READ-Bench: Benchmarking Historical Instance Retrieval for Time-Series Diagnosis

Gerardo Pastrana; Haojun Li; Dhruv Mehta; Anoushka Vyas; Sina Khoshfetrat Pakazad; Henrik Ohlsson; John Paparrizos

cs.AISep 26, 2026

READ-Bench: Benchmarking Historical Instance Retrieval for Time-Series Diagnosis

Authors: Gerardo Pastrana, Haojun Li, Dhruv Mehta, Anoushka Vyas, Sina Khoshfetrat Pakazad, Henrik Ohlsson, John Paparrizos

Organizations: C3 AI · The Ohio State University

Abstract

Time-series diagnostic systems rarely rely on retrieving relevant historical cases, and when they do, retrieval is evaluated only indirectly through downstream prediction. We introduce READ-Bench, a benchmark for historical-case retrieval across 12 diagnostic datasets, centered on multivariate time series, that defines relevance by shared fault or event type rather than signal shape, so visually different traces of the same fault count as relevant while similar-looking traces of different faults do not. Treating retrieval as a base retriever followed by a reranker, we evaluate classical distances, symbolic retrievers, self-supervised and foundation-model embedders, and their fusion, plus label-aware and language-model rerankers, under one protocol that varies supervision, pollution, and corpus scale with significance testing. Under a common channel-independent interface, pretrained representations offer no statistically detectable advantage over strong classical and symbolic baselines for search alone. The decisive factor is a small amount of resolved-case supervision at reranking, namely a Gaussian-process reranker that propagates a few neighbor labels in embedding space, which helps far more than more sophisticated representations or language-model reasoning and holds under pollution and at full corpus scale. Guided by these findings, we fuse a normal-residual-scored embedder with a dynamic time warping leg via reciprocal-rank fusion, then rerank with the Gaussian-process reranker, improving NDCG@10 over its own search stage on all 12 datasets, by +0.11 from reranking and +0.16 over the strongest single base retriever.

Figures & tables

Figure 1: Per-dataset NDCG@10 over the 12 READ-Bench datasets (spokes grouped by diagnostic domain). The base retrievers, coloured by family, form a tight inner band with no clear winner, while our system (CHARM

+

|

DTW-I

+

GPC) improves over them on most datasets.

Figure 2: The correct retrieval shares the query’s fault type but differs in shape, while a wrong retrieval of a different fault is closer in raw shape, so relevance is not signal similarity.

Method	Family	$\rho{=}0\%$	$\rho{=}10\%$	$\rho{=}20\%$
Random	Trivial floor	0.189	0.168	0.155
Majority		0.246	0.242	0.166
SBD-D	Raw-window distance	0.328	0.271	0.245
DTW-I		0.480	0.414	0.395
SAX-BM25	Symbolic	0.489	0.457	0.431
SFA-BM25		0.409	0.386	0.376

Table 1: NDCG@10 of every base retriever, our best composed system, and the supervised reference, macro-averaged over the 12 datasets at each pollution level

\rho

, the fraction of the corpus diluted with normal-class series (

0\%/\!\approx\!10\%/\!\approx\!20\%

). The supervised rows are a classification reference rather than retrieval systems. In each column the best value is bold and the second best underlined, over the base retrievers and our system (the supervised reference excluded).

Reranking the CHARM pool
Method	NDCG@10
CHARM (base)	0.486
Toto-1.0-QA	0.493
ChatTS	0.475
Claude	0.496
Claude $+$ TSAD	0.492

Table 2: Language-model methods on the CHARM pool (NDCG@10). The first block reranks the retrieved pool and the second classifies the fault type with no retrieval. Best per block bold, second best underlined.

Figure 3: The NR embedder variants against the non-NR field as the corpus is diluted with normal series (NDCG@10, macro-averaged over the 12 datasets). The shaded band spans the non-NR retrievers (raw distances, symbolic BM25, and ED embedders) with their mean dashed; the four NR embedder variants (bold) sit above it and stay high at every pollution level while the field degrades. MantisV2

+

NR is the strongest base retriever throughout.

Figure 4: Effect of GPC reranking across base retrievers (NDCG@10, macro-averaged over the 12 datasets), the full grid is in Table 14 . Each arrow runs from a base retriever to its GPC-reranked version, and the label-aware reranker gives a consistent gain on every base retriever. The dashed line marks the supervised reference (MR-Hydra), which no single-pool system reaches.

Appendix figures & tables74 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$C$	$T$	Classes	$\|\mathcal{C}\|$ ( $0/10/20\%$ )	$\|\mathcal{Q}\|$	rel $/q$	Imb.
CTSR ( Yeh et al., 2023 )	1	512	94	1000/1111/1250	100	10.6	1.2
DAMADICS ( Bartyś et al., 2006 )	32	256	4	1000/1111/1250	18	308.8	83.9
Exathlon ( Jacob et al., 2021 )	51	256	6	1000/1111/1250	100	215.9	13.2
HAI ( Shin et al., 2020 )	79	256	6	305/339/381	28	60.8	10.0
MIT-BIH ( Moody and Mark, 2001 )	2	256	4	1000/1111/1250	100	336.9	10.7
Petrobras 3W ( Vargas et al., 2019 )	2–7	256	9	1000/1111/1250	100	146.8	23.7

Appendix

Table 3: Per-dataset composition of READ-Bench.

C

is the channel count,

T

the window length, and “classes” the number of non-normal fault types.

|\mathcal{C}|

is the corpus size at the three pollution levels (

0/10/20\%

normal windows),

|\mathcal{Q}|

the query count, and “rel

/q

” the mean number of same-class corpus entries per query at

\rho{=}0\%

. “Imb.” is the class-imbalance ratio (most frequent over least frequent fault class, at

\rho{=}0\%

Dataset	Label source	Label rule	Split unit
CTSR	UCR dataset and class label	native label	random hold-out
DAMADICS	actuator-fault schedule	majority or coverage	fault instance
Exathlon	disturbance interval	majority or coverage	trace
HAI	attack schedule	majority or coverage	attack instance
MIT-BIH	per-beat class	window majority	patient recording
Petrobras 3W	per-timestep event class	window majority	instance

Appendix

Table 4: How each dataset’s fault-type label is derived. “Label source” is the native annotation the label is read from; “label rule” is how a window’s fault type is resolved from it, one of “window majority” (a fault holds over half the window), “majority or coverage” (window majority or the short-fault relaxation of Section 3.2 ), or “native label” (the window inherits its as-shipped annotation); “split unit” is the granularity at which train and test are kept disjoint so that queries are held out from their corpus.

Figure 5: Example fault signatures across four READ-Bench domains, one per panel, with the anomalous span shaded and the most informative channel shown.

Figure 6: Representative anomaly classes across the READ-Bench datasets. Each row is one dataset and each cell shows a representative window of one fault type (red) against a normal window (grey).

Method	Hyperparameter	Value
DTW	Sakoe–Chiba band radius	$\max(1,\operatorname{round}(0.1\,T))$
NR	nearest-normal count $k$	30
SAX-BM25	window / stride	96/8
	PAA word length	6
	alphabet size	4
SFA-BM25	window / stride	128/8

Appendix

Table 5: Distance and symbolic baseline settings.

T

is the window length, so the DTW band scales with the window. The symbolic vocabulary size is alphabet

{}^{\text{word length}}

Encoder	Hyperparameter	Value
CHARM	embedding dimension	384
CHARM	pooling / batch	mean / 32
Chronos-2	embedding dimension	768
Chronos-2	pooling / batch	mean / 32
MantisV2	per-channel width	256
	patch grid	32 patches ( $T$ a multiple of 32, $T\geq 64$ )

Appendix

Table 6: Foundation-model embedder settings. The embedding dimension and any patch-grid constraint are fixed by the released checkpoint; pooling and inference batch size are set by us.

Method	Hyperparameter	Value
RRF	fusion constant $k_{0}$	60
Weighted sum	weight $\alpha$	0.5
Two-step cascade	shortlist size $K^{\prime}$	100
PRF	neighbourhood $k$ / weight $\beta$	10/0.5
Purity	neighbourhood $k$ / weight $\alpha$	10/0.5
Majority-Vote	neighbourhood $k$	10

Appendix

Table 7: Fusion and reranking settings. The weighted sum and the purity-based rerankers combine per-query

z

-scored terms with weight

\alpha

Panel of Figure 7 — Figure 7: Base retrievers and the supervised reference over the 12 datasets, both at $\rho{=}0\%$ and family-coloured. ( 7(a) ) Retrieval accuracy against wall-clock scoring time (log scale, median per-dataset runtime). The base retrievers span two orders of magnitude in scoring time yet cluster in a narrow accuracy band well below the supervised reference, so among label-free methods the choice of a classical distance, a symbolic retriever, or a pretrained embedder trades more compute than retrieval quality. ( 7(b) ) The matching per-dataset NDCG@10 distribution, one box per method ordered by median (dashed line is the mean, solid the median, dots the 12 per-dataset scores), where the label-free retrievers overlap heavily and remain below the supervised reference.

Method	Family	P@1	P@5	HR@10	NDCG@5	NDCG@10	NDCG@20	macro-F1	$\sigma$
Random	Trivial floor	0.172	0.183	0.699	0.182	0.189	0.192	0.125	0.10
Majority		0.246	0.246	0.246	0.246	0.246	0.247	0.060	0.13
SBD-D	Raw-window distance	0.353	0.340	0.739	0.345	0.328	0.317	0.330	0.17
DTW-I		0.568	0.492	0.857	0.510	0.480	0.470	0.464	0.18
DTW-D		0.507	0.434	0.787	0.451	0.422	0.412	0.435	0.18
SAX-BM25	Symbolic	0.553	0.498	0.897	0.512	0.489	0.466	0.503	0.20

Appendix

Table 8: Full base-retriever metric grid at

\rho{=}0\%

, macro-averaged over the 12 datasets.

\sigma

is the standard deviation of NDCG@10 across datasets. The supervised rows are a classification reference rather than retrieval systems.

Method	Capped	Full	$\Delta$
CHARM	0.486	0.437	$-0.049$
CHARM $+$ NR	0.523	0.442	$-0.081$
Chronos-2	0.471	0.437	$-0.034$
MantisV2	0.510	0.447	$-0.063$
TiRex	0.483	0.434	$-0.049$

Appendix

Table 9: Capped versus full-corpus NDCG@10, macro-averaged over the 12 datasets (five are already at or below the cap, so their full corpus is the one already reported). The elastic distances, the symbolic BM25 retrievers, and the supervised reference do not scale to the largest corpora and are omitted. In each column the best value is bold and the second best underlined.

Embedder	Scoring	$\rho{=}0\%$	$\rho{=}10\%$	$\rho{=}20\%$	$\Delta$
CHARM	ED	0.486	0.461	0.437	$-0.049$
CHARM	NR	0.523	0.511	0.502	$-0.021$
MantisV2	ED	0.510	0.474	0.454	$-0.056$
MantisV2	NR	0.555	0.539	0.527	$-0.028$
Chronos-2	ED	0.471	0.455	0.432	$-0.039$
Chronos-2	NR	0.474	0.467	0.452	$-0.022$

Appendix

Table 10: NR scoring versus ED on each embedder, NDCG@10 at each pollution level

\rho

, where

\Delta

is the drop from

\rho{=}0\%

\rho{=}20\%

and a smaller-magnitude

\Delta

means more robust.

Embedder	Leg	RRF	WS	2-step
CHARM	DTW-I	0.536	0.521	0.498
	DTW-D	0.492	0.477	0.444
	SBD-D	0.445	0.454	0.397
	SAX-BM25	0.519	0.500	0.497
	SFA-BM25	0.485	0.448	0.456
CHARM $+$ NR	DTW-I	0.575	0.543	0.538

Appendix

Table 11: Choosing the fusion leg and mode. NDCG@10 at

\rho{=}0\%

, macro-averaged over the 12 datasets, for every complementary leg (rows) crossed with every fusion mode (columns), per embedder. Modes are RRF, WS (

\alpha{=}0.5

), and 2-step cascade (shortlist 100). Marks are per embedder block, and the DTW-I leg under RRF is the joint optimum in every block.

	$\|$ DTW-I			$\|$ DTW-D			$\|$ SBD-D			$\|$ SAX-BM25			$\|$ SFA-BM25
Embedder	$\rho{=}0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$
CHARM	0.536	0.493	0.453	0.492	0.461	0.425	0.445	0.426	0.397	0.519	0.492	0.459	0.485	0.459	0.430
CHARM $+$ NR	0.575	0.549	0.514	0.543	0.524	0.491	0.476	0.456	0.432	0.551	0.540	0.521	0.510	0.503	0.480
MantisV2	0.532	0.490	0.452	0.507	0.468	0.432	0.462	0.440	0.408	0.531	0.500	0.472	0.496	0.466	0.447
MantisV2 $+$ NR	0.590	0.556	0.534	0.562	0.533	0.509	0.493	0.475	0.454	0.582	0.560	0.540	0.543	0.522	0.507
Chronos-2 $+$ NR	0.551	0.521	0.498	0.528	0.508	0.485	0.442	0.422	0.404	0.538	0.527	0.510	0.499	0.490	0.470

Appendix

Table 12: Complete fusion grid. RRF of each embedder (rows) with each shape or symbolic leg (column groups), NDCG@10 at pollution levels

\rho{=}0/\!\approx\!10/\!\approx\!20\%

, macro-averaged over the 12 datasets. Both the ED and NR variant of each embedder are shown, so the fusion gains are not specific to a single embedder. Marks are on the clean (

\rho{=}0\%

) sub-column of each leg; DTW-I is the best complementary leg and MantisV2

+

NR the strongest fused pool.

Pool	Base	$+$ GPC	$+$ MajVote	$+$ QPurity	$+$ PRF	$+$ Purity
CHARM	0.486	0.556	0.538	0.511	0.455	0.386
CHARM $+$ NR	0.523	0.608	0.612	0.558	0.528	0.461
DTW-I	0.480	0.573	0.525	0.500	0.483	0.342
MantisV2	0.510	0.604	0.571	0.528	0.501	0.445
MantisV2 $+$ NR	0.555	0.645	0.616	0.562	0.511	0.489
Chronos-2 $+$ NR	0.474	0.582	0.545	0.512	0.463	0.432

Appendix

Table 13: Reranking a fixed pool, NDCG@10. Label-free rerankers are PRF and Purity; label-aware ones are GPC, MajVote, and QPurity. Language-model rerankers in Table 2 .

	$+$ GPC			$+$ MajVote			$+$ QPurity			$+$ PRF			$+$ Purity
Pool	$\rho{=}0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$	$0\%$	$10\%$	$20\%$
CHARM	0.556	0.562	0.517	0.538	0.536	0.500	0.511	0.513	0.481	0.455	0.443	0.407	0.386	0.383	0.332
CHARM $+$ NR	0.608	0.608	0.578	0.612	0.614	0.579	0.558	0.566	0.535	0.528	0.524	0.499	0.461	0.464	0.427
MantisV2	0.604	0.556	0.532	0.571	0.548	0.522	0.528	0.507	0.486	0.501	0.470	0.450	0.445	0.431	0.413
MantisV2 $+$ NR	0.645	0.613	0.582	0.616	0.610	0.589	0.562	0.549	0.538	0.511	0.498	0.490	0.489	0.478	0.458
Chronos-2 $+$ NR	0.582	0.573	0.563	0.545	0.546	0.534	0.512	0.508	0.492	0.463	0.464	0.451	0.432	0.440	0.417

Appendix

Table 14: Complete reranker grid. Each reranker on each pool, NDCG@10 at pollution levels

\rho{=}0/\!\approx\!10/\!\approx\!20\%

, macro-averaged over the 12 datasets. Pools include both the ED and NR variant of each embedder, so the reranking gains are shown to hold across embedders and are not specific to the CHARM-based NR. The label-free rerankers are PRF and Purity; the label-aware ones are GPC, MajVote, and QPurity. Marks are on the clean (

\rho{=}0\%

) sub-column of each reranker.

	Fusion	Fusion $+$ GPC
Fusion	$\rho{=}0\%$	$\rho{=}0\%$	$\rho{=}10\%$	$\rho{=}20\%$
CHARM $+$ NR $\|$ DTW-I	0.575	0.687	0.640	0.619
MantisV2 $+$ NR $\|$ DTW-I	0.590	0.665	0.641	0.604
CHARM $\|$ DTW-I	0.536	0.631	0.582	0.535
MantisV2 $\|$ DTW-I	0.532	0.609	0.564	0.552

Appendix

Table 15: Composing fusion and reranking. NDCG@10 of the fused pool (fusion only, at

\rho{=}0\%

) and of the same pool reranked with GPC across pollution levels

\rho

, macro-averaged over the 12 datasets. In each column the best value is bold and the second best underlined.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.000	0.008	0.012	0.000	0.040	0.120	0.000	0.005	0.009	0.014	0.000
DAMADICS	0.333	0.300	0.317	0.333	0.611	0.667	0.333	0.299	0.309	0.325	0.170
Exathlon	0.180	0.224	0.219	0.180	0.740	0.890	0.180	0.221	0.219	0.211	0.134
HAI	0.179	0.214	0.225	0.179	0.679	0.893	0.179	0.203	0.215	0.199	0.163
MIT-BIH	0.300	0.340	0.364	0.300	0.840	0.960	0.300	0.334	0.353	0.351	0.214

Appendix

Table 16: Full metric grid for Random , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.010	0.010	0.010	0.010	0.010	0.010	0.010	0.010	0.010	0.014	0.000
DAMADICS	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.091
Exathlon	0.420	0.420	0.420	0.420	0.420	0.420	0.420	0.420	0.420	0.420	0.099
HAI	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.059
MIT-BIH	0.370	0.370	0.370	0.370	0.370	0.370	0.370	0.370	0.370	0.370	0.135

Appendix

Table 17: Full metric grid for Majority , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.650	0.516	0.430	0.650	0.850	0.890	0.650	0.540	0.472	0.552	0.637
DAMADICS	0.389	0.411	0.333	0.389	0.722	0.889	0.389	0.411	0.370	0.406	0.302
Exathlon	0.860	0.760	0.732	0.860	0.920	0.950	0.860	0.782	0.755	0.720	0.744
HAI	0.286	0.350	0.354	0.286	0.679	0.786	0.286	0.337	0.349	0.337	0.514
MIT-BIH	0.430	0.496	0.543	0.430	0.710	0.790	0.430	0.484	0.521	0.548	0.224

Appendix

Table 18: Full metric grid for ED , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.010	0.454	0.403	0.010	0.830	0.840	0.010	0.386	0.375	0.454	0.572
DAMADICS	0.389	0.233	0.217	0.389	0.444	0.556	0.389	0.263	0.247	0.255	0.095
Exathlon	0.370	0.362	0.306	0.370	0.450	0.500	0.370	0.367	0.327	0.275	0.427
HAI	0.286	0.286	0.279	0.286	0.643	0.750	0.286	0.279	0.280	0.272	0.269
MIT-BIH	0.830	0.832	0.808	0.830	0.950	0.950	0.830	0.832	0.816	0.777	0.673

Appendix

Table 19: Full metric grid for SBD-D , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.730	0.572	0.473	0.730	0.840	0.890	0.730	0.605	0.526	0.583	0.610
DAMADICS	0.667	0.622	0.583	0.667	0.722	0.833	0.667	0.641	0.635	0.651	0.506
Exathlon	0.350	0.318	0.307	0.350	0.400	0.500	0.350	0.325	0.315	0.329	0.372
HAI	0.107	0.121	0.154	0.107	0.464	0.786	0.107	0.115	0.141	0.171	0.032
MIT-BIH	0.560	0.510	0.522	0.560	0.800	0.900	0.560	0.523	0.527	0.529	0.445

Appendix

Table 20: Full metric grid for DTW-I , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.730	0.572	0.473	0.730	0.840	0.890	0.730	0.605	0.526	0.583	0.610
DAMADICS	0.500	0.356	0.300	0.500	0.667	0.667	0.500	0.384	0.350	0.318	0.295
Exathlon	0.400	0.388	0.357	0.400	0.460	0.500	0.400	0.394	0.370	0.361	0.408
HAI	0.107	0.114	0.136	0.107	0.393	0.571	0.107	0.114	0.131	0.156	0.123
MIT-BIH	0.470	0.442	0.426	0.470	0.630	0.720	0.470	0.450	0.436	0.430	0.389

Appendix

Table 21: Full metric grid for DTW-D , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.410	0.302	0.257	0.410	0.730	0.810	0.410	0.324	0.285	0.333	0.268
DAMADICS	0.333	0.311	0.289	0.333	0.722	0.889	0.333	0.322	0.315	0.333	0.231
Exathlon	0.870	0.856	0.849	0.870	0.970	1.000	0.870	0.860	0.854	0.822	0.825
HAI	0.429	0.421	0.393	0.429	0.750	0.857	0.429	0.425	0.403	0.371	0.360
MIT-BIH	0.680	0.644	0.639	0.680	0.900	0.950	0.680	0.654	0.647	0.621	0.697

Appendix

Table 22: Full metric grid for SAX-BM25 , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.360	0.308	0.262	0.360	0.680	0.750	0.360	0.318	0.283	0.335	0.313
DAMADICS	0.444	0.333	0.306	0.444	0.667	0.833	0.444	0.362	0.345	0.346	0.300
Exathlon	0.790	0.732	0.695	0.790	0.960	0.980	0.790	0.744	0.714	0.658	0.800
HAI	0.357	0.364	0.325	0.357	0.750	0.786	0.357	0.368	0.340	0.320	0.282
MIT-BIH	0.480	0.468	0.433	0.480	0.800	0.940	0.480	0.468	0.444	0.429	0.379

Appendix

Table 23: Full metric grid for SFA-BM25 , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.650	0.516	0.430	0.650	0.850	0.890	0.650	0.540	0.472	0.552	0.637
DAMADICS	0.389	0.411	0.333	0.389	0.722	0.889	0.389	0.411	0.370	0.406	0.302
Exathlon	0.860	0.760	0.732	0.860	0.920	0.950	0.860	0.782	0.755	0.720	0.744
HAI	0.286	0.350	0.354	0.286	0.679	0.786	0.286	0.337	0.349	0.337	0.514
MIT-BIH	0.430	0.496	0.543	0.430	0.710	0.790	0.430	0.484	0.521	0.548	0.224

Appendix

Table 24: Full metric grid for CHARM , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.630	0.512	0.424	0.630	0.800	0.870	0.630	0.541	0.469	0.544	0.629
DAMADICS	0.556	0.489	0.478	0.556	0.778	1.000	0.556	0.499	0.518	0.521	0.354
Exathlon	0.920	0.850	0.818	0.920	0.990	1.000	0.920	0.863	0.837	0.782	0.886
HAI	0.393	0.321	0.332	0.393	0.643	0.786	0.393	0.331	0.338	0.326	0.462
MIT-BIH	0.460	0.588	0.638	0.460	0.790	0.880	0.460	0.565	0.608	0.650	0.328

Appendix

Table 25: Full metric grid for CHARM

+

NR , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.510	0.452	0.365	0.510	0.810	0.830	0.510	0.465	0.401	0.453	0.496
DAMADICS	0.111	0.189	0.200	0.111	0.611	0.611	0.111	0.176	0.191	0.210	0.029
Exathlon	0.960	0.912	0.881	0.960	0.990	0.990	0.960	0.921	0.896	0.858	0.897
HAI	0.321	0.321	0.311	0.321	0.750	0.964	0.321	0.325	0.317	0.322	0.346
MIT-BIH	0.720	0.708	0.719	0.720	0.860	0.890	0.720	0.708	0.716	0.705	0.393

Appendix

Table 26: Full metric grid for Chronos-2 , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.444	0.357	0.600	0.770	0.820	0.600	0.474	0.403	0.463	0.535
DAMADICS	0.167	0.233	0.267	0.167	0.611	0.889	0.167	0.230	0.261	0.292	0.163
Exathlon	0.910	0.852	0.810	0.910	0.950	0.980	0.910	0.860	0.828	0.781	0.748
HAI	0.357	0.364	0.354	0.357	0.714	0.786	0.357	0.363	0.357	0.359	0.369
MIT-BIH	0.680	0.752	0.759	0.680	0.870	0.920	0.680	0.739	0.748	0.745	0.416

Appendix

Table 27: Full metric grid for Chronos-2

+

NR , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.590	0.522	0.411	0.590	0.780	0.850	0.590	0.539	0.456	0.528	0.596
DAMADICS	0.500	0.578	0.611	0.500	0.944	1.000	0.500	0.555	0.610	0.622	0.359
Exathlon	0.950	0.934	0.908	0.950	1.000	1.000	0.950	0.939	0.919	0.897	0.964
HAI	0.536	0.429	0.396	0.536	0.714	0.964	0.536	0.460	0.432	0.397	0.539
MIT-BIH	0.310	0.328	0.333	0.310	0.400	0.450	0.310	0.325	0.329	0.324	0.355

Appendix

Table 28: Full metric grid for MantisV2 , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.610	0.508	0.410	0.610	0.770	0.850	0.610	0.534	0.458	0.527	0.580
DAMADICS	0.611	0.578	0.472	0.611	0.778	0.889	0.611	0.596	0.544	0.540	0.389
Exathlon	0.940	0.938	0.924	0.940	0.970	0.980	0.940	0.939	0.929	0.914	0.958
HAI	0.357	0.307	0.357	0.357	0.464	0.786	0.357	0.312	0.350	0.354	0.351
MIT-BIH	0.510	0.538	0.546	0.510	0.700	0.830	0.510	0.528	0.537	0.550	0.356

Appendix

Table 29: Full metric grid for MantisV2

+

NR , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.670	0.518	0.456	0.670	0.840	0.880	0.670	0.548	0.494	0.565	0.543
DAMADICS	0.444	0.322	0.333	0.444	0.611	0.833	0.444	0.345	0.353	0.361	0.208
Exathlon	0.860	0.840	0.821	0.860	0.990	1.000	0.860	0.849	0.832	0.786	0.860
HAI	0.429	0.321	0.314	0.429	0.786	0.964	0.429	0.344	0.335	0.333	0.350
MIT-BIH	0.580	0.614	0.633	0.580	0.820	0.880	0.580	0.607	0.622	0.641	0.458

Appendix

Table 30: Full metric grid for TiRex , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.670	0.516	0.435	0.670	0.850	0.870	0.670	0.552	0.483	0.550	0.596
DAMADICS	0.444	0.322	0.344	0.444	0.778	0.889	0.444	0.349	0.373	0.407	0.215
Exathlon	0.880	0.828	0.809	0.880	0.990	0.990	0.880	0.838	0.822	0.758	0.820
HAI	0.321	0.329	0.339	0.321	0.679	0.786	0.321	0.328	0.335	0.333	0.374
MIT-BIH	0.650	0.662	0.656	0.650	0.800	0.850	0.650	0.661	0.657	0.663	0.367

Appendix

Table 31: Full metric grid for TiRex

+

NR , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.711	0.572
DAMADICS	0.278	0.278	0.278	0.278	0.278	0.333	0.278	0.278	0.289	0.305	0.217
Exathlon	0.820	0.820	0.820	0.820	0.820	0.820	0.820	0.820	0.820	0.820	0.677
HAI	0.429	0.429	0.421	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.388
MIT-BIH	0.400	0.400	0.400	0.400	0.400	0.400	0.400	0.400	0.400	0.400	0.243

Appendix

Table 32: Full metric grid for CHARM

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.631	0.516
DAMADICS	0.278	0.278	0.272	0.278	0.278	0.333	0.278	0.278	0.285	0.289	0.217
Exathlon	0.780	0.780	0.780	0.780	0.780	0.780	0.780	0.780	0.780	0.780	0.541
HAI	0.464	0.464	0.457	0.464	0.464	0.464	0.464	0.464	0.464	0.464	0.417
MIT-BIH	0.440	0.440	0.440	0.440	0.440	0.440	0.440	0.440	0.440	0.440	0.275

Appendix

Table 33: Full metric grid for CHARM

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.570	0.510	0.417	0.570	0.730	0.840	0.570	0.523	0.454	0.502	0.525
DAMADICS	0.222	0.289	0.278	0.222	0.444	0.444	0.222	0.273	0.274	0.282	0.167
Exathlon	0.810	0.786	0.781	0.810	0.810	0.810	0.810	0.792	0.786	0.781	0.608
HAI	0.429	0.407	0.414	0.429	0.429	0.536	0.429	0.414	0.417	0.404	0.299
MIT-BIH	0.440	0.434	0.434	0.440	0.440	0.490	0.440	0.435	0.435	0.435	0.275

Appendix

Table 34: Full metric grid for CHARM

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.480	0.444	0.412	0.480	0.790	0.880	0.480	0.452	0.428	0.508	0.428
DAMADICS	0.389	0.311	0.289	0.389	0.611	0.778	0.389	0.327	0.315	0.360	0.274
Exathlon	0.740	0.728	0.703	0.740	0.860	0.920	0.740	0.731	0.713	0.681	0.513
HAI	0.429	0.379	0.364	0.429	0.571	0.786	0.429	0.395	0.383	0.367	0.331
MIT-BIH	0.440	0.450	0.454	0.440	0.480	0.520	0.440	0.447	0.451	0.464	0.275

Appendix

Table 35: Full metric grid for CHARM

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.130	0.106	0.106	0.130	0.170	0.220	0.130	0.110	0.109	0.138	0.061
DAMADICS	0.056	0.156	0.189	0.056	0.278	0.333	0.056	0.133	0.164	0.191	0.075
Exathlon	0.750	0.728	0.719	0.750	0.850	0.910	0.750	0.735	0.727	0.688	0.512
HAI	0.250	0.271	0.246	0.250	0.429	0.571	0.250	0.267	0.252	0.254	0.143
MIT-BIH	0.380	0.446	0.469	0.380	0.580	0.660	0.380	0.431	0.453	0.476	0.275

Appendix

Table 36: Full metric grid for CHARM

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.661	0.499
DAMADICS	0.444	0.444	0.444	0.444	0.444	0.444	0.444	0.444	0.444	0.444	0.224
Exathlon	0.900	0.900	0.900	0.900	0.900	0.900	0.900	0.900	0.900	0.900	0.776
HAI	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.339
MIT-BIH	0.710	0.710	0.710	0.710	0.710	0.710	0.710	0.710	0.710	0.710	0.444

Appendix

Table 37: Full metric grid for CHARM

+

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.632	0.520
DAMADICS	0.556	0.556	0.522	0.556	0.556	0.722	0.556	0.556	0.566	0.596	0.342
Exathlon	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.858
HAI	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.374
MIT-BIH	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.355

Appendix

Table 38: Full metric grid for CHARM

+

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.522	0.430	0.600	0.730	0.860	0.600	0.540	0.470	0.519	0.529
DAMADICS	0.389	0.444	0.472	0.389	0.778	0.944	0.389	0.429	0.461	0.481	0.205
Exathlon	0.890	0.854	0.838	0.890	0.930	0.960	0.890	0.861	0.847	0.829	0.624
HAI	0.464	0.436	0.414	0.464	0.607	0.750	0.464	0.443	0.427	0.391	0.394
MIT-BIH	0.570	0.616	0.653	0.570	0.720	0.830	0.570	0.609	0.637	0.674	0.348

Appendix

Table 39: Full metric grid for CHARM

+

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.530	0.480	0.418	0.530	0.800	0.850	0.530	0.495	0.447	0.523	0.433
DAMADICS	0.556	0.511	0.461	0.556	0.833	1.000	0.556	0.518	0.510	0.530	0.354
Exathlon	0.870	0.846	0.838	0.870	0.970	1.000	0.870	0.854	0.845	0.806	0.749
HAI	0.393	0.371	0.329	0.393	0.643	0.786	0.393	0.372	0.343	0.343	0.446
MIT-BIH	0.700	0.674	0.691	0.700	0.760	0.830	0.700	0.679	0.689	0.704	0.360

Appendix

Table 40: Full metric grid for CHARM

+

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.280	0.230	0.206	0.280	0.460	0.490	0.280	0.236	0.218	0.259	0.148
DAMADICS	0.167	0.344	0.372	0.167	0.611	0.611	0.167	0.314	0.346	0.344	0.222
Exathlon	0.890	0.796	0.755	0.890	0.910	0.940	0.890	0.814	0.779	0.733	0.576
HAI	0.286	0.329	0.318	0.286	0.571	0.750	0.286	0.320	0.315	0.313	0.211
MIT-BIH	0.410	0.578	0.638	0.410	0.790	0.870	0.410	0.547	0.600	0.649	0.291

Appendix

Table 41: Full metric grid for CHARM

+

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.690	0.690	0.690	0.690	0.690	0.690	0.690	0.690	0.690	0.765	0.624
DAMADICS	0.611	0.611	0.594	0.611	0.611	0.667	0.611	0.611	0.622	0.638	0.416
Exathlon	0.330	0.330	0.330	0.330	0.330	0.330	0.330	0.330	0.330	0.330	0.322
HAI	0.214	0.214	0.200	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.230
MIT-BIH	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.544

Appendix

Table 42: Full metric grid for DTW-I

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.610	0.610	0.610	0.610	0.610	0.610	0.610	0.610	0.610	0.662	0.539
DAMADICS	0.556	0.556	0.544	0.556	0.556	0.611	0.556	0.556	0.559	0.572	0.346
Exathlon	0.350	0.318	0.307	0.350	0.400	0.500	0.350	0.325	0.315	0.329	0.372
HAI	0.071	0.071	0.071	0.071	0.071	0.071	0.071	0.071	0.071	0.071	0.033
MIT-BIH	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.535

Appendix

Table 43: Full metric grid for DTW-I

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.580	0.534	0.435	0.580	0.710	0.810	0.580	0.547	0.475	0.507	0.532
DAMADICS	0.500	0.533	0.533	0.500	0.556	0.556	0.500	0.525	0.528	0.530	0.300
Exathlon	0.340	0.348	0.413	0.340	0.350	0.550	0.340	0.347	0.392	0.431	0.212
HAI	0.071	0.100	0.136	0.071	0.179	0.464	0.071	0.093	0.121	0.149	0.032
MIT-BIH	0.610	0.616	0.633	0.610	0.630	0.700	0.610	0.614	0.626	0.633	0.535

Appendix

Table 44: Full metric grid for DTW-I

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.730	0.576	0.474	0.730	0.840	0.890	0.730	0.607	0.526	0.584	0.610
DAMADICS	0.667	0.622	0.583	0.667	0.722	0.833	0.667	0.641	0.635	0.649	0.506
Exathlon	0.350	0.318	0.307	0.350	0.400	0.500	0.350	0.325	0.315	0.329	0.372
HAI	0.107	0.121	0.154	0.107	0.464	0.786	0.107	0.115	0.141	0.171	0.032
MIT-BIH	0.500	0.576	0.574	0.500	0.850	0.910	0.500	0.562	0.566	0.564	0.458

Appendix

Table 45: Full metric grid for DTW-I

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.140	0.100	0.105	0.140	0.200	0.310	0.140	0.110	0.110	0.139	0.058
DAMADICS	0.500	0.500	0.494	0.500	0.556	0.611	0.500	0.499	0.496	0.476	0.295
Exathlon	0.330	0.386	0.376	0.330	0.600	0.710	0.330	0.383	0.378	0.377	0.288
HAI	0.107	0.150	0.182	0.107	0.286	0.571	0.107	0.139	0.165	0.168	0.046
MIT-BIH	0.500	0.574	0.570	0.500	0.830	0.930	0.500	0.560	0.563	0.563	0.403

Appendix

Table 46: Full metric grid for DTW-I

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.667	0.573
DAMADICS	0.611	0.611	0.628	0.611	0.611	0.833	0.611	0.611	0.655	0.719	0.359
Exathlon	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.964
HAI	0.429	0.429	0.421	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.494
MIT-BIH	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.295

Appendix

Table 47: Full metric grid for MantisV2

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.599	0.482
DAMADICS	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.389
Exathlon	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.920	0.792
HAI	0.393	0.393	0.386	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.465
MIT-BIH	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.340	0.227

Appendix

Table 48: Full metric grid for MantisV2

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.580	0.520	0.438	0.580	0.730	0.850	0.580	0.531	0.470	0.515	0.538
DAMADICS	0.667	0.700	0.656	0.667	0.944	0.944	0.667	0.688	0.680	0.704	0.750
Exathlon	0.930	0.926	0.922	0.930	0.930	0.930	0.930	0.927	0.924	0.922	0.800
HAI	0.429	0.371	0.361	0.429	0.500	0.643	0.429	0.384	0.376	0.380	0.436
MIT-BIH	0.360	0.340	0.339	0.360	0.370	0.390	0.360	0.343	0.342	0.348	0.227

Appendix

Table 49: Full metric grid for MantisV2

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.540	0.480	0.405	0.540	0.750	0.830	0.540	0.495	0.439	0.504	0.528
DAMADICS	0.556	0.678	0.600	0.556	0.944	0.944	0.556	0.648	0.624	0.624	0.776
Exathlon	0.940	0.922	0.913	0.940	0.960	0.980	0.940	0.926	0.918	0.904	0.958
HAI	0.500	0.471	0.393	0.500	0.643	0.857	0.500	0.481	0.429	0.395	0.569
MIT-BIH	0.330	0.328	0.333	0.330	0.350	0.370	0.330	0.329	0.332	0.336	0.232

Appendix

Table 50: Full metric grid for MantisV2

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.270	0.248	0.223	0.270	0.380	0.420	0.270	0.250	0.232	0.265	0.168
DAMADICS	0.611	0.600	0.583	0.611	0.667	0.667	0.611	0.603	0.592	0.564	0.360
Exathlon	0.940	0.902	0.888	0.940	0.940	0.940	0.940	0.910	0.897	0.884	0.609
HAI	0.464	0.500	0.454	0.464	0.643	0.714	0.464	0.498	0.465	0.438	0.484
MIT-BIH	0.350	0.346	0.348	0.350	0.410	0.450	0.350	0.348	0.349	0.344	0.223

Appendix

Table 51: Full metric grid for MantisV2

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.600	0.602	0.600	0.600	0.610	0.600	0.600	0.601	0.660	0.546
DAMADICS	0.556	0.556	0.506	0.556	0.556	0.556	0.556	0.556	0.556	0.556	0.342
Exathlon	0.950	0.950	0.950	0.950	0.950	0.950	0.950	0.950	0.950	0.950	0.914
HAI	0.536	0.536	0.529	0.536	0.536	0.536	0.536	0.536	0.536	0.536	0.616
MIT-BIH	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.428

Appendix

Table 52: Full metric grid for MantisV2

+

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.590	0.590	0.590	0.590	0.590	0.590	0.590	0.590	0.590	0.632	0.526
DAMADICS	0.500	0.500	0.483	0.500	0.500	0.500	0.500	0.500	0.500	0.508	0.257
Exathlon	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.789
HAI	0.357	0.357	0.350	0.357	0.357	0.357	0.357	0.357	0.357	0.357	0.342
MIT-BIH	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.361

Appendix

Table 53: Full metric grid for MantisV2

+

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.570	0.508	0.440	0.570	0.680	0.820	0.570	0.521	0.470	0.512	0.477
DAMADICS	0.722	0.600	0.522	0.722	0.778	0.889	0.722	0.633	0.591	0.577	0.622
Exathlon	0.920	0.922	0.915	0.920	0.970	0.970	0.920	0.923	0.918	0.906	0.840
HAI	0.357	0.379	0.368	0.357	0.429	0.464	0.357	0.377	0.372	0.371	0.405
MIT-BIH	0.560	0.548	0.552	0.560	0.650	0.730	0.560	0.551	0.552	0.559	0.352

Appendix

Table 54: Full metric grid for MantisV2

+

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.460	0.434	0.379	0.460	0.720	0.830	0.460	0.441	0.401	0.474	0.374
DAMADICS	0.500	0.467	0.428	0.500	0.667	0.778	0.500	0.479	0.464	0.444	0.426
Exathlon	0.870	0.886	0.877	0.870	0.930	0.940	0.870	0.883	0.878	0.876	0.664
HAI	0.321	0.343	0.354	0.321	0.464	0.643	0.321	0.334	0.349	0.326	0.477
MIT-BIH	0.550	0.534	0.544	0.550	0.560	0.600	0.550	0.537	0.543	0.548	0.358

Appendix

Table 55: Full metric grid for MantisV2

+

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.260	0.240	0.204	0.260	0.360	0.390	0.260	0.242	0.216	0.248	0.157
DAMADICS	0.556	0.522	0.494	0.556	0.667	0.722	0.556	0.532	0.510	0.450	0.330
Exathlon	0.950	0.936	0.927	0.950	0.960	0.970	0.950	0.938	0.932	0.920	0.730
HAI	0.286	0.307	0.304	0.286	0.536	0.750	0.286	0.303	0.302	0.327	0.233
MIT-BIH	0.540	0.554	0.567	0.540	0.710	0.820	0.540	0.553	0.562	0.569	0.372

Appendix

Table 56: Full metric grid for MantisV2

+

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.660	0.662	0.661	0.660	0.670	0.670	0.660	0.662	0.661	0.706	0.592
DAMADICS	0.389	0.389	0.372	0.389	0.389	0.389	0.389	0.389	0.389	0.389	0.274
Exathlon	0.890	0.890	0.890	0.890	0.890	0.890	0.890	0.890	0.890	0.890	0.886
HAI	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.168
MIT-BIH	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.570	0.321

Appendix

Table 57: Full metric grid for TiRex

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.655	0.534
DAMADICS	0.278	0.278	0.278	0.278	0.278	0.278	0.278	0.278	0.278	0.278	0.109
Exathlon	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.818
HAI	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.250	0.161
MIT-BIH	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.630	0.334

Appendix

Table 58: Full metric grid for TiRex

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.590	0.514	0.444	0.590	0.680	0.820	0.590	0.531	0.477	0.526	0.483
DAMADICS	0.333	0.300	0.256	0.333	0.444	0.556	0.333	0.308	0.283	0.299	0.114
Exathlon	0.820	0.828	0.830	0.820	0.860	0.890	0.820	0.827	0.828	0.827	0.580
HAI	0.179	0.257	0.243	0.179	0.464	0.571	0.179	0.246	0.242	0.263	0.097
MIT-BIH	0.620	0.606	0.605	0.620	0.650	0.680	0.620	0.608	0.607	0.608	0.331

Appendix

Table 59: Full metric grid for TiRex

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.508	0.445	0.600	0.820	0.870	0.600	0.526	0.477	0.557	0.490
DAMADICS	0.278	0.289	0.328	0.278	0.500	0.833	0.278	0.285	0.322	0.329	0.109
Exathlon	0.860	0.818	0.814	0.860	0.920	0.990	0.860	0.826	0.821	0.783	0.780
HAI	0.286	0.314	0.286	0.286	0.750	0.857	0.286	0.325	0.302	0.302	0.166
MIT-BIH	0.620	0.602	0.591	0.620	0.670	0.670	0.620	0.607	0.597	0.598	0.333

Appendix

Table 60: Full metric grid for TiRex

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.190	0.188	0.171	0.190	0.310	0.370	0.190	0.190	0.178	0.209	0.103
DAMADICS	0.222	0.267	0.250	0.222	0.333	0.333	0.222	0.265	0.254	0.269	0.151
Exathlon	0.830	0.788	0.764	0.830	0.900	0.940	0.830	0.795	0.776	0.744	0.548
HAI	0.214	0.214	0.236	0.214	0.321	0.500	0.214	0.211	0.226	0.228	0.051
MIT-BIH	0.590	0.602	0.606	0.590	0.710	0.810	0.590	0.600	0.603	0.624	0.334

Appendix

Table 61: Full metric grid for TiRex

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.610	0.610	0.610	0.610	0.640	0.650	0.610	0.610	0.610	0.668	0.536
DAMADICS	0.389	0.389	0.372	0.389	0.389	0.389	0.389	0.389	0.389	0.389	0.274
Exathlon	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.874
HAI	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.214	0.159
MIT-BIH	0.720	0.710	0.709	0.720	0.760	0.780	0.720	0.712	0.710	0.708	0.461

Appendix

Table 62: Full metric grid for TiRex

+

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.600	0.643	0.539
DAMADICS	0.389	0.389	0.389	0.389	0.389	0.500	0.389	0.389	0.400	0.428	0.232
Exathlon	0.930	0.930	0.930	0.930	0.930	0.930	0.930	0.930	0.930	0.930	0.836
HAI	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.393	0.392
MIT-BIH	0.680	0.680	0.680	0.680	0.680	0.680	0.680	0.680	0.680	0.680	0.357

Appendix

Table 63: Full metric grid for TiRex

+

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.620	0.536	0.451	0.620	0.750	0.820	0.620	0.556	0.491	0.532	0.554
DAMADICS	0.444	0.367	0.339	0.444	0.667	0.833	0.444	0.388	0.373	0.384	0.296
Exathlon	0.850	0.856	0.848	0.850	0.920	0.940	0.850	0.856	0.850	0.828	0.584
HAI	0.393	0.371	0.368	0.393	0.536	0.643	0.393	0.371	0.370	0.345	0.390
MIT-BIH	0.680	0.674	0.672	0.680	0.700	0.720	0.680	0.676	0.674	0.666	0.357

Appendix

Table 64: Full metric grid for TiRex

+

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.470	0.438	0.415	0.470	0.750	0.860	0.470	0.442	0.426	0.500	0.344
DAMADICS	0.444	0.367	0.278	0.444	0.667	0.667	0.444	0.382	0.324	0.369	0.214
Exathlon	0.850	0.840	0.838	0.850	0.950	0.960	0.850	0.842	0.840	0.812	0.682
HAI	0.321	0.329	0.311	0.321	0.643	0.857	0.321	0.335	0.322	0.299	0.259
MIT-BIH	0.690	0.708	0.702	0.690	0.750	0.760	0.690	0.706	0.703	0.702	0.379

Appendix

Table 65: Full metric grid for TiRex

+

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.290	0.254	0.223	0.290	0.440	0.500	0.290	0.261	0.237	0.276	0.170
DAMADICS	0.333	0.333	0.267	0.333	0.500	0.556	0.333	0.333	0.285	0.273	0.242
Exathlon	0.850	0.844	0.791	0.850	0.970	0.990	0.850	0.845	0.809	0.749	0.659
HAI	0.250	0.221	0.243	0.250	0.500	0.679	0.250	0.221	0.235	0.234	0.107
MIT-BIH	0.660	0.672	0.662	0.660	0.750	0.820	0.660	0.670	0.663	0.664	0.367

Appendix

Table 66: Full metric grid for TiRex

+

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.550	0.622	0.488
DAMADICS	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.222	0.240	0.095
Exathlon	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.940	0.877
HAI	0.500	0.500	0.500	0.500	0.500	0.500	0.500	0.500	0.500	0.500	0.477
MIT-BIH	0.770	0.770	0.770	0.770	0.770	0.770	0.770	0.770	0.770	0.770	0.406

Appendix

Table 67: Full metric grid for Chronos-2

+

+

GPC , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.540	0.540	0.540	0.540	0.540	0.540	0.540	0.540	0.540	0.591	0.465
DAMADICS	0.278	0.278	0.283	0.278	0.278	0.333	0.278	0.278	0.281	0.310	0.139
Exathlon	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.870	0.715
HAI	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.429	0.401
MIT-BIH	0.760	0.760	0.760	0.760	0.760	0.760	0.760	0.760	0.760	0.760	0.396

Appendix

Table 68: Full metric grid for Chronos-2

+

+

MajVote , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.550	0.444	0.368	0.550	0.670	0.790	0.550	0.468	0.407	0.446	0.466
DAMADICS	0.389	0.333	0.300	0.389	0.500	0.611	0.389	0.346	0.325	0.329	0.243
Exathlon	0.860	0.842	0.837	0.860	0.870	0.900	0.860	0.844	0.840	0.821	0.629
HAI	0.357	0.393	0.371	0.357	0.643	0.679	0.357	0.393	0.378	0.363	0.352
MIT-BIH	0.750	0.734	0.744	0.750	0.760	0.810	0.750	0.737	0.743	0.738	0.396

Appendix

Table 69: Full metric grid for Chronos-2

+

+

QPurity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.420	0.372	0.318	0.420	0.700	0.780	0.420	0.382	0.341	0.405	0.396
DAMADICS	0.222	0.278	0.294	0.222	0.667	0.667	0.222	0.266	0.282	0.289	0.095
Exathlon	0.840	0.854	0.846	0.840	0.930	0.940	0.840	0.850	0.846	0.822	0.714
HAI	0.393	0.393	0.361	0.393	0.750	0.821	0.393	0.389	0.369	0.335	0.398
MIT-BIH	0.750	0.738	0.740	0.750	0.780	0.810	0.750	0.739	0.740	0.734	0.392

Appendix

Table 70: Full metric grid for Chronos-2

+

+

PRF , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.270	0.206	0.176	0.270	0.370	0.430	0.270	0.219	0.194	0.231	0.142
DAMADICS	0.167	0.178	0.217	0.167	0.389	0.611	0.167	0.171	0.203	0.214	0.079
Exathlon	0.830	0.826	0.804	0.830	0.930	0.960	0.830	0.828	0.812	0.781	0.573
HAI	0.357	0.343	0.325	0.357	0.679	0.750	0.357	0.354	0.337	0.331	0.324
MIT-BIH	0.730	0.726	0.717	0.730	0.820	0.850	0.730	0.727	0.720	0.715	0.396

Appendix

Table 71: Full metric grid for Chronos-2

+

+

Purity , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.750	0.770	0.768	0.750	0.800	0.800	0.750	0.767	0.767	0.783	0.743
DAMADICS	0.667	0.667	0.667	0.667	0.667	0.778	0.667	0.667	0.689	0.712	0.478
Exathlon	0.960	0.966	0.965	0.960	0.970	0.970	0.960	0.964	0.964	0.964	0.927
HAI	0.321	0.379	0.382	0.321	0.571	0.571	0.321	0.367	0.374	0.414	0.200
MIT-BIH	0.760	0.838	0.823	0.760	0.910	0.910	0.760	0.824	0.819	0.826	0.640

Appendix

Table 72: Full metric grid for MR-Hydra , per dataset and macro-averaged, at each pollution level.

$\rho{=}0\%$
	P@1	P@5	P@10	HR@1	HR@5	HR@10	N@1	N@5	N@10	N@20	mF1
CTSR	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.670	0.673	0.605
DAMADICS	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.667	0.392
Exathlon	0.980	0.980	0.980	0.980	0.980	0.980	0.980	0.980	0.980	0.980	0.989
HAI	0.357	0.357	0.357	0.357	0.357	0.357	0.357	0.357	0.357	0.357	0.251
MIT-BIH	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.580	0.628

Appendix

Table 73: Full metric grid for RDST , per dataset and macro-averaged, at each pollution level.

Explore similar work

Which Histories Matter for Time Series Forecasting? Learning Predictive Relevance with Future Supervision
Aug 24, 2026Yong-Hoon Choi, Kwang-Hyun Park, Youngjin ChoLearning to Rank Time Series Forecasting
Division of Robotics, Kwangwoon University Seoul 01897, Republic of Korea
Revitalizing Medical Time Series with Vision-Informed Retrieval: A Vision-Language Perspective
Sep 28, 2026Guoqi Yu, Juncheng Wang, Shujun WangMultivariate Time Series Classification Time Series Classification
Department of Biomedical Engineering and Sports Technology The Hong Kong Polytechnic University
TS-Haystack: A Multi-Task Retrieval Benchmark for Long-Context Time-Series Reasoning
Feb 15, 2026Nicolas Zumarraga, Thomas Kaar, Ning Wang +12Time Series Reasoning Time Series QA
1Agentic Systems Lab, ETH Zurich · 2Stanford University · 3Traffic Engineering Group, Institute for Transport Planning and Systems, ETH Zurich +4

Aug 24, 2026, cs.IRJ/K move · Enter open · S save

Which Histories Matter for Time Series Forecasting? Learning Predictive Relevance with Future Supervision

Yong-Hoon Choi, Kwang-Hyun Park, Youngjin Cho

Division of Robotics, Kwangwoon University Seoul 01897, Republic of Korea

Learning to Rank Time Series Forecasting Retrieval-Augmented Time Series Forecasting

Open paper AskPDF arXiv

Historical retrieval for time-series prediction commonly treats past similarity as a proxy for usefulness. We ask a different question: which historical examples should be expected to matter for a query? We define predictive relevance as expected future utility conditioned on inference-time information, using realized futures only during training as privileged supervision. A normalized-pattern retriever first forms a coarse candidate set, and a lightweight residual multilayer perceptron (MLP) learns a listwise future-compatibility target while keeping inference-time scoring strictly past-only. Our method retains similarity-based candidate generation but reranks its candidates by a more predictive relevance criterion. Optimal relevance decomposes into candidate-level utility and query-specific compatibility, motivating Candidate-Prior and Shuffled-Future controls. Across six benchmarks, the reranker improves Pattern retrieval while revealing candidate-global, query-specific, and mixed relevance regimes. On all 12 confirmatory tasks, it improves Pattern and outperforms a matched-protocol Stationarity-Aware Retrieval-Augmented Time Series Forecasting (SARAF) retrieval rule. Architecture-matched ablations show that correct future supervision, rather than the MLP or added context alone, drives gains in query-specific regimes. Alternative-similarity experiments show that a strong last-value-anchored L2 rule remains superior in some domains, whereas future-supervised relevance is particularly strong where our diagnostics indicate query-specific relevance, especially on Solar. Candidate-pool diagnostics show that this contrast is not explained solely by coarse Pattern retrieval. Overall, historical relevance is structured and domain dependent rather than governed by a universally superior retrieval rule.

Abstract

Explore similar work

Which Histories Matter for Time Series Forecasting? Learning Predictive Relevance with Future Supervision

Revitalizing Medical Time Series with Vision-Informed Retrieval: A Vision-Language Perspective

TS-Haystack: A Multi-Task Retrieval Benchmark for Long-Context Time-Series Reasoning