RACE: Residual-Aware Test-Time Adaptation for Neighbor-Rich Time-Series Foundation Model Forecasting

Hao-Nan Shi; Tong Wu; Chen-Cong Sun; Yuan Jiang; Han-Jia Ye; De-Chuan Zhan

cs.LGSep 30, 2026

RACE: Residual-Aware Test-Time Adaptation for Neighbor-Rich Time-Series Foundation Model Forecasting

Authors: Hao-Nan Shi, Tong Wu, Chen-Cong Sun, Yuan Jiang, Han-Jia Ye, De-Chuan Zhan

Organizations: School of Artificial Intelligence, Nanjing University · National Key Laboratory for Novel Software Technology, Nanjing University · Nanjing, China

Abstract

Time-series foundation models (TSFMs) perform strongly across forecasting tasks, but their per-series inference is ill-suited to neighbor-rich forecasting, where each query has access to related but nonidentical historical series. Continuous glucose monitoring (CGM) and Web/cloud workloads exemplify this setting: CGM trajectories share physiological patterns but vary across individuals, devices, and conditions, while Web/cloud workloads combine common operating regimes with non-stationarity, heavy tails, and bursts. These histories share useful structure, yet neighbors are not equally relevant. Existing methods either fine-tune TSFMs for each target domain, incurring additional costs and offering limited transferability across backbones, or append retrieved series without verifying whether they support the current forecast. The key challenges are conflicting residual evidence from neighboring series and residual patterns that vary across TSFMs and forecasting tasks. We formulate test-time neighborhood scaling: using same-domain neighbor evidence without modifying the backbone. We propose RACE (Residual-Aware Correction of Forecasting Errors), a two-stage framework for using historical neighbors. We first retrieve query-compatible neighbors, align their residuals to the query scale, and aggregate coherent evidence into the training-free RACE-TF correction. Full RACE then uses a lightweight, domain-specific Gate to determine when applying the correction is beneficial, with a reusable training workflow across TSFM backbones. Across four TSFMs, RACE improves all three domain-aggregate metrics on both primary domains, with the largest gains on high-error queries. Within each domain, a Gate trained on one TSFM transfers to other backbones without adaptation, and the resulting pipeline improves all 72 cross-backbone metric comparisons over the matched frozen targets.

Figures & tables

Figure 1: Neighbor-rich forecasting turns reusable domain histories into selective corrections. Each row uses outcome-observed windows and frozen-TSFM residuals from the same domain. Conflicting Web evidence is rejected, whereas coherent CGM neighbors improve the forecast.

Figure 2: The pipeline of RACE. A domain library supplies historical residual records; Residual Evidence Construction retrieves, aligns, and aggregates query-compatible neighbors, and the Utility-Aware Correction Gate decides whether to apply the proposal.

TSFM + correction	Method status	Web: MSE / MAE / CRPS	CGM: MSE / MAE / CRPS	Best Count
Chronos-2	frozen base	8.676e+05 / 70.35 / 51.27	2.925 / 1.149 / 0.9099	–
+ RAF	training-free	+2.503 / +2.423 / +0.216	+6.826 / +4.869 / +4.585	2
+ RACE-TF (ours)	training-free	+1.159 / +0.099 / -0.186	+20.752 / +10.531 / +9.849	0
+ RACE (ours)	fitted (XGB)	+0.982 / +1.091 / +1.182	+20.814 / +10.677 / +9.939	4
Chronos-Bolt Base	frozen base	9.277e+05 / 74.87 / 59.49	3.170 / 1.233 / 0.9708	–
+ RAF	training-free	+5.527 / +1.230 / +1.244	+6.923 / +4.904 / +4.394	1

Table 1: Main results across four frozen TSFMs. Base rows report raw errors; other rows report percentage improvements, with positive values indicating lower error. Best results are bolded. TS-RAG uses its supported Chronos-Bolt Base interface only. Comparisons use each backbone’s matched base and test windows (Appendix F).

Source Gate	Web/cloud					CGM (ShanghaiT2DM target)
	Chronos-2 Base	Chronos-Bolt Base	TimesFM 2.5 200M	Toto-2.0- 313m	Gate uplift	Chronos-2 Base	Chronos-Bolt Base	TimesFM 2.5 200M	Toto-2.0- 313m	Gate uplift
Chronos-2 Base	+1.09	+1.94	+1.98	+1.04	+0.74	+13.81	+15.46	+12.99	+14.68	+0.10
Chronos-Bolt Base	+0.88	+1.81	+1.21	+0.87	+0.35	+13.74	+15.46	+12.89	+14.68	+0.03
TimesFM 2.5 200M	+1.09	+1.93	+1.97	+1.02	+0.78	+13.63	+15.44	+12.95	+14.47	-0.04
Toto-2.0-313m	+1.06	+1.92	+1.98	+0.99	+0.67	+13.72	+15.44	+12.95	+14.54	+0.06

Table 2: Cross-backbone Gate transfer: equal-weight mean MSE/MAE/CRPS improvement (%) over matched frozen targets. Diagonal values are target-trained RACE references from Table 1 . Backbone variants follow Section 5.2 . Gate uplift excludes diagonal entries.

Panel A: evidence-construction ablations (Gate retained)
Variant	Change	Web (%)	CGM (%)
RACE	None	+1.410 / +1.431 / +1.557	+21.327 / +11.200 / +10.043
Global bias + Gate	Without local retrieval	-0.040 / -0.945 / -1.332	+3.802 / +1.320 / +1.369
History-only retrieval	Without forecast	+1.214 / +1.209 / +1.336	+16.484 / +8.943 / +8.021
Residual mean + Gate	Without consensus	+2.247 / +0.984 / +0.917	+21.716 / +10.154 / +9.086
Panel B: Gate ablation and combined control

Table 3: Mechanism ablation. History-only retrieval excludes the forecast; residual mean (Mean) averages unfiltered residuals. Global bias replaces local retrieval with a horizon-wise correction. Metrics and four-backbone averaging are defined in Section 6.3 .

Figure 3: Residual-evidence construction audit. The first two columns show calibration-time mean residual-medoid cosine and the share of residuals with negative cosine to the medoid. The third shows post-hoc test gains averaged over MSE, MAE, and CRPS for ungated GlobalBias, Retrieved Mean, and Consensus Mean (RACE-TF). Error bars show 95% target-parent bootstrap intervals.

Panel of Figure 4 — Figure 4: Selective residual-transfer audit for Chronos-Bolt Base. (a) Test-set precision versus acceptance rate; dashed lines mark acceptance rates at calibration-fixed thresholds. (b) Five within-dataset base-MSE bins: correction gains (top), acceptance and precision (bottom). Error bars show 95% target-parent bootstrap intervals.

Figure 5: Chronos-2 case studies: CGM success (top) and Web failure (bottom). Left: the last 256 steps of the query and top-neighbor histories, followed by the 64-step RACE forecast window. Right: RAF and Residual Mean (without consensus or Gate) on the same future. Panels share the time-step scale and each row’s vertical scale. Values are history-standardized; shading marks forecast changes. Percentages average MSE, MAE, and CRPS improvements.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Symbols	Meaning
Windows and forecasts $q,i;\ f_{\theta}$ $\mathbf{c}_{q},\mathbf{c}_{i};\ \mathbf{y}_{q},\mathbf{y}_{i}$ $\hat{\mathbf{y}}_{q},\hat{\mathbf{y}}_{i},\hat{\mathbf{y}}_{q}^{\mathrm{final}}$	$q$ is a query window and $i$ a historical window in $q$ ’s eligible residual pool. $f_{\theta}$ is the frozen time-series foundation model (TSFM); $\mathbf{c}_{q},\mathbf{c}_{i}$ are visible contexts; $\mathbf{y}_{q},\mathbf{y}_{i}$ are realized futures; and $\hat{\mathbf{y}}_{q},\hat{\mathbf{y}}_{i},\hat{\mathbf{y}}_{q}^{\mathrm{final}}$ are the base, neighbor, and deployed forecasts.
Legal retrieval $\mathcal{C}_{q},\mathcal{N}_{q};\ K,P$ $t_{i}^{\mathrm{end}},t_{q}^{\mathrm{forecast\ start}}$	$\mathcal{C}_{q}$ is the eligible residual pool and $\mathcal{N}_{q}$ its final subset of at most $K$ query-compatible neighbors. $P=600$ is the initial exact-ranking width; a short query is expanded exactly after legality filtering. A parent is an original full series; a same-parent candidate’s future must end no later than the query future starts: $t_{i}^{\mathrm{end}}\leq t_{q}^{\mathrm{forecast\ start}}$ .
Retrieval representations $\mathbf{z}_{q};\ \phi_{j},\operatorname{LN}_{j},w_{j}$ $\operatorname{concat},\lVert\cdot\rVert_{2},\langle\cdot,\cdot\rangle$ $\mathbf{e}_{q},\mathbf{e}_{i},\mathbf{e}_{q}^{\mathrm{HM}}$	$\mathbf{z}_{q}$ is the normalized visible history–forecast trajectory. For $j\in\{1,\ldots,6\}$ , $\phi_{j}$ , $\operatorname{LN}_{j}$ , and $w_{j}$ are feature extractor $j$ , its layer-normalization transform, and its fixed weight. $\operatorname{concat}$ concatenates vectors, $\lVert\cdot\rVert_{2}$ is the Euclidean norm, $\langle\cdot,\cdot\rangle$ is the Euclidean inner product, and $\mathbf{e}_{q},\mathbf{e}_{i}$ are query and neighbor representations; $\mathbf{e}_{q}^{\mathrm{HM}}$ is the HybridMath unit-norm representation.
Residual scale alignment $\mathbf{r}_{i};\ s_{i},s_{q}$ $\widetilde{\mathbf{r}}_{qi};\ \operatorname{clip}(\cdot,\tfrac{1}{4},4)$	$\mathbf{r}_{i}=\mathbf{y}_{i}-\hat{\mathbf{y}}_{i}$ is the realized neighbor residual. $s_{i},s_{q}$ are mean absolute observed-context magnitudes, floored at $10^{-6}$ ; $\operatorname{clip}(x,\tfrac{1}{4},4)$ clamps $x$ to $[\tfrac{1}{4},4]$ , and $\widetilde{\mathbf{r}}_{qi}$ is the dimensionless, clipped residual aligned to the query scale.
Evidence sets $\mathcal{V}_{q},\mathcal{V}_{qhc},\mathcal{H}_{qi}$ $\mathcal{I}_{q}^{(0)},\mathcal{I}_{qhc};\ h,c$	$\mathcal{V}_{q},\mathcal{V}_{qhc},\mathcal{H}_{qi}$ denote valid neighbors, coordinate-available neighbors, and usable coordinates. $\mathcal{I}_{q}^{(0)}$ and $\mathcal{I}_{qhc}$ are the initial and coordinate-specific retained sets; $(h,c)$ indexes a forecast-horizon coordinate and channel.
Evidence summaries and screening $\bar{r}_{qhc},d_{qihc},m_{qi},v_{qi}$ $\mu_{m},\mu_{v},S(\cdot);\ c_{\mathrm{MAD}},\mathbf{a},a_{j}$ $z^{(m)}_{qi},z^{(v)}_{qi},\gamma$	$\bar{r},d,m,v$ are the residual center, coordinate deviation, signed mean deviation, and within-neighbor standard deviation. $\mu_{m},\mu_{v},S$ are robust centers and median-absolute-deviation scales; $c_{\mathrm{MAD}}$ is the fixed MAD consistency coefficient; $\mathbf{a}$ is an across-neighbor scalar collection and $a_{j}$ one entry. $z^{(m)},z^{(v)}$ are MAD-standardized displacement and one-sided variability-excess scores; $\gamma$ is their fixed screening threshold.

Appendix

Table 4: Global mathematical notation used by RACE.

Frozen TSFM	MSE gain	MAE gain	CRPS gain	Acceptance rate
Chronos-2	+0.168	+0.052	-0.089	36.15%
Chronos-Bolt Base	+3.533	+1.409	+0.955	62.26%
TimesFM-2.5	+0.010	+0.086	+0.122	13.79%
Toto-2.0-313m	+4.724	+1.267	+0.971	41.29%
Four-TSFM mean	+2.109	+0.704	+0.490	38.37%

Appendix

Table 5: Broader GIFT-Eval diagnostic on 34 subsets satisfying the fixed-window requirements. Values are percentage reductions in raw-observed error; positive is better.

Source Gate	Target TSFM	MSE gain (%)	MAE gain (%)	CRPS gain (%)	Mean Gate uplift (pp)
Chronos-2	Chronos-Bolt Base	+1.633	+2.022	+2.164	+0.746
Chronos-2	TimesFM-2.5	+1.982	+1.767	+2.205	+0.578
Chronos-2	Toto-2.0-313m	+1.035	+1.101	+0.984	+0.888
Chronos-Bolt Base	Chronos-2	+1.019	+0.804	+0.829	+0.526
Chronos-Bolt Base	TimesFM-2.5	+1.189	+1.091	+1.339	-0.200
Chronos-Bolt Base	Toto-2.0-313m	+1.116	+0.849	+0.632	+0.714

Appendix

Table 6: Complete raw-z zero-shot Gate transfer on Web. Gains are relative to the target’s matched frozen TSFM; the final column is the equal-weight MSE/MAE/CRPS difference between the transferred Gate and that target’s RACE-TF proposal. Positive values are better.

Source Gate	Target TSFM	MSE gain (%)	MAE gain (%)	CRPS gain (%)	Mean Gate uplift (pp)
Chronos-2	Chronos-Bolt Base	+22.801	+12.895	+10.672	+0.129
Chronos-2	TimesFM-2.5	+19.862	+9.879	+9.244	+0.117
Chronos-2	Toto-2.0-313m	+22.235	+11.364	+10.426	+0.048
Chronos-Bolt Base	Chronos-2	+20.708	+10.609	+9.895	+0.027
Chronos-Bolt Base	TimesFM-2.5	+19.684	+9.818	+9.177	+0.015
Chronos-Bolt Base	Toto-2.0-313m	+22.228	+11.388	+10.418	+0.051

Appendix

Table 7: Complete raw-z zero-shot Gate transfer on CGM. Gains are relative to the target’s matched frozen TSFM; the final column is the equal-weight MSE/MAE/CRPS difference between the transferred Gate and that target’s RACE-TF proposal. Positive values are better.

Group	Configuration and changed factor	Web (%)	CGM (%)
		MSE / MAE / CRPS	MSE / MAE / CRPS
Deployed route	RACE: raw-z representation, Consensus Mean, $K=60$ , Gate with MSE target	+1.410 / +1.431 / +1.557	+21.327 / +11.200 / +10.043
Encoder sensitivity	HybridMath representation replaces raw-z	+1.807 / +1.256 / +1.337	+20.637 / +10.725 / +9.300
Query control	Chronos-T5 final-token representation; history-only query	-0.071 / -0.312 / -0.406	+6.330 / +2.145 / +1.815
Query control	History only; frozen forecast removed from raw-z representation	+1.214 / +1.209 / +1.336	+16.484 / +8.943 / +8.021
Neighborhood	$K=10$ replaces $K=60$	+1.486 / +1.035 / +1.052	+14.336 / +6.549 / +5.735

Appendix

Table 8: Complete primary mechanism ablations. These controls are diagnostic variants, not additional RACE variants.

Task	Frozen TSFM	RACE-TF (all queries)	CERR (direct learned)	RACE (Gate-selected)
		MSE / MAE / CRPS	MSE / MAE / CRPS	MSE / MAE / CRPS
Web	Chronos-2	+1.159 / +0.099 / -0.186	+1.445 / -8.179 / -12.516	+0.982 / +1.091 / +1.182
Web	Chronos-Bolt Base	+1.782 / +0.954 / +0.846	+4.290 / -3.513 / -4.963	+1.647 / +1.838 / +1.945
Web	TimesFM-2.5	+2.342 / +0.845 / +1.033	+2.826 / -4.014 / -4.782	+1.986 / +1.748 / +2.187
Web	Toto-2.0-313m	+1.301 / -0.126 / -0.719	-2.226 / -10.163 / -13.909	+1.025 / +1.047 / +0.912
Web	Four-TSFM mean	+1.646 / +0.443 / +0.243	+1.584 / -6.467 / -9.042	+1.410 / +1.431 / +1.557

Appendix

Table 9: Matched direct residual-regression control. All rows use the same raw-z context-plus-frozen-forecast representation, eligible residual pool derived from the domain library,

K=60

, evidence retained by mean-centred MAD-based residual consensus, raw-observed evaluator, test windows, and frozen TSFM. CERR replaces only the consensus-mean output interface with a learned residual regressor applied to every query; it has no Gate. Values are percentage reductions in error relative to the matched frozen base, ordered as MSE / MAE / CRPS; positive is better. Four-TSFM means are unweighted.

Group	Configuration and changed factor	Web (%)
		MSE / MAE / CRPS
Scope diagnostic	RACE with raw-z representation; same logical-subset eligible residual pool only	-0.277 / +0.378 / +0.288
Scope diagnostic	RACE with HybridMath representation; same logical-subset eligible residual pool only	+0.106 / +0.501 / +0.523
Scope diagnostic	RACE with raw-z representation; hard Top-3 subset-card eligible residual pool	-0.769 / +0.016 / -0.198
Scope diagnostic	RACE with HybridMath representation; hard Top-3 subset-card eligible residual pool	+0.938 / +1.004 / +1.056
Unconditional transfer	Same-subset mean, Web only; no consensus or Gate	-39.280 / -26.983 / -35.841

Appendix

Table 10: Web-only retrieval-scope diagnostics and restricted-pool controls. CGM is omitted because its external domain library and held-out target split differ from the Web protocol. These controls isolate the effect of restricting the eligible residual pool relative to the default full-domain construction. The default Gate uses MSE utility; the sMAPE-targeted Gate in Table 8 is a separate sensitivity control.

Task	Frozen TSFM	$\tau$	RACE-TF beneficial (%)	Acceptance rate (%)	Benefit among accepted (%)	AP	Spearman
CGM	Chronos-2	0.514	68.09	95.25	69.05	0.773	0.336
CGM	Chronos-Bolt Base	0.601	69.91	95.74	71.63	0.780	0.400
CGM	TimesFM-2.5	0.581	66.35	90.55	68.32	0.744	0.306
CGM	Toto-2.0-313m	0.644	69.42	90.67	70.96	0.787	0.368
Web	Chronos-2	0.453	37.85	52.02	52.69	0.548	0.520
Web	Chronos-Bolt Base	0.564	40.23	51.75	53.48	0.529	0.421

Appendix

Table 11: Complete four-backbone Figure 4 Gate audit. RACE-TF beneficial and Benefit among accepted are post-hoc fractions of test windows whose RACE-TF proposal lowers per-window MSE. AP is average precision for this beneficial-correction label; Spearman is the rank correlation between Gate probability and per-window RACE-TF MSE percentage gain. All diagnostics are computed after thresholds are frozen.

Frozen TSFM	Error bin	RACE MSE reduction (%)	RACE-TF MSE reduction (%)	Acceptance rate (%)	Benefit among accepted (%)	$n$
Chronos-2	1	-0.02	-30.99	34.48	36.34	4820
Chronos-2	2	-0.12	-60.11	42.13	50.89	4814
Chronos-2	3	-0.52	-51.59	49.74	53.86	4815
Chronos-2	4	+0.07	-26.84	67.30	57.13	4814
Chronos-2	5	+60.92	+42.34	66.48	56.96	4812
Chronos-Bolt Base	1	-0.50	-35.68	39.98	34.61	4820

Appendix

Table 12: Complete four-backbone Figure 4 error-stratified audit on Web. Error bins are within-dataset equal-count quintiles of frozen-base MSE (1=lowest, 5=highest). MSE reductions are dataset-macro averages after normalizing each dataset by its mean frozen-base MSE. Acceptance rate and Benefit among accepted are post-hoc test diagnostics;

n

is the number of windows in the bin.

Frozen TSFM	Error bin	RACE MSE reduction (%)	RACE-TF MSE reduction (%)	Acceptance rate (%)	Benefit among accepted (%)	$n$
Chronos-2	1	-1.74	-1.77	96.36	47.80	165
Chronos-2	2	+1.21	+1.02	94.51	65.16	164
Chronos-2	3	+7.20	+6.91	92.07	69.54	164
Chronos-2	4	+22.32	+21.96	98.17	80.75	164
Chronos-2	5	+75.22	+75.76	95.12	82.05	164
Chronos-Bolt Base	1	-4.08	-4.16	95.76	43.04	165

Appendix

Table 13: Complete four-backbone Figure 4 error-stratified audit on CGM. Error bins are within-dataset equal-count quintiles of frozen-base MSE (1=lowest, 5=highest). MSE reductions are dataset-macro averages after normalizing each dataset by its mean frozen-base MSE. Acceptance rate and Benefit among accepted are post-hoc test diagnostics;

n

is the number of windows in the bin.

Panel of Figure 6 — Figure 6: Selective residual-transfer audit extension: Chronos-2. The left and right panels reproduce the respective body diagnostic procedures without model, threshold, or test-time selection changes.

Panel of Figure 7 — Figure 7: Selective residual-transfer audit extension: TimesFM-2.5. Definitions, threshold provenance, and bootstrap units are identical to Figure 4 .

Panel of Figure 8 — Figure 8: Selective residual-transfer audit extension: Toto-2.0-313m. The panels retain the same task encodings and outcome definitions as all other supplementary extensions.

Method	MSE W/L	MAE W/L	CRPS W/L
RACE-TF	21/19	10/30	11/29
RACE	26/14	17/23	17/23
RAF	13/27	4/36	8/32

Appendix

Table 14: Web subset–backbone win/loss summary. For each metric, W and L count the 40 logical-subset–frozen-TSFM cells with, respectively, positive and non-positive matched-base percentage improvement. Thus W+L equals 40 for every method and metric. These frozen post-hoc counts do not select a route, model state, or Gate threshold.

Feature block (dimension)	Exact construction	Decision information
Evidence availability (2)	Consensus-retained valid-neighbor count and valid-neighbor fraction.	Amount of legal, consensus-retained evidence.
Aligned residual evidence (9)	Mean and standard deviation of values and of their absolute values, median, and root-mean-square over valid aligned residual coordinates; mean and standard deviation of within-neighbor residual standard deviations; and $1/(1+\operatorname{mean}\text{ within-neighbor standard deviation})$ .	Residual level, magnitude, variability, and temporal coherence.
Retrieval similarity (8)	Mean, standard deviation, minimum, maximum, 10th/50th/90th percentiles, and Gini coefficient of $1/(1+\lVert\mathbf{e}_{i}-\mathbf{e}_{q}\rVert_{2})$ over valid neighbors.	Closeness and concentration of retrieved support.
Frozen forecast shape (2)	Standard deviation and mean absolute first difference of the query-scale-normalized frozen forecast, $\operatorname{clip}(\hat{\mathbf{y}}_{q}/s_{q},-50,50)$ .	Volatility of the backbone forecast.
Observed context shape (2)	Standard deviation and mean absolute first difference of observed $\mathbf{c}_{q}/s_{q}$ .	Recent target dynamics in comparable units.
Consensus proposal (3)	Mean absolute value, standard deviation, and maximum absolute value of $\boldsymbol{\Delta}_{q}$ on $\mathcal{H}_{q}$ .	Size and temporal variation of the proposed correction.

Appendix

Table 15: Exact 30-dimensional, inference-visible schema for the Utility-Aware Correction Gate. The default target is positive full-correction MSE utility.

Method	Measurement boundary	Chronos-2	Chronos-Bolt Base	TimesFM-2.5	Toto-2.0-313m
		Web / CGM (s)	Web / CGM (s)	Web / CGM (s)	Web / CGM (s)
RAF	Encoding, Top-1 retrieval, augmented-input forecasting, evaluation, and audit output.	38.7 / 19.6	50.3 / 20.0	208.0 / 43.2	95.6 / 49.1
RACE-TF	Post-cache route replay; residual evidence construction and one final test pass; no fitted Gate.	91.0 / 8.7	93.2 / 8.4	92.8 / 8.5	93.7 / 8.7
RACE	Post-cache route replay; residual evidence construction, Gate fitting, model-select, calibration, and one final test pass.	156.3 / 19.5	156.7 / 19.5	157.2 / 19.6	170.8 / 20.7
TS-RAG (PT)	Frozen supplied ARM; Chronos-Bolt Base interface only.	–	134.2 / 9.2	–	–
TS-RAG (FT)	Fine-tuned ARM; Chronos-Bolt Base interface only.	–	814.0 / 178.7	–	–

Appendix

Table 16: Recorded runtime across four frozen backbones, in wall-clock seconds (Web / CGM), with measurement boundaries. RAF includes encoding, retrieval, and augmented-input backbone inference; RACE-TF and RACE use post-cache replays, with Gate fitting and selection included for RACE. TS-RAG uses its Chronos-Bolt Base interface only; PT and FT denote pretrained and fine-tuned augmentation modules. These workflow times have different measurement boundaries and are not a matched latency benchmark.

Explore similar work

Jun 12, 2026cs.LG

Learning the Context of Errors: Black-Box Online Adaptation of Time Series Foundation Models

The rapid evolution of Time Series Foundation Models (TSFMs) has advanced zero-shot forecasting across diverse domains. Inspired by the current form of Large Language Models, future TSFMs may be offered as commercialized, closed-source API services. However, many existing online adaptation methods still rely on white-box access for parameter fine-tuning or gradient backpropagation. This paradigm mismatch raises a question: In black-box online adaptation for TSFMs, what should we learn? We answer this with an insight: the predictive errors of the base model are conditioned on both the input and output of the base model (i.e., the context of errors). To validate this insight, we propose ORCA (Online Residual Contextual Adaptation). We conduct extensive experiments across 5 state-of-the-art TSFMs and 8 datasets to demonstrate the effectiveness of our approach. Furthermore, through ablation studies, we quantitatively analyze the impact of different adapter learning hypotheses on the final adaptation performance in black-box online adaptation. Code available at https://github.com/Fifthky/ORCA.

Xilin Dai, Yiding Liu, Hongjie Xia +4

Ant International · The Chinese University of Hong Kong

Time Series Foundation Models Zero-Shot Forecasting Wireless Foundation Models

Ask PDF

Apr 13, 2026cs.LG

TempusBench: An Evaluation Framework for Time-Series Forecasting

Foundation models have transformed natural language processing and computer vision, and a rapidly growing literature on time-series foundation models (TSFMs) seeks to replicate this success in forecasting. While recent open-source models demonstrate the promise of TSFMs, the field lacks a comprehensive and community-accepted model evaluation framework. We see at least four major issues impeding progress on the development of such a framework. First, existing evaluation frameworks comprise benchmark forecasting tasks derived from often outdated datasets (e.g., M3), many of which lack clear metadata and overlap with the corpora used to pre-train TSFMs. Second, these frameworks evaluate models along a narrowly defined set of benchmark forecasting tasks, such as forecast horizon length or domain, but overlook core statistical properties such as non-stationarity and seasonality. Third, domain-specific models (e.g., XGBoost) are often compared unfairly, as existing frameworks do not enforce a systematic and consistent hyperparameter tuning convention for all models. Fourth, visualization tools for interpreting comparative performance are lacking. To address these issues, we introduce TempusBench, an open-source evaluation framework for TSFMs. TempusBench consists of 1) new datasets which are not included in existing TSFM pretraining corpora, 2) a set of novel benchmark tasks that go beyond existing ones, 3) a model evaluation pipeline with a standardized hyperparameter tuning protocol, and 4) a tensorboard-based visualization interface. We provide access to our code on GitHub: https://github.com/Smlcrm/TempusBench and maintain a live leaderboard at https://smlcrm.com/tempusbench.

Denizalp Goktas, Gerardo Riaño-Briceño, Omkar Tekawade +14

Simulacrum New York City, NY, USA

Time Series Foundation Models

Ask PDF

Aug 8, 2026cs.LG

Ground-Truth Neighborhood Regularization for Reinforcement Learning Post-Training of Time Series Foundation Models

Time series forecasting (TSF) plays an important role in a wide range of real-world applications. Recently, time series foundation models (TSFMs), pretrained on large-scale datasets, have demonstrated strong generalization capabilities and emerged as an important paradigm for TSF. Reinforcement learning (RL) post-training has consequently attracted growing attention as a means of further improving their performance on downstream tasks. However, we find that, in certain forecast regions, RL post-training may gradually shift the output distributions of TSFMs away from the ground truth, thereby limiting their performance. We refer to this phenomenon as \textbf{suboptimal collapse}. Our analysis suggests that difficulty in initially sampling high-quality trajectories near the ground truth is an important contributing factor to suboptimal collapse. To address this issue, we propose Ground-Truth Neighborhood Regularization (GTN-R) for RL post-training of TSFMs. GTN-R uses the ground truth as a reference for locating high-quality regions and guides the model's probability mass toward the ground-truth neighborhood. This increases the probability of sampling high-quality trajectories, mitigates suboptimal collapse, and improves performance. Moreover, GTN-R can be flexibly integrated into various RL methods for TSFMs. Extensive experiments show its effectiveness.

Jianqi Zhang, Xingyu Zhang, Zeen Song +3

Institute of Software, Chinese Academy of Sciences, Beijing, China

Time Series Foundation Models Reinforcement Learning Post-Training Tabular Prior-Data Fitted Network Collapse

Ask PDF