that achieves fine-grained specialization by leveraging structural priors from the target instance's own history. This shift allows the fine-tuning process to be entirely transparent to the end-user, which significantly enhances the ease of adoption. To implement this framework, SWIFT synergizes Self-Retrieval for historical mining and Frequency-Aware Loss for spectral consistency, while employing Deviation-Aware Dynamic Fusion to balance adaptation with general knowledge preservation. Extensive experiments across four TSFM backbones and three long-history benchmark subsets validate SWIFT, yielding average MASE and WQL reductions of 6.41% and 4.83% across 24 combinations of four backbones and six Chronos Benchmark II datasets. To the best of our knowledge, SWIFT is the first framework to investigate sample-wise instance fine-tuning of time-series foundation models, specializing the model for an individual prediction window using only the corresponding forecasting sample's own observed history.

Introduction

Time Series Foundation Models (TSFMs), trained on massive corpora of diverse temporal data, have ushered in a new era of zero-shot forecasting capabilities [10, 11, 18, 31]. These models excel at capturing general temporal patterns, across a wide range of domains [13, 12]. Despite their outstanding zero-shot capabilities, adapting TSFMs to a specific forecasting instance remains an important research question [8, 20]. Real-world time series exhibit heterogeneous temporal patterns and non-stationarity [27, 41, 40], with local trends and seasonal behavior varying across sequences and over time. These characteristics motivate investigating whether the target instance’s own observed history can provide useful supervision for specialization.

Existing research explores adaptation through few-shot fine-tuning [9], transfer learning [14], and retrieval-augmented forecasting [30]. Recent studies have also investigated fine-tuning TSFMs such as Chronos-2 on task-specific datasets, highlighting the potential and limitations of dataset-level adaptation [46]. These approaches leverage target-domain training data or relevant support examples to improve forecasting. However, such data may not be readily available when a forecasting request provides only the target instance’s observed history. This constraint motivates a complementary question: can a pretrained TSFM be specialized for an individual prediction window using context–target pairs constructed solely from that instance’s own history?

The practical challenge is the availability and preparation of adaptation data beyond the current forecasting sample. While a target-domain training split may already exist in some applications, others require additional data collection or curation. In streaming or privacy-restricted environments, access restrictions may further limit the use of other sequences [17]. Even when support data are available, their relevance to the current forecasting instance must be assessed, since distribution mismatch can limit transfer benefits [44, 24]. These constraints motivate a complementary history-only adaptation setting.

To address this setting, we propose Sample-Wise Instance Fine-Tuning (SWIFT). A forecasting sample comprises an observed history, a forecast origin, and a requested future prediction window. SWIFT treats the sample’s own history as an intrinsic adaptation dataset and specializes the pretrained model for the corresponding prediction window. All adaptation targets are observed before the forecast origin. To the best of our knowledge, SWIFT is the first framework to investigate history-only sample-wise fine-tuning of TSFMs at this prediction-window.

As illustrated in Figure 1, SWIFT fundamentally shifts the adaptation objective from “learning from others” to “learning from self”. It posits that by learning from the historical behaviors inherent to the specific instance itself, we can specialize a generalist foundation model into an instance-specific expert at pre-inference adaptation time, without requiring a separate adaptation dataset or access to other sequences, while introducing model-side computation for historical retrieval and parameter updates.

Figure 1. Learning from others versus learning from self. The illustrated support-set adaptation learns from a separately supplied support set Dsupport\mathcal{D}_{\mathrm{support}} of similar sequences, leveraging cross-instance information. In contrast, SWIFT constructs an intrinsic support set Dintr\mathcal{D}_{\mathrm{intr}} by mining context–target pairs from the target instance’s own history. Both approaches update only LoRA parameters, while SWIFT constructs its support set from the current sample’s own history to specialize the model for its prediction window.

To implement this paradigm, we address data scarcity and overfitting through three key components. First, we employ Self-Retrieval to mitigate data limitations by mining the sample’s own history for relevant subsequences, unlike standard retrieval methods that query external databases [21, 30]. Second, we refine the spectral consistency approach [42] with Frequency-Aware Loss (FreAL), using adaptive weighting to ensure the model captures global dynamics rather than local fluctuations. Finally, to prevent catastrophic forgetting, we introduce a Deviation-Aware Dynamic Fusion mechanism, adapting trust-region principles [36, 37] as an inference-time safeguard to balance the specialized expert with the robust baseline.

In summary, our specific contributions are as follows:

  • We pioneer sample-wise instance fine-tuning for TSFMs at the granularity of individual forecasting samples. SWIFT specializes a pretrained model for each requested prediction window using only the corresponding sample’s observed history, eliminating the need for a separate adaptation dataset.

  • We propose a unified framework to materialize this paradigm by synergizing three components to address core adaptation challenges: Self-Retrieval, Frequency-Aware Loss, and a Deviation-Aware Dynamic Fusion.

  • We validate our method on curated long-history subsets of the fev-bench, GIFT-Eval, and Chronos Benchmark II. Notably, SWIFT achieves a 6.99% average WQL improvement gain on the Chronos Benchmark II compared to the Chronos-2 zero-shot baseline.

Related Work

Time-Series Forecasting Models

Classical forecasting methods, such as ARIMA [6] and related statistical models [15, 16, 4], are commonly fitted to individual series and can capture relatively simple temporal dynamics. Deep forecasting models instead share parameters across collections of series, using architectures such as RNNs [35, 29], CNNs [28], and Transformers [26, 50, 49, 7]. Although such global models improve cross-series generalization, shared representations may not fully capture instance-specific dynamics.

Recent TSFMs further scale this global modeling paradigm through large and heterogeneous pre-training corpora [25, 43, 39, 33]. Chronos-2 [2] supports zero-shot univariate and multivariate forecasting, while Moirai 2.0 [23] and TimesFM [10] explore efficient decoder-based architectures and large-scale pre-training. TiRex [5] further advances large-scale temporal pre-training with strong zero-shot forecasting capabilities. These advances provide strong zero-shot generalization, while also motivating adaptation methods for target instances whose local dynamics differ from those captured during pre-training.

Adaptation Paradigms for Time Series Forecasting Foundation Models

Zero-shot. Zero-shot forecasting serves as the baseline for evaluating TSFMs, utilizing pretrained weights without parameter updates [10, 2, 23, 11]. While attractive for its simplicity, zero-shot inference may exhibit reduced performance under distribution shifts, as generic representations may not fully capture target-specific dynamics.

Few-shot adaptation. Few-shot fine-tuning updates a pre-trained TSFM using a support set relevant to the target task [9]. Recent strategies such as Multi-Scale Fine-Tuning (MSFT) incorporate multi-scale temporal modeling and outperform full fine-tuning and representative parameter-efficient methods across multiple TSFM backbones [32]. Retrieval-augmented and in-context methods instead use examples from external time series as additional forecasting context or demonstrations [34, 47, 48, 9]. Although these approaches can improve target-specific predictions, they require target-domain or external adaptation data, and their effectiveness depends on the quality and alignment of the selected examples. Poorly aligned data may cause negative transfer, and fine-tuning Chronos-2 can degrade performance on certain datasets [45].

Test-time adaptation. Test-time adaptation (TTA) updates a pre-trained forecaster during inference to track distribution shifts in an evolving test stream [22]. Recent methods leverage progressively revealed observations or partially observed ground truth to construct self-supervised objectives and continually recalibrate the model under non-stationary conditions [19]. Although this online adaptation can improve responsiveness to recent changes, it requires future test-stream observations to become available and may introduce additional waiting time. Moreover, repeated parameter updates can lead to error accumulation, parameter drift, and unstable performance when the incoming observations are noisy or unrepresentative.

In this work, we introduce SWIFT, a sample-wise instance fine-tuning framework for TSFMs. A sample denotes a forecasting instance comprising an observed history, a forecast origin, and a requested prediction window. SWIFT specializes the pretrained model for this window using context–target pairs drawn exclusively from the sample’s own history. Unlike dataset-level fine-tuning, its specialization targets an individual forecasting instance; unlike continual TTA, it completes adaptation before forecasting using only observations available at the current origin. Table I summarizes the representative protocols.

Adaptation StrategyUpdate StageExternal DataAdaptation DataSupervision TypeAdaptation Granularity
Zero-shotNoneNoneNoneNoneN/A
Few-shot adaptationOfflineTypically requiredSupport SetObserved targetsDataset-level
TTAOnlineNoneTest StreamSelf-supervisionStream/batch-level
SWIFT (Ours)Pre-inferenceNoneTarget HistoryInternal PriorsInstance-level

Table I. Comparison of SWIFT with representative adaptation paradigms for TSFMs.

Problem Formulation

Preliminaries

Let X={x1:T}∈RC×T\mathcal{X} = \{x_{1:T}\} \in\mathbb{R}^{C \times T} denote a multivariate time-series instance with CC channels and length TT. Given a foundation model fθ0f_{\theta_0} pre-trained on Dsource\mathcal{D}_{\mathrm{source}}, zero-shot forecasting predicts the future horizon HH from a context window of length LL: x^T+1:T+H=fθ0(xT−L+1:T)\hat{x}_{T+1:T+H} = f_{\theta_0}(x_{T-L+1:T}). The ideal target-specific adaptation objective is

θ∗=arg⁡min⁡θ L(fθ(xT−L+1:T),xT+1:T+H).(1)\theta^{*} = \underset{\theta}{\arg\min}\ \mathcal{L}\bigl(f_{\theta}(x_{T-L+1:T}), x_{T+1:T+H}\bigr). \tag*{(1)}

Since xT+1:T+Hx_{T+1:T+H} is unavailable at inference time, (1) is an ideal objective rather than an executable adaptation procedure.

External-Support Fine-Tuning

Conventional fine-tuning assumes access to an external support set

Dsupport={(xctx(i),xtgt(i))}i=1N,(2)\mathcal{D}_{\mathrm{support}} = \left\{\left(x_{\mathrm{ctx}}^{(i)},x_{\mathrm{tgt}}^{(i)}\right)\right\}_{i=1}^{N}, \tag*{(2)}

where each context–target pair is sampled from a support distribution Psupport\mathcal{P}_{\mathrm{support}} assumed to be aligned with the target instance. Starting from θ0\theta_0, adaptation minimizes

θ∗=arg min⁡θ E(xctx,xtgt)∼Dsupport[L(fθ(xctx),xtgt)](3)\theta^{*} = \underset{\theta}{\operatorname{arg\,min}}\ \mathbb{E}_{(x_{\mathrm{ctx}},x_{\mathrm{tgt}})\sim\mathcal{D}_{\mathrm{support}}}\left[\mathcal{L}\left(f_{\theta}(x_{\mathrm{ctx}}),x_{\mathrm{tgt}}\right)\right] \tag*{(3)}

This formulation becomes unavailable when Dsupport=∅\mathcal{D}_{\mathrm{support}} = \varnothing and may be suboptimal when the external support distribution is misaligned with the target. Because different sequences can exhibit distinct non-stationary dynamics, optimizing over Psupport\mathcal{P}_{\mathrm{support}} does not necessarily minimize the risk on Ptarget\mathcal{P}_{\mathrm{target}}:

Ex∼Psupport[L(fθ(x))]≠Ex∼Ptarget[L(fθ(x))].(4)\mathbb{E}_{x\sim\mathcal{P}_{\mathrm{support}}}\left[\mathcal{L}(f_{\theta}(x))\right] \ne\mathbb{E}_{x\sim\mathcal{P}_{\mathrm{target}}}\left[\mathcal{L}(f_{\theta}(x))\right]. \tag*{(4)}

Moreover, sufficiently long target histories can themselves contain predictive temporal dependencies, reflected by the non-negligible mutual information I(x1:T;xT+1:T+H)I(x_{1:T};x_{T+1:T+H}). External-support methods therefore face two practical limitations: distribution misalignment and the unavailability or retrieval cost of peer sequences.

Sample-Wise Instance Fine-Tuning

SWIFT operates under data isolation and uses only the target history x1:Tx_{1:T}. We assume that a sufficiently long sequence contains recurrent temporal patterns that can provide instance-specific supervision. SWIFT therefore constructs an intrinsic support set

Dintr=MinePriors⁡(x1:T)={(xctx(i),xtgt(i))}i=1N,(5)\mathcal{D}_{\mathrm{intr}} = \operatorname{MinePriors}(x_{1:T}) = \left\{\left(x_{\mathrm{ctx}}^{(i)},x_{\mathrm{tgt}}^{(i)}\right)\right\}_{i=1}^{N}, \tag*{(5)}

where MinePriors⁡(⋅)\operatorname{MinePriors}(\cdot) extracts informative context–target pairs directly from the target timeline.

SWIFT then performs parameter-efficient adaptation within a constrained subspace around the pre-trained model:

θswift∗=arg min⁡θ∈Ω(θ0) E(x,y)∼Dintr[Ltask(fθ(x),y)],(6)\theta_{\mathrm{swift}}^{*} = \underset{\theta\in\Omega(\theta_0)}{\operatorname{arg\,min}}\ \mathbb{E}_{(x,y)\sim\mathcal{D}_{\mathrm{intr}}}\left[\mathcal{L}_{\mathrm{task}}(f_{\theta}(x),y)\right], \tag*{(6)}

where Ω(θ0)\Omega(\theta_0) denotes a parameter-efficient adaptation subspace, such as LoRA, centered at θ0\theta_0. By learning from the target instance’s own historical patterns, SWIFT specializes the foundation model without external support data.

Realizing this formulation requires addressing three challenges: (1) constructing a high-quality Dintr\mathcal{D}_{\mathrm{intr}} from the available history; (2) learning structural dynamics without overfitting transient noise; and (3) controlling harmful over-adaptation. The next section addresses these challenges through Self-Retrieval, Frequency-Aware Loss, and Deviation-Aware Dynamic Fusion, respectively.

Methodology

In this section, we present the realization of the SWIFT paradigm. As illustrated in Figure 2, SWIFT consists of three modules: (1) Self-Retrieval, which mines structurally relevant historical and recent subsequences to construct the intrinsic support set Dintr\mathcal{D}_{\mathrm{intr}}; (2) Frequency-Aware Loss (FreAL), which guides parameter-efficient fine-tuning by emphasizing spectral discrepancies while suppressing noisy frequency components; and (3) Deviation-Aware Dynamic Fusion, which adaptively combines the adapted and zero-shot predictions to mitigate over-adaptation. During fine-tuning, an additional prediction-space anchor regularizer constrains the adapted model’s quantile forecasts toward those of the frozen foundation model.

Figure 2. Overview of SWIFT. The framework consists of three stages: (a) Self-Retrieval constructs an intrinsic support set from similar historical and recent subsequences; (b) Frequency-Aware Loss adapts the model with FreAL to emphasize informative spectral discrepancies, together with an anchor regularizer that limits prediction drift; and (c) Deviation-Aware Dynamic Fusion combines adapted and zero-shot predictions to reduce harmful over-adaptation.

Self-Retrieval

To address the challenge of data scarcity in sample-wise instance fine-tuning, we propose a Self-Retrieval mechanism.

Query and search space construction. Let xquery=xT−Lquery+1:Tx_{\mathrm{query}} = x_{T-L_{\mathrm{query}}+1:T} denote the current lookback window used as the retrieval query. The query length LqueryL_{\mathrm{query}} is determined from the dataset’s known seasonality or, when such prior information is unavailable, through a spectral heuristic. In the latter case, we apply a Fast Fourier Transform (FFT) to the local context and select the period associated with the largest non-DC spectral peak.

We construct the historical search space using a sliding window of length LqueryL_{\mathrm{query}}. Specifically, the window is moved along the observed history x1:Tx_{1:T} with stride one, producing candidate contexts

xcand(i)=xi:i+Lquery−1,(7)x_{\mathrm{cand}}^{(i)} = x_{i:i+L_{\mathrm{query}}-1}, \tag*{(7)}

where only windows satisfying i+Lquery+H−1≤Ti + L_{\mathrm{query}} + H - 1 \le T are retained. This constraint ensures that each candidate context has a fully observed subsequent target window

ycand(i)=xi+Lquery:i+Lquery+H−1,(8)y_{\mathrm{cand}}^{(i)} = x_{i+L_{\mathrm{query}}:i+L_{\mathrm{query}}+H-1}, \tag*{(8)}

which can be used as supervision during adaptation. The resulting search space is

S={xi:i+Lquery−1∣1≤i≤T−Lquery−H+1}.(9)S = \left\{x_{i:i+L_{\mathrm{query}}-1} \mid1 \le i \le T - L_{\mathrm{query}} - H + 1\right\}. \tag*{(9)}

To identify candidate windows that exhibit temporal patterns similar to the query, we independently apply local Z-score normalization to each window:

x~=x−μxσx+ϵ,(10)\tilde{x} = \frac{x-\mu_x}{\sigma_x+\epsilon}, \tag*{(10)}

where μx\mu_x and σx\sigma_x are computed within the corresponding window. This normalization reduces the influence of level and scale differences, allowing the retrieval process to focus on temporal shape. The similarity between the query and a candidate window is then measured using the negative normalized L1L_1 distance:

s(xquery,xcand(i))=−1Lquery∑t=1Lquery∣x~query(t)−x~cand(i,t)∣.(11)s\left(x_{\mathrm{query}}, x_{\mathrm{cand}}^{(i)}\right) = -\frac{1}{L_{\mathrm{query}}}\sum_{t=1}^{L_{\mathrm{query}}}\left|\tilde{x}_{\mathrm{query}}^{(t)}-\tilde{x}_{\mathrm{cand}}^{(i,t)}\right|. \tag*{(11)}

A larger similarity score therefore indicates a closer match between the historical candidate and the current query.

Temperature-controlled sampling. Instead of greedily selecting the top-kk matches, which may lead to redundancy, we apply a temperature-controlled probabilistic sampling strategy to ensure diversity in the intrinsic support set. The probability of selecting the ii-th candidate is given by

P(i)=Softmax⁡(s(xquery,xi)τret),(12)P(i) = \operatorname{Softmax}\left(\frac{s(x_{\mathrm{query}},x_i)}{\tau_{\mathrm{ret}}}\right), \tag*{(12)}

where τret\tau_{\mathrm{ret}} is a temperature hyperparameter. We sample KK subsequences based on P(i)P(i) to form the intrinsic support set Dintr\mathcal{D}_{\mathrm{intr}}.

Dual-focus augmentation. Finally, to ensure that the model remains responsive to recent distribution shifts and local trends, we augment Dintr\mathcal{D}_{\mathrm{intr}} with a deterministic subset of the most recent sliding windows. The retrieved and recent samples are jointly used for adaptation and are further provided as covariates during forecasting, enabling the model to combine deep historical motifs with evolving local dynamics.

Frequency-Aware Loss (FreAL)

Recent work such as FreDF [42] has demonstrated the benefit of minimizing forecasting divergence in the frequency domain. However, existing frequency-domain objectives are primarily designed for data-rich supervised training. In our sample-wise adaptation setting, directly optimizing all frequency components may overfit high-frequency noise and transient anomalies, thereby disrupting the underlying periodic structure.

To address this issue, we propose Frequency-Aware Loss (FreAL), which adapts the spectral objective to the intrinsic support set Dintr\mathcal{D}_{\mathrm{intr}} constructed by Self-Retrieval. FreAL first identifies the frequency components that contribute most to the current forecasting error, then constructs a weighted frequency-domain loss, and finally combines it with the time-domain objective. Diagnostic spectral weighting strategy. Given the prediction y^\hat{y} and ground truth yy, we first apply the Fast Fourier Transform to obtain their frequency representations F^\hat{F} and FF, and compute the absolute spectral residual E=∣F^−F∣∈RC×LfreqE = \lvert\hat{F} - F \rvert\in\mathbb{R}^{C \times L_{\mathrm{freq}}}, where LfreqL_{\mathrm{freq}} denotes the spectral length. The DC component is separated from the AC components and is always assigned a weight of 1.0 to preserve mean-level consistency.

For the AC components, we sort their residual magnitudes in descending order, E(1)≥E(2)≥⋯≥E(Lfreq−1)E_{(1)} \ge E_{(2)} \ge\cdots\ge E_{(L_{\mathrm{freq}}-1)}, and select the smallest number of components whose cumulative residual reaches a predefined threshold τspec\tau_{\mathrm{spec}}:

K=min⁡{k | ∑i=1kE(i)∑j=1Lfreq−1E(j)≥τspec}.(13)K = \min\left\{ k \,\middle|\, \frac{\sum_{i=1}^{k} E_{(i)}}{\sum_{j=1}^{L_{\mathrm{freq}}-1} E_{(j)}} \ge\tau_{\mathrm{spec}} \right\}. \tag*{(13)}

The selected top-KK components are treated as critical frequencies and assigned whigh=1.0w_{\mathrm{high}} = 1.0, while the remaining components are assigned wlow=0.1w_{\mathrm{low}} = 0.1 to suppress less informative frequency variations. The resulting weights are then restored to their original frequency indices to form the final spectral mask WfinalW_{\mathrm{final}}.

Based on this mask, we define the frequency-domain loss as the weighted L1L_1 distance between the predicted and target spectra. To reduce the effect of random fluctuations during sample-wise adaptation, the loss is averaged over the batch of retrieved context–target pairs:

Lfreq=1B∑b=1B∥(F^b−Fb)⊙Wfinal(b)∥1,(14)\mathcal{L}_{\mathrm{freq}} = \frac{1}{B} \sum_{b=1}^{B} \left\| \left(\hat{F}_b - F_b\right) \odot W_{\mathrm{final}}^{(b)} \right\|_1, \tag*{(14)}

where BB denotes the number of retrieved training pairs.

Finally, FreAL combines the weighted frequency-domain loss with the time-domain forecasting loss:

LFreAL=(1−α)Ltime(y^,y)+αLfreq(F^,F),(15)\mathcal{L}_{\mathrm{FreAL}} = (1-\alpha)\mathcal{L}_{\mathrm{time}}(\hat{y}, y) + \alpha\mathcal{L}_{\mathrm{freq}}(\hat{F}, F), \tag*{(15)}

where α\alpha controls the trade-off between time-domain accuracy and spectral consistency. For quantile forecasting, Ltime\mathcal{L}_{\mathrm{time}} is the pinball loss over the predicted quantiles, while the spectral objective is computed on the median forecast. Through this progressive construction, FreAL focuses adaptation on dominant spectral discrepancies while reducing overfitting to noisy tail-frequency components.

To limit prediction drift during adaptation, we introduce an anchor loss that encourages the adapted model’s outputs to remain close to those of the frozen pretrained model:

Lanchor=MSE⁡(y^adapt,sg⁡[y^base]),(16)\mathcal{L}_{\mathrm{anchor}} = \operatorname{MSE}\left(\hat{y}_{\mathrm{adapt}}, \operatorname{sg}[\hat{y}_{\mathrm{base}}]\right), \tag*{(16)}

where both predictions use the same historical support inputs, and MSE is averaged over all quantiles and output positions. Here, sg denotes stop-gradient, ensuring that only the LoRA parameters are updated.

The complete adaptation objective is

Ladapt=LFreAL+λanchorLanchor,(17)\mathcal{L}_{\mathrm{adapt}} = \mathcal{L}_{\mathrm{FreAL}} + \lambda_{\mathrm{anchor}}\mathcal{L}_{\mathrm{anchor}}, \tag*{(17)}

where λanchor≥0\lambda_{\mathrm{anchor}} \ge0 controls the strength of regularization. Setting λanchor=0\lambda_{\mathrm{anchor}} = 0 recovers adaptation with FreAL alone; increasing it penalizes departures from the frozen baseline more strongly. Anchor regularization operates during parameter updates, while deviation-aware fusion controls how the resulting forecasts are combined at inference.

Deviation-Aware Dynamic Fusion

A critical risk in sample-wise adaptation is “over-adaptation”, where the model becomes too specialized to the retrieved history and fails to generalize to the immediate future, especially under sudden distribution shifts. To address this, we introduce the Deviation-Aware Dynamic Fusion mechanism. During inference, we maintain both the frozen foundation model fθ0f_{\theta_0} and the adapted expert fθswiftf_{\theta_{\mathrm{swift}}}. We generate two sets of predictions: the robust baseline prediction y^base\hat{y}_{\mathrm{base}} and the specialized prediction y^adapt\hat{y}_{\mathrm{adapt}}. We quantify their normalized forecast deviation as

Δ=1H∑t=1H∣y^adapt(t)−y^base(t)∣∣y^base(t)∣+ϵ.(18)\Delta= \frac{1}{H}\sum_{t=1}^{H}\frac{\left|\hat{y}_{\mathrm{adapt}}^{(t)}-\hat{y}_{\mathrm{base}}^{(t)}\right|}{\left|\hat{y}_{\mathrm{base}}^{(t)}\right|+\epsilon}. \tag*{(18)}

We posit that a small Δ\Delta indicates a safe adaptation where the expert refines the baseline, whereas a large Δ\Delta suggests potential overfitting or instability. We design a dynamic gating function to calculate the weight of the baseline model:

wbase=min⁡(λmax⁡,σmax⁡(0,Δ−τfus)),(19)w_{\mathrm{base}}=\min\left(\lambda_{\max},\sigma\max(0,\Delta-\tau_{\mathrm{fus}})\right), \tag*{(19)}

where τfus\tau_{\mathrm{fus}} is a tolerance threshold and σ\sigma is a scaling factor. The final prediction is a deviation-weighted combination:

y^final=(1−wbase)y^adapt+wbasey^base.(20)\hat{y}_{\mathrm{final}}=(1-w_{\mathrm{base}})\hat{y}_{\mathrm{adapt}}+w_{\mathrm{base}}\hat{y}_{\mathrm{base}}. \tag*{(20)}

This mechanism acts as an adaptive safety valve: it allows the specialized model to take the lead when predictions are consistent, but increases reliance on the foundation model when the adaptation deviates substantially. The same fusion weight is applied to the point and quantile forecasts. When λmax⁡<1\lambda_{\max}<1, the baseline contribution is capped rather than providing a complete fallback.

Overall Procedure

Algorithm 1 SWIFT: Sample-Wise Instance Fine-Tuning

Require: Observed history x1:Tx_{1:T}, pretrained model fθ0f_{\theta_0}, initial LoRA parameters ϕ0\phi_0, context length LL, horizon HH, adaptation steps SS, anchor weight λanchor\lambda_{\mathrm{anchor}}

Ensure: Final forecast y^final\hat{y}_{\mathrm{final}}

1. Self-Retrieval

  1. Construct query qq from the latest observations

  2. Extract eligible historical context–target pairs

  3. Score historical contexts against qq using Eq. (11)

  4. Sample KK historical pairs using temperature-controlled top-pp sampling

  5. Select RR recent context–target pairs

  6. Concatenate both subsets into Dintr\mathcal{D}_{\mathrm{intr}}

2. Instance-Specific Adaptation

  1. Set ϕ←ϕ0\phi\leftarrow\phi_0 and freeze θ0\theta_0

  2. (X,Y)←Dintr(X, Y) \leftarrow\mathcal{D}_{\mathrm{intr}}

  3. for s=1,…,Ss = 1, \ldots, S do

  4. Predict support quantiles Y^Q←fθ0,ϕ(X)\hat{Y}^{Q} \leftarrow f_{\theta_0,\phi}(X)

  5. Obtain frozen reference quantiles Y^0Q←sg⁡[fθ0(X)]\hat{Y}^{Q}_{0} \leftarrow\operatorname{sg}[f_{\theta_0}(X)]

  6. Compute LFreAL\mathcal{L}_{\mathrm{FreAL}} using Eq. (15)

  7. Compute Lanchor\mathcal{L}_{\mathrm{anchor}} from Y^Q\hat{Y}^{Q} and Y^0Q\hat{Y}^{Q}_{0} using Eq. (16)

  8. Ladapt←LFreAL+λanchorLanchor\mathcal{L}_{\mathrm{adapt}} \leftarrow\mathcal{L}_{\mathrm{FreAL}} + \lambda_{\mathrm{anchor}}\mathcal{L}_{\mathrm{anchor}}

  9. Update only ϕ\phi by minimizing Ladapt\mathcal{L}_{\mathrm{adapt}}

  10. end for

3. Deviation-Aware Fusion

  1. xctx←xT−L+1:Tx_{\mathrm{ctx}} \leftarrow x_{T-L+1:T}

  2. y^base←fθ0(xctx)\hat{y}_{\mathrm{base}} \leftarrow f_{\theta_0}(x_{\mathrm{ctx}})

  3. y^adapt←fθ0,ϕ(xctx)\hat{y}_{\mathrm{adapt}} \leftarrow f_{\theta_0,\phi}(x_{\mathrm{ctx}})

  4. Compute normalized forecast deviation Δ\Delta

  5. wbase←min⁡(λmax⁡,σmax⁡(0,Δ−τfus))w_{\mathrm{base}} \leftarrow\min(\lambda_{\max}, \sigma\max(0, \Delta-\tau_{\mathrm{fus}}))

  6. y^final←(1−wbase)y^adapt+wbasey^base\hat{y}_{\mathrm{final}} \leftarrow(1-w_{\mathrm{base}})\hat{y}_{\mathrm{adapt}} + w_{\mathrm{base}}\hat{y}_{\mathrm{base}}

  7. return y^final\hat{y}_{\mathrm{final}}

Algorithm 1 summarizes SWIFT for an individual forecasting instance. Using only the observed history, SWIFT constructs an intrinsic support set, updates the LoRA parameters while freezing the pretrained backbone, and combines the adapted and zero-shot forecasts through deviation-aware fusion. The LoRA updates minimize FreAL together with the prediction-space anchor regularizer in Eq. (17). The default setting uses 15 retrieved pairs, five recent pairs, and 15 adaptation steps. All support targets are observed before the requested forecast.

Experiments

Experimental Setup

Datasets and forecasting tasks. To support the construction of historical context–target pairs for instance-wise adaptation, we restrict our primary evaluation to tasks with available history length T≥2048T \ge2048. Under this criterion, we select six long-history tasks from Chronos Benchmark II [3]: Traffic, Electricity, Exchange, ETTm, ETTh, and ERCOT. These tasks cover transportation, electricity demand, transformer measurements, and exchange rates, with sampling frequencies ranging from 15 minutes to one business day. Here, Electricity refers to the Australian Electricity dataset, while ETTm and ETTh denote the 15-minute and hourly ETT tasks, respectively. We use the benchmark-specific prediction horizons: 48 steps for Electricity, 30 steps for Exchange, and 24 steps for Traffic, ETTm, ETTh, and ERCOT. Within each task, all compared methods are evaluated on the same forecasting targets and prediction horizons.

Evaluation metrics. We report Mean Absolute Scaled Error (MASE) and Weighted Quantile Loss (WQL) to assess point forecasting accuracy and probabilistic forecasting quality, respectively. Lower values indicate better performance for both metrics.

For a time series with observed history y1:Ty_{1:T} and forecast horizon HH, MASE is defined as

MASE⁡=1H∑h=1H∣yT+h−y^T+h∣1T−m∑t=m+1T∣yt−yt−m∣(21)\operatorname{MASE} = \frac{\frac{1}{H}\sum_{h=1}^{H}\lvert y_{T+h}-\hat{y}_{T+h}\rvert}{\frac{1}{T-m}\sum_{t=m+1}^{T}\lvert y_t-y_{t-m}\rvert} \tag*{(21)}

where y^T+h\hat{y}_{T+h} denotes the point forecast and mm is the seasonal period specified by the benchmark. The denominator scales the forecast error by the in-sample mean absolute error of a seasonal naive predictor.

Probabilistic forecasts are evaluated at nine quantile levels, Q={0.1,0.2,…,0.9}Q=\{0.1,0.2,\ldots,0.9\}. Let y^T+h(q)\hat{y}_{T+h}^{(q)} denote the predicted qq-quantile. The pinball loss at quantile level qq is

ℓq(y,y^)={q(y−y^),y≥y^,(1−q)(y^−y),y<y^.(22)\ell_q(y,\hat{y}) = \begin{cases} q(y-\hat{y}), & y \ge\hat{y},\\ (1-q)(\hat{y}-y), & y < \hat{y}. \end{cases} \tag*{(22)}

WQL averages the normalized quantile losses over QQ:

WQL⁡=1∣Q∣∑q∈Q2∑h=1Hℓq(yT+h,y^T+h(q))∑h=1H∣yT+h∣.(23)\operatorname{WQL} = \frac{1}{\lvert Q\rvert}\sum_{q\in Q}\frac{2\sum_{h=1}^{H}\ell_q\left(y_{T+h},\hat{y}_{T+h}^{(q)}\right)}{\sum_{h=1}^{H}\lvert y_{T+h}\rvert}. \tag*{(23)}

The above expressions describe a single forecast window; dataset-level errors are computed following the benchmark’s aggregation protocol.

For each metric, we first compute the relative error reduction on each dataset and then average these reductions across datasets. Specifically, the improvement over a reference method on dataset dd is

Imp.⁡d=Edref−EdSWIFTEdref×100%,(24)\operatorname{Imp.}_d = \frac{E_d^{\mathrm{ref}}-E_d^{\mathrm{SWIFT}}}{E_d^{\mathrm{ref}}}\times100\%, \tag*{(24)}

where EdE_d denotes the corresponding dataset-level forecasting error. The overall improvement across DD datasets is

Imp.⁡‾=1D∑d=1DImp.⁡d.(25)\overline{\operatorname{Imp.}} = \frac{1}{D}\sum_{d=1}^{D}\operatorname{Imp.}_d. \tag*{(25)}

This aggregation assigns equal weight to each dataset rather than weighting results by the number of constituent series.

To complement average relative error reductions, we report pairwise win rates across the 24 backbone–dataset configurations. For metric k∈{MASE,WQL}k\in\{\mathrm{MASE},\mathrm{WQL}\}, the win rate of method AA against method BB is

WR⁡k(A,B)=124∑c=124I[Ec,kA<Ec,kB]×100%,(26)\operatorname{WR}_k(A,B)=\frac{1}{24}\sum_{c=1}^{24}\mathbb{I}\left[E_{c,k}^{A}<E_{c,k}^{B}\right]\times100\%, \tag*{(26)}

where Ec,kAE_{c,k}^{A} denotes the error of method A under configuration cc, and I[⋅]\mathbb{I}[\cdot] is the indicator function. A win requires a strictly lower error. We also report a pooled win rate over the 48 backbone–dataset–metric comparisons:

WRpooled(A,B)=WRMASE(A,B)+WRWQL(A,B)2.(27)\mathrm{WR}_{\mathrm{pooled}}(A,B)=\frac{\mathrm{WR}_{\mathrm{MASE}}(A,B)+\mathrm{WR}_{\mathrm{WQL}}(A,B)}{2}. \tag*{(27)}

This pooled rate counts the two metrics separately; it does not measure the fraction of configurations in which both metrics improve simultaneously.

Backbone models and inference setting. We evaluate SWIFT on four TSFM backbones: Chronos-2 [2], TimesFM-2.5, Moirai-2 [23], and TiRex [5]. The combination of four backbones and six datasets yields 24 backbone–dataset configurations. Evaluation instances are processed independently, without using other series as contextual demonstrations. In particular, we set predict_batches_jointly=False for Chronos-2. This setting separates the effects of parameter adaptation from those of cross-sample contextual information.

Baseline adaptation protocols. For each backbone, we compare SWIFT with zero-shot inference, few-shot fine-tuning, and test-time adaptation (TTA). The zero-shot baseline directly applies the pretrained model without parameter updates. The few-shot fine-tuning baseline performs dataset-level LoRA adaptation using the full training split of the target dataset, without an additional external support set. Training examples are extracted independently from each univariate series using a context length of 2048, the dataset-specific prediction horizon, and a sliding-window stride of 10. Each model is fine-tuned for one epoch and subsequently evaluated with independent per-instance inference. For TTA, we implement the method of Kim et al. [19] on all four backbones. The pretrained forecaster remains frozen, while a lightweight Gated Calibration Module (GCM) calibrates its inputs and outputs. Before forecasting each benchmark target window, the GCM undergoes ten gradient updates on ten preceding pseudo-windows with a unit sliding stride. Each update minimizes mean squared error over the observed portion of the corresponding pseudo-window target, using a learning rate of 10−510^{-5} and weight decay of 10−410^{-4}. The calibrated model then generates the complete target horizon in one forecasting operation. The ground truth of the final target window is reserved exclusively for evaluation.

These baselines represent different adaptation regimes: dataset-level fine-tuning, rolling online calibration, and instance-wise pre-inference adaptation. Their adaptation data sources and update budgets therefore differ; the comparison evaluates these protocols under their respective information requirements.

Implementation and runtime measurement. Experiments are conducted on a single NVIDIA RTX A6000 GPU with 48 GB of memory. The implementation uses PyTorch and Hugging Face Transformers, with the random seed fixed to 123. Runtime measurements exclude model loading and dataset downloading. For few-shot fine-tuning, the measured computation includes dataset-level LoRA training and subsequent forecasting; for TTA, it includes the ten GCM updates and target forecasting. For SWIFT, it includes support-set construction, 15 LoRA updates, generation of the frozen and adapted forecasts, and deviation-aware fusion. These measurements quantify computation after the required data are available and exclude data-collection time and the waiting time for online observations.

SWIFT across Diverse TSFM Backbones

As shown in Table 2, SWIFT achieves positive average improvements over zero-shot inference on all four backbones. The dataset-wise relative reductions in MASE and WQL average

MethodERCOT MASEERCOT WQLElectricity MASEElectricity WQLExchange MASEExchange WQLETTm MASEETTm WQLETTh MASEETTh WQLTraffic MASETraffic WQL
Chronos-20.62750.03071.82930.01170.58130.04560.75990.07360.83880.02620.81890.2416
+Few-shot0.55860.02841.81450.01150.61660.05050.77290.07620.71540.02450.84410.2558
+TTA0.55320.02941.62730.01090.65210.06610.72190.06610.76090.02580.81820.2291
+SWIFT0.54590.02781.59380.00940.57330.04320.71630.06860.79220.02600.82460.2415
TimesFM-2.50.91270.04141.91230.01190.59220.05320.82570.08140.83810.02840.79430.2320
+Few-shot0.78940.04371.82820.01460.65240.07050.83010.10780.80170.03240.79220.2737
+TTA0.76340.04401.53180.01080.59460.05520.74070.06820.77780.02860.85380.2711
+SWIFT0.80380.03811.35160.01060.58020.05060.80510.07840.76410.02580.79530.2329
Moirai-20.56910.02891.78890.01090.63100.05050.75840.07370.56060.01750.82140.2509
+Few-shot0.59920.03721.77580.01450.64080.06900.77840.09960.63110.01940.85060.2908
+TTA0.52750.02901.45450.01160.58280.05830.75270.07000.82140.02650.77350.2483
+SWIFT0.56920.02871.38800.01080.63090.04970.74230.07280.55960.01710.78840.2409
TiRex0.98660.05382.38440.01620.63800.05400.80300.08070.67790.02000.78360.2310
+Few-shot0.89620.04251.81310.01600.61300.06390.77860.09400.74400.02170.79790.2751
+TTA0.99230.05441.82810.01540.79640.09920.78540.07940.60950.02200.69340.2741
+SWIFT0.98080.05241.74810.01460.63040.05290.77160.07860.67670.01890.78230.2209

Table 2. SWIFT across diverse TSFM backbones. Forecasting performance of zero-shot inference, few-shot fine-tuning, TTA, and SWIFT on the long-history subset of Chronos Benchmark II. The backbone-name rows report zero-shot results, while the remaining rows report the corresponding adaptation methods. Lower values indicate better performance. The best and second-best results within each backbone are marked in bold and underline, respectively.

6.31% and 6.99% for Chronos-2, 9.08% and 6.04% for TimesFM-2.5, 4.79% and 1.78% for Moirai-2, and 5.45% and 4.50% for TiRex, respectively. Across all 24 backbone–dataset configurations, the corresponding average reductions are approximately 6.41% and 4.83%.

Comparison with adaptation baselines. Figure 3 shows that SWIFT outperforms few-shot fine-tuning in 41 of 48 comparisons, with win rates of 79.2% for MASE and 91.7% for WQL, and TTA in 34 of 48 comparisons, with corresponding rates of 62.5% and 79.2%. Its advantage is therefore more pronounced for probabilistic forecasting. In comparison, TTA outperforms zero-shot inference in 75.0% of MASE comparisons but only 45.8% of WQL comparisons; the corresponding rates for few-shot fine-tuning are 50.0% and 20.8%. These results show that improvements in point forecasting do not necessarily translate into better probabilistic forecasts. The comparisons follow the distinct adaptation protocols specified in § 5.1.

Figure 3. Pairwise win rates across four TSFM backbones and six long-history datasets from Chronos Benchmark II. Each off-diagonal cell reports the percentage and count of comparisons in which the row method achieves a lower error than the column method. Panels (a) and (b) summarize 24 comparisons for MASE and WQL, respectively. Panel (c) pools the two metrics into 48 comparisons, rather than requiring simultaneous improvements in both metrics. Darker colors indicate higher win rates. Diagonal entries are omitted.

Overall comparison. SWIFT achieves the uniquely lowest error among all four methods in 29 of 48 comparisons (60.4%), including 12 MASE and 17 WQL results. Relative to zero-shot inference, it improves MASE in 21 of 24 configurations (87.5%) and WQL in 23 of 24 configurations (95.8%). The four degradations occur in Moirai-2 ERCOT MASE, Chronos-2 Traffic MASE, and TimesFM-2.5 Traffic MASE and WQL, indicating that the benefits of adaptation vary across backbones, datasets, and metrics.

SWIFT across Forecasting Benchmarks

We further evaluate SWIFT with Chronos-2 on Chronos Benchmark II, fev-bench [38], and GIFT-Eval [1]. The first row of Table 3 summarizes the Chronos-2 results from Table 2, while the remaining rows extend the evaluation to two additional benchmark collections.

Benchmark Subset# TasksWQLMASE
Chronos Benchmark II6+6.99%+6.31%
fev-bench46+1.66%+2.14%
GIFT-Eval24+1.23%+1.25%

Table 3. SWIFT across forecasting benchmarks. Average relative improvements over Chronos-2 zero-shot inference on history-qualified tasks satisfying T≥2048T \ge2048.

SWIFT achieves positive average improvements on all three benchmark subsets. The largest gains are observed on Chronos Benchmark II, with WQL and MASE reductions of 6.99% and 6.31%, respectively. It also improves WQL and MASE by 1.66% and 2.14% on fev-bench, and by 1.23% and 1.25% on GIFT-Eval.

fev-bench covers diverse real-world forecasting applications, including tasks with covariates, while GIFT-Eval evaluates models across varying domains, sampling frequencies, numbers of variates, and prediction lengths [38, 1]. Despite these heterogeneous and challenging settings, SWIFT consistently reduces both WQL and MASE, demonstrating the effectiveness of history-only adaptation across varied forecasting scenarios.

Overall, these results demonstrate that SWIFT generalizes across multiple forecasting benchmarks rather than being limited to a specific dataset collection.

Adaptation and Forecasting Latency

Table 4 reports the model-side execution time of few-shot fine-tuning, TTA, and SWIFT under their respective adaptation protocols. Few-shot fine-tuning includes dataset-level LoRA training and subsequent forecasting. TTA performs ten sequential updates of a lightweight Gated Calibration Module (GCM) using preceding pseudo-windows, followed by a single forecasting operation that produces the complete target horizon. SWIFT includes intrinsic support-set construction, 15 LoRA adaptation steps, generation of the frozen and adapted forecasts, and deviation-aware fusion. Across the 24 backbone–dataset combinations, SWIFT averages 6.835 seconds, compared with 51.405 seconds for few-shot fine-tuning and 6.365 seconds for TTA. These results show that SWIFT requires less computation time than the evaluated dataset-level fine-tuning protocol, while its average runtime is close to that of TTA.

All timings measure computation after the required data are available and exclude model loading and dataset downloading. TTA updates use only the observed portions of preceding pseudo-window targets; the final target-window ground truth is reserved exclusively for evaluation. The unit stride specifies the movement between adaptation windows, rather than autoregressive prediction within the target horizon. The ten sequential calibration updates introduce computational overhead, which is included in the reported TTA runtime.

TTA may additionally incur waiting time when the observations needed for successive calibration updates have not yet arrived. This delay depends on the sampling interval and observation availability and is not included in Table 4. When all observations required for the preceding pseudo-window updates are already available at the forecast origin, no additional observation-waiting delay is incurred. SWIFT performs adaptation directly from the history available at the current forecast origin, incurring retrieval and optimization costs without requiring subsequent observations. Likewise, any time needed to collect or curate the training data for few-shot fine-tuning is excluded from the measurements. Because the methods differ in adaptation data, trainable components, and update budgets, the reported results characterize the computational costs of their respective protocols rather than a matched-budget efficiency comparison.

BackboneMethodTrafficElectricityExchangeETTmETThERCOTAvg.
Chronos-2Few-shot3.70855.7321.43716.2623.72236.76619.604
TTA0.6613.0411.6091.1531.0071.7481.537
SWIFT2.7485.7504.7122.2362.9534.9963.899
TimesFM-2.5Few-shot9.333135.6683.29340.1189.02290.87448.051
TTA1.4039.6476.0814.4314.5636.5355.443
SWIFT1.3678.3953.8114.4874.4208.1385.103
Moirai-2Few-shot1.48522.2270.5506.5611.51314.7557.849
TTA0.2091.5180.6230.5110.3370.9570.693
SWIFT0.9110.2892.5020.6661.5470.6721.098
TiRexFew-shot9.064627.1853.35042.2489.18289.668130.116
TTA4.04451.54816.2099.2859.26116.38917.789
SWIFT6.95211.59522.18822.55722.66717.47617.239

Table 4. Adaptation and forecasting latency across TSFM backbones. Model-side execution time in seconds for few-shot fine-tuning, TTA, and SWIFT on the long-history subset of Chronos Benchmark II. Few-shot latency includes dataset-level LoRA fine-tuning and subsequent forecasting. TTA latency includes ten sequential GCM updates on preceding pseudo-windows and one complete target-window forecast. SWIFT latency includes support-set construction, 15 LoRA updates, base and adapted forecasting, and prediction fusion. Model loading, dataset downloading, and data-availability waiting time are excluded. All results are rounded to three decimal places.

Ablation Studies

We evaluate the contribution of individual SWIFT components using Chronos-2 on the six long-history datasets of Chronos Benchmark II. Table 5 compares Full SWIFT with variants that remove the anchor term, frequency-domain objective, prediction fusion, self-retrieval, or recent-window augmentation. The zero-shot model serves as the reference, with its relative improvements set to zero.

SettingMASE Imp.WQL Imp.
Zero-shot0.00%0.00%
Full SWIFT+6.31%+6.99%
w/o Anchor+6.00%+6.19%
w/o Frequency+4.40%+5.04%
w/o Fusion+6.32%+5.90%
w/o RAG+2.05%+2.38%
w/o Recent+5.12%+4.92%

Table 5. Component ablation of SWIFT on Chronos Benchmark II. MASE and WQL improvements are measured relative to Chronos-2 zero-shot inference. All adapted variants use 20 adaptation pairs and 15 gradient steps, with all non-ablated settings held fixed. Dataset-wise relative improvements are computed first and then averaged across the six long-history datasets. The best result for each metric is marked in bold.

All adapted variants are evaluated under strictly matched settings, except for the component being ablated. Specifically, they use the same pretrained checkpoint, evaluation instances, available historical observations, context lengths, and prediction horizons. Each variant uses 20 adaptation pairs, a batch size of 20, and 15 gradient update steps. The optimizer, learning rate,

LoRA configuration, initialization procedure, and random seed are also kept unchanged. For support-set ablations, the sample composition changes while the total number of adaptation pairs remains fixed. All non-ablated loss and fusion settings retain their Full SWIFT values, without variant-specific hyperparameter tuning. Relative improvements are computed separately for each dataset and then averaged across the six datasets. From the results, we make the following observations.

  • Self-retrieval provides the largest contribution among the evaluated component ablations. Removing self-retrieved historical windows reduces the average MASE improvement from 6.31% to 2.05% and the WQL improvement from 6.99% to 2.38%, corresponding to decreases of 4.26 and 4.61 percentage points, respectively. The comparison supports the value of selecting historical patterns relevant to the current forecasting query. Nevertheless, the variant without retrieval still improves over zero-shot inference, indicating that adaptation remains useful even without the full historical selection mechanism.

  • Recent-window augmentation complements self-retrieved historical patterns. Removing recent windows reduces the MASE and WQL improvements to 5.12% and 4.92%, respectively, compared with 6.31% and 6.99% for Full SWIFT. These differences amount to 1.19 percentage points for MASE and 2.07 percentage points for WQL. Historical retrieval retains substantial benefits on its own, while incorporating recent observations further improves adaptation. This result is consistent with the dual-focus design: retrieved windows supply recurring historical patterns, whereas recent windows provide observations closer to the current forecast origin.

  • Frequency-aware supervision improves both point and probabilistic forecasting. Without the frequency-domain objective, the MASE improvement decreases from 6.31% to 4.40%, while the WQL improvement decreases from 6.99% to 5.04%. The corresponding reductions of 1.91 and 1.95 percentage points indicate that the frequency-domain term provides benefits beyond the remaining adaptation objective. These results support combining time-domain supervision with spectral supervision when adapting the model using a limited historical support set.

  • The anchor term provides a modest additional benefit. Removing the anchor term yields improvements of 6.00% in MASE and 6.19% in WQL, below the Full SWIFT results by

0.31 and 0.80 percentage points, respectively. It improves both aggregate metrics in this evaluation. This suggests that the anchor constraint complements the adaptation objective, although these aggregate results alone do not establish its effect on every individual dataset.

  • Deviation-aware fusion benefits WQL without improving average MASE. Returning the adapted forecast directly produces a MASE improvement of 6.32%, slightly higher than the 6.31% obtained by Full SWIFT. However, its WQL improvement decreases from 6.99% to 5.90%. Thus, fusion improves the aggregate WQL result by 1.09 percentage points while introducing a negligible 0.01-percentage-point reduction in MASE improvement. This comparison supports fusion as a mechanism for improving probabilistic forecast quality, rather than as a component that necessarily improves both metrics.

Overall, Full SWIFT achieves the highest average WQL improvement and a nearly identical MASE improvement to the best ablated variant. The matched-budget comparisons support the complementary roles of historical retrieval, recent-window augmentation, frequency-aware supervision, and the anchor term. Prediction fusion provides an additional benefit for WQL, while its effect on average MASE is minimal. These findings support the combined framework without implying that every component improves every metric.

Conclusion

In this work, we challenged the prevailing dataset-centric adaptation paradigm for Time Series Foundation Models by proposing SWIFT (Sample-Wise Instance Fine-Tuning), a novel approach that treats individual series as self-contained datasets to bridge the generalization-specialization gap without external corpora. SWIFT combines Self-Retrieval, Frequency-Aware Loss, and Deviation-Aware Dynamic Fusion to construct informative supervision, capture instance-specific dynamics, and mitigate over-adaptation. Experiments across four TSFM backbones and three forecasting benchmarks demonstrate consistent improvements over zero-shot inference and competitive performance against few-shot fine-tuning and TTA without requiring external support data or online test-stream observations. Future work will focus on reducing the computational cost of instance-wise adaptation.

Acknowledgement

This work is supported in part by the National Natural Science Foundation of China (62576160, 62192783) and Nanjing University-Siemens Joint Research Center for Industrial AI.

References

  1. [1]Aksu, T., Woo, G., Liu, J., Liu, X., Liu, C., Savarese, S., Xiong, C., Sahoo, D., 2024. Gift-eval: A benchmark for general time series forecasting model evaluation. arXiv preprint arXiv:2410.10393.arxiv.org/abs/2410.10393
  2. [2]Ansari, A.F., Shchur, O., Küken, J., Auer, A., Han, B., Mercado, P., Rangapuram, S.S., Shen, H., Stella, L., Zhang, X., et al., 2025. Chronos-2: From univariate to universal forecasting. arXiv preprint arXiv:2510.15821.arxiv.org/abs/2510.15821
  3. [3]Ansari, A.F., Stella, L., Turkmen, C., Zhang, X., Mercado, P., Shen, H., Schchur, O., Rangapuram, S.S., Arango, S.P., Kapoor, S., et al., 2024. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815 .arxiv.org/abs/2403.07815
  4. [4]Assimakopoulos, V., Nikolopoulos, K., 2000. The theta model: a decomposition approach to forecasting. International journal of forecasting 16, 521–530.DOI
  5. [5]Auer, A., Podest, P., Klotz, D., Böck, S., Klambauer, G., Hochreiter, S., 2025. Tirex: Zero-shot forecasting across long and short horizons with enhanced in-context learning. URL: https://arxiv.org/abs/2505.23719, arXiv:2505.23719.
  6. [6]Box, G.E., Jenkins, G.M., 1968. Some recent advances in forecasting and control. Journal of the Royal Statistical Society. Series C (Applied Statistics) 17, 91–109.
  7. [7]Chen, Y., Liu, S., Yang, J., Jing, H., Zhao, W., Yang, G., 2024. A joint time-frequency domain transformer for multivariate time series forecasting. Neural Networks 176, 106334.DOI
  8. [8]Dange, A., Sarawagi, S., 2025. Tfmadapter: Lightweight instance-level adaptation of foundation models for forecasting with covariates, in: Proceedings of the 34th ACM International Conference on Information and Knowledge Management, pp. 498–507.DOI
  9. [9]Das, A., Faw, M., Sen, R., Zhou, Y., 2024a. In-context fine-tuning for time-series foundation models. arXiv preprint arXiv:2410.24087 .arxiv.org/abs/2410.24087
  10. [10]Das, A., Kong, W., Sen, R., Zhou, Y., 2024b. A decoder-only foundation model for time-series forecasting, in: Forty-first International Conference on Machine Learning.arxiv.org/abs/2310.10688
  11. [11]Feng, C., Huang, L., Krompass, D., 2024. General time transformer: an encoder-only foundation model for zero-shot multivariate time series forecasting, in: Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pp. 3757–3761.DOI
  12. [12]Gao, S., Koker, T., Queen, O., Hartvigsen, T., Tsiligkaridis, T., Zitnik, M., 2024. Units: Building a unified time series model. CoRR .arxiv.org/abs/2403.00131
  13. [13]Garza, A., Challu, C., Mergenthaler-Canseco, M., 2023. Timegpt-1. arXiv preprint arXiv:2310.03589 .arxiv.org/abs/2310.03589
  14. [14]He, H., Queen, O., Koker, T., Cuevas, C., Tsiligkaridis, T., Zitnik, M., 2023. Domain adaptation for time series under feature and label shifts, in: International conference on machine learning, PMLR. pp. 12746–12744.arxiv.org/abs/2302.03133
  15. [15]Hyndman, R.J., Athanasopoulos, G., 2018. Forecasting: principles and practice. OTexts.
  16. [16]Hyndman, R.J., Koehler, A.B., Snyder, R.D., Grose, S., 2002. A state space framework for automatic forecasting using exponential smoothing methods. International Journal of forecasting 18, 439–454.DOI
  17. [17]Imtiaz, S., Horchidan, S.F., Abbas, Z., Arsalan, M., Chaudhry, H.N., Vlassov, V., 2020. Privacy preserving time-series forecasting of user health data streams, in: 2020 IEEE International Conference on Big Data (Big Data), IEEE. pp. 3428–3437.DOI
  18. [18]Khwaja, E., Lettieri, C., Woo, G., Belouadah, E., Cenac, M., Jarry, G., Paquin, E., Zhao, X., Zhukov, V., Abou-Amal, O., et al., 2026. Toto 2.0: Time series forecasting enters the scaling era. arXiv preprint arXiv:2605.20119 .arxiv.org/abs/2605.20119
  19. [19]Kim, H., Kim, S., Mok, J., Yoon, S., 2025. Battling the non-stationarity in time series forecasting via test-time adaptation, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 17868–17876.arxiv.org/abs/2501.04970
  20. [20]Lee, T.L., Toner, W., Singh, R., Joosen, A., Asenov, M., 2025. Lightweight online adaptation for time series foundation model forecasts. arXiv preprint arXiv:2502.12920 .arxiv.org/abs/2502.12920
  21. [21]Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.t., Rocktäschel, T., et al., 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems 33, 9459–9474.arxiv.org/abs/2005.11401
  22. [22]Liang, J., He, R., Tan, T., 2025. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision 133, 31–64.DOI
  23. [23]Liu, C., Aksu, T., Liu, J., Liu, X., Yan, H., Pham, Q., Savarese, S., Sahoo, D., Xiong, C., Li, J., 2025. Moirai 2.0: When less is more for time series forecasting. arXiv preprint arXiv:2511.11698 .arxiv.org/abs/2511.11698
  24. [24]Liu, L., Zhou, B., Zhao, Z., Liu, Z., 2024a. Active dynamic weighting for multi-domain adaptation. Neural Networks 177, 106398.DOI
  25. [25]Liu, Q., Liu, X., Liu, C., Wen, Q., Liang, Y., 2024b. Time-ffm: Towards lm-empowered federated foundation model for time series forecasting. Advances in Neural Information Processing Systems 37, 94512–94538.arxiv.org/abs/2405.14252
  26. [26]Liu, Y., Hu, T., Zhang, H., Wu, H., Wang, S., Ma, L., Long, M., 2023. itransformer: Inverted transformers are effective for time series forecasting. ArXiv abs/2310.06625. URL: https://api.semanticscholar.org/CorpusID:263830644.arxiv.org/abs/2310.06625
  27. [27]Liu, Y., Wu, H., Wang, J., Long, M., 2022. Non-stationary transformers: Exploring the stationarity in time series forecasting. Advances in neural information processing systems 35, 9881–9893.arxiv.org/abs/2205.14415
  28. [28]Luo, D., Wang, X., 2024. Moderntcn: A modern pure convolution structure for general time series analysis, in: The twelfth international conference on learning representations, pp. 1–43.
  29. [29]Mienye, I.D., Swart, T.G., Obaido, G., 2024. Recurrent neural networks: A comprehensive review of architectures, variants, and applications. Information 15, 517.DOI
  30. [30]Ning, K., Pan, Z., Liu, Y., Jiang, Y., Zhang, J.Y., Rasul, K., Schneider, A., Ma, L., Nevmyvaka, Y., Song, D., 2025. T5-rag: Retrieval-augmented generation based time series foundation models are stronger zero-shot forecaster. arXiv preprint arXiv:2503.07649 .arxiv.org/abs/2503.07649
  31. [31]Podest, P., Pichler, M., Bürger, E., Zólyomi, L., Voggenberger, B., Berghammer, W., Klotz, D., Böck, S., Klambauer, G., Hochreiter, S., 2026. Tirex-2: Generalizing tirex to multivariate data and streaming. arXiv preprint arXiv:2607.01204 .arxiv.org/abs/2607.01204
  32. [32]Qiao, Z., Liu, C., Zhang, Y., Jin, M., Pham, Q., Wen, Q., Suganthan, P., Jiang, X., Ramasamy, S., 2026. Multi-scale finetuning for encoder-based time series foundation models. Advances in Neural Information Processing Systems 38, 22313–22345.arxiv.org/abs/2506.14087
  33. [33]Rasul, K., Ashok, A., Williams, A.R., Khorasani, A., Adamopoulos, G., Bhagwatkar, R., Biloš, M., Ghonia, H., Hassen, N., Schneider, A., et al., 2023. Lag-llama: Towards foundation models for time series forecasting, in: R0-FoMo: Robustness of Few-shot and Zero-shot Learning in Large Foundation Models.arxiv.org/abs/2310.08278
  34. [34]Ravuru, C., Sakhinana, S.S., Runkana, V., 2024. Agentic retrieval-augmented generation for time series analysis. arXiv preprint arXiv:2408.14484 .arxiv.org/abs/2408.14484
  35. [35]Salinas, D., Flunkert, V., Gasthaus, J., Januschowski, T., 2020. DeepAR: Probabilistic forecasting with autoregressive recurrent networks. International journal of forecasting 36, 1181–1191.arxiv.org/abs/1704.04110
  36. [36]Schulman, J., Levine, S., Abbeel, P., Jordan, M., Moritz, P., 2015. Trust region policy optimization, in: International conference on machine learning, PMLR. pp. 1889–1897.arxiv.org/abs/1502.05477
  37. [37]Schulman, J., Wolski, F., Dhariwal, P., Radford, A., Klimov, O., 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 .arxiv.org/abs/1707.06347
  38. [38]Shchur, O., Ansari, A.F., Turkmen, C., Stella, L., Erickson, N., Guerron, P., Bohlke-Schneider, M., Wang, Y., 2025. fev-bench: A realistic benchmark for time series forecasting. arXiv preprint arXiv:2509.26468 .arxiv.org/abs/2509.26468
  39. [39]Shi, X., Wang, S., Nie, Y., Li, D., Ye, Z., Wen, Q., Jin, M., 2024. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040 .arxiv.org/abs/2409.16040
  40. [40]Sonmezer, M., Ertekin, S., 2025. Canet: Chronoadaptive network for enhanced long-term time series forecasting under non-stationarity. Neural Networks , 108495.DOI
  41. [41]Wan, J., Xia, N., Yin, Y., Pan, X., Hu, J., Yi, J., 2024. Tcdformer: A transformer framework for non-stationary time series forecasting based on trend and change-point detection. Neural Networks 173, 106196.DOI
  42. [42]Wang, H., Pan, L., Chen, Z., Yang, D., Zhang, S., Yang, Y., Liu, X., Li, H., Tao, D., 2024a. Fredf: Learning to forecast in the frequency domain. arXiv preprint arXiv:2402.02399 .arxiv.org/abs/2402.02399
  43. [43]Wang, S., Li, J., Shi, X., Ye, Z., Mo, B., Lin, W., Ju, S., Chu, Z., Jin, M., 2024b. Timenixer++: A general time series pattern machine for universal predictive analysis. arXiv preprint arXiv:2410.16032 .arxiv.org/abs/2410.16032
  44. [44]Wang, Z., Dai, Z., Póczos, B., Carbonell, J., 2019. Characterizing and avoiding negative transfer, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 11293–11302.arxiv.org/abs/1811.09751
  45. [45]Xie, S., Odonnat, A., Xiao, Z., Zan, L., Tiomoko, M., Pan, L., Palpanas, T., Oreshkin, B.N., Liu, C., Zhang, K., 2026. Post-training in time series foundation models: A unifying framework. URL: https://arxiv.org/abs/2607.20002, arXiv:2607.20002.
  46. [46]Yu, Y., Peeperkorn, J., De Smedt, J., De Weerdt, J., 2026. Time series foundation models for process model forecasting, in: Advanced Information Systems Engineering, Springer. pp. 78–97.
  47. [47]Zhang, H., Xu, C., Zhang, Y.F., Zhang, Z., Wang, L., Bian, J., 2025. Timeraf: Retrieval-augmented foundation model for zero-shot time series forecasting. IEEE Transactions on Knowledge and Data Engineering .arxiv.org/abs/2412.20810
  48. [48]Zhou, S., Schöner, H., Wu, Z., Fouché, E., Wilson, I., Wang, S., 2026. Stationarity-aware retrieval-augmented time series forecasting. arXiv preprint arXiv:2606.04135 .arxiv.org/abs/2606.04135
  49. [49]Zhou, T., Ma, Z., Wen, Q., Wang, X., Sun, L., Jin, R., 2022. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting, in: International Conference on Machine Learning. URL: https://api.semanticscholar.org/CorpusID:246430171.arxiv.org/abs/2201.12740
  50. [50]Zhou, T., Niu, P., Wang, X., Sun, L., Jin, R., 2023. One fits all: Power general time series analysis by pretrained lm, in: Neural Information Processing Systems. URL: https://api.semanticscholar.org/CorpusID:258741419.arxiv.org/abs/2302.11939

Paper details

vlnc:W_f2bc4j6q
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

Not recorded for this paper

View statement
Looking for related papers…
Contents