Time Series Classification Benchmark: Foundation Models vs Classical Methods
We benchmarked 13 time series classification methods, from pretrained time series foundation models to a 22-feature baseline from 2019, on 33 UCR/UEA datasets under one frozen protocol. That is 14,638 recorded method-dataset-resample cells, 11,874 of them scored.
TSC benchmark results
The chart shows the benchmark’s main comparison: 12 methods on the 15 univariate datasets where every one of them produced a result, each dataset run on the same 5 train/test splits. Every method is scored on exactly the same data, which is what makes the ranks comparable. Each point’s tooltip carries a 95% range showing how far that method’s rank moves when the collection of datasets is resampled 5,000 times at seed 1729. That range measures stability, not statistical significance.
No single winner. HIVE-COTE 2 ranks first on mean rank (2.400) and RocketPFN first on mean accuracy (0.9047). Under a 5,000-draw dataset bootstrap, first place goes to HIVE-COTE 2 in 46.6% of draws, RocketPFN in 45.5% and MultiRocket in 7.9%. The two orderings disagree and the bootstrap splits nearly evenly, so naming either one would be an artifact of the metric chosen.
The top group is smaller than the table suggests. The Wilcoxon-Holm clique is HIVE-COTE 2, RocketPFN, MultiRocket and Hydra. MiniRocket, six tenths of a rank behind Hydra at 4.700, is not in it: membership comes from the corrected pairwise tests, not from where the rank numbers fall.
On this panel the transforms take ranks 3, 4 and 5 and the pretrained models take 6, 7, 8 and 10. MOMENT, the highest-ranked pretrained model at 6.633, sits almost two full ranks below MiniRocket.
catch22, published in 2019 with 22 fixed features, lands inside the pretrained band at 8.700, ahead of Mantis-V2 (9.733) and behind TiRex (8.433). At full 30-resample depth it stays in that band, both on the 18-dataset panel below at 7.222 and on a 27-dataset, eight-method variant of it that drops RocketPFN.
Rank and accuracy disagree at the bottom too. DTW-1NN has worse mean accuracy than TiCT (0.7563 against 0.7759) and a better mean rank (9.900 against 10.200). Which metric you report changes the order in both halves of the table.
Only the middle of the table is stable. Removing any single dataset changes the ordering in 10 of 15 cases. Six of those are the HIVE-COTE 2 and RocketPFN swap; Mantis-V2 and DTW-1NN exchange places in six and DTW-1NN and TiCT in two. Positions 3 through 9 never move.
The comparison metric is accuracy because the declared primary metric, ROC-AUC, is unavailable for 5 of the 7 reference foundation models. The panel is univariate-only: TiCT and RocketPFN have no native multivariate support, so the complete-case rule removes all 10 multivariate datasets. Foundation-model rows measure those models under one fixed transfer protocol. TimEE carries no rank, for the reason below. A bootstrap first-place share is the frequency of ranking first across resampled dataset collections, and nothing more.
The ordering itself replicates the 2024 bake-off, which already established that HIVE-COTE 2 and the ROCKET family lead this problem class and reports HIVE-COTE 2 as much slower.1 The contribution here is an independently controlled measurement on a 2026 roster under one common protocol, with the separations quantified after familywise correction.
RocketPFN is ROCKET features fed to TabPFN,2 an in-context tabular model rather than a pretrained time series foundation model, so its position above the pretrained rows says nothing about that class.
Bradley-Terry rank stability
Post hoc Bradley-Terry summary of accuracy wins on the primary panel. Rank ranges are empirical dataset-bootstrap stability summaries, not pairwise significance intervals. HIVE-COTE 2, RocketPFN, MultiRocket and Hydra are the frozen Wilcoxon-Holm top clique: no within-clique difference was detected, which is not evidence of equivalence. Hydra’s rank range of [4, 5] therefore does not establish separation from the top three. No unique winner is supported.
A ratings view cannot manufacture a winner the significance tests did not find. HIVE-COTE 2 at 1384.6, RocketPFN at 1380.6 and MultiRocket at 1338.3 all carry a bootstrap rank range of [1, 3], and the same first-place shares reappear.
The 1000 anchor is relative to this panel, and the ratings encode modeled odds of winning on a dataset rather than differences in accuracy points.
Which gaps survive correction
Nineteen pairwise contrasts were inspected and six survive Holm correction. Each point carries its 95% interval and Holm-adjusted p-value in the tooltip; those intervals are dataset-bootstrap summaries and are not multiplicity-adjusted, so only the p-values carry Holm.
Every surviving contrast is a convolutional transform beating a pretrained model.
MultiRocket sweeps two of them without losing a dataset. 28 wins to 0 against both Chronos-2 and TiRex, at p = 1.4e-07.
Thirteen contrasts are inconclusive, which is not evidence of no difference. Nothing involving DTW-1NN survives correction, and nothing on the multivariate subset does.
Each row is computed on the datasets that pair shares, not on a common panel, so the rows are not a league table and cannot be read against each other.
Foundation-model results are conditional on one transfer protocol: final-layer mean pooling with a logistic probe, applied identically to MOMENT, Chronos-2, TiRex and Mantis-V2 and frozen before any scored cell. A post hoc, rank-gratuito arm changed only TiRex’s layer extraction and moved mean accuracy from 0.7933 to 0.8046, 1.13 percentage points, with 23 of 30 datasets improving and seven worsening. The largest loss, Heartbeat at 7.98 points, improved on none of its 30 resamples, so the mean hides a result that reverses by dataset. Details are in the appendix.
Six times the split depth
Every table above runs 5 train/test splits, because HIVE-COTE 2 and DTW-1NN run only those five under the frozen expensive-cell exception and the matched-resample rule pulls the rest down to match. Drop those two and the remaining nine methods run all 30 seeded stratified resamples on 18 datasets, the depth the bake-off uses.
The ordering holds. RocketPFN 1.694, MultiRocket 2.250, Hydra 2.889 and MiniRocket 3.833 take the first four places, all four built on random convolutional features. Then MOMENT 5.889, Chronos-2 6.444 and TiRex 7.056, with catch22 at 7.222 and Mantis-V2 at 7.722. Friedman chi-squared is 105.067 at p = 3.9e-19. Five splits did not manufacture it.
The two blocks separate cleanly at this depth. MiniRocket’s 95% bootstrap rank range reaches 4.278 and MOMENT’s starts at 5.111, and across all 5,000 bootstrap draws no pretrained model’s mean rank crossed into the convolutional block’s range.
Ten of the twelve transform-against-pretrained pairs survive Holm correction here. The two that do not, MiniRocket over MOMENT at p = 0.108 and Hydra over MOMENT at p = 0.105, both survive on the primary panel. Holm covers this panel’s own family of 36 pairs, so that count and the primary panel’s six survivors answer different questions.
First place is no more settled. RocketPFN takes it in 92.9% of 5,000 dataset-bootstrap draws, against 46.6% and 45.5% on the primary panel, but HIVE-COTE 2, the method it splits first place with there, is absent here. The top clique is still three wide: RocketPFN, MultiRocket and Hydra. MiniRocket falls outside it again, sharing a clique with MOMENT.
This is a depth check, not an independent replication. The 18 datasets contain all 15 of the primary panel plus ElectricDevices, StarLightCurves and UWaveGestureLibraryAll, so the leave-one-dataset-out fragility above stands. It is univariate-only for the same reason, RocketPFN being present. What it rules out is an ordering that depends on five particular splits.
Batch composition changes TimEE’s predictions
TimEE is an in-context classifier whose published headline is first place in ROC-AUC on the UCR benchmark.3 We ran it on two machines and compared cell-level results.
Agreement breaks exactly at the harness batch size of 512. Accuracy is identical on 270 of 270 cells where the test set fits in one call, and on 146 of 325 cells where it does not. For ROC-AUC the split is 221 of 270 against 12 of 308. The two machines differ in hardware in both strata, so neither stratum alone proves anything; the contrast between them is the evidence.
The mechanism is in the source. The statistical embedding function z-scores each feature across every row in the inference call, support and query rows alike, with no guard separating training from evaluation positions. Thirty lines later the same file applies exactly such a guard to a different normalisation step.4 A controlled peer swap confirms it: holding one anchor query fixed and replacing the other 511 queries in its call shifts the anchor’s predicted probability by 1.163e-04, against a repeat-run noise of exactly zero. TiCT and RocketPFN, tested identically, are bit-identical under the same swap.
A TimEE prediction is therefore not a function of the input series alone. Its scores are reproducible only at fixed batch composition, so it cannot receive a like-for-like rank against inductive classifiers and this benchmark reports it without one. This qualifies the published ROC-AUC claim, since ROC-AUC is computed from exactly the probabilities shown to be call-dependent. It does not show the scores are inflated in either direction, it does not change the published rank, which was earned on a leaderboard we do not hold, and it covers the version we inspected.
Three of the four detection routes need no source access. Repeat a cell twice, swap batch peers around a fixed anchor, and compare results across two machines stratified by the batch boundary. Related published work batches prior-fitted-network inference by clustered test queries for efficiency and does not report composition dependence.5
Multivariate panel results
A shortlist built on the broad panel does not carry to multivariate data. MOMENT is last of ten on the 9-dataset multivariate panel while it is the highest-ranked pretrained model on the 24-dataset wide panel at 6.729. Mantis-V2 runs the other way: 5.444 on the multivariate panel against 7.208 on the wide one.
This is not a modality effect. The wide panel contains the multivariate datasets as a subset, so the movement is a contrast between two rosters of different size and task composition, not something measured inside either method. With N=9 and no contrast surviving correction, read it as a warning against carrying a broad-panel shortlist into a multivariate deployment, not as a selection result.
Subject-grouped split results
The UCR and UEA archives ship fixed train/test splits defined over individual series, with no subject or source annotation to split on,6 7 so resampling them reshuffles series rather than subjects. On SleepEDF sleep staging, where one subject contributes many windows, we also split by subject. MiniRocket carries 8 of 10 random folds because of the freeze cutoff.
Random splitting reports higher accuracy, by 0.060 for catch22 and 0.030 for MiniRocket.
The variance compression is the more transferable half. Fold-to-fold variability is roughly seven to nine times smaller under random splitting than under subject grouping, on both window accuracy and subject-macro accuracy. A model selected on random splits looks better and more stable than the same model evaluated the way it will be deployed, on unseen subjects.
We report a split-associated gap, never leakage. The data cannot say whether the gap comes from subject-identity leakage or from genuine variation in subject difficulty and stage mix, and random splits distribute that variation across every test set, which stabilises accuracy mechanically. The experiment that would separate the two, identical test epochs under training sets that do and do not contain the test subjects, ran on PTB-XL and was dropped from SleepEDF for compute. PTB-XL itself retains 14.9% same-subject overlap under random splitting, so it is a low-overlap comparator, not a null control.
Imbalanced data: Earthquakes
Accuracy conceals near-chance ranking. Every CPU-lane method reports 0.72 to 0.78 accuracy on Earthquakes while balanced accuracy runs from 0.4996 to 0.5453. Read 0.5 as chance on the balanced-accuracy axis only; chance for ordinary accuracy depends on class prevalence.
The method with the highest accuracy has the lowest balanced accuracy. catch22 at 0.7835 accuracy scores 0.4996 balanced, numerically just below 0.5. Whether it is literally predicting the majority class was not measured; the figure quoted carries no uncertainty and no test against 0.5.
The reversal is local, not a property of the method. Across the full panel catch22 is clearly last on balanced accuracy, 0.7260 against MultiRocket’s 0.8062. Balanced accuracy is a post hoc metric here, computed after the freeze from archived predictions. DTW-1NN carries 5 resamples against the other four methods’ 30, so its point rests on less evidence.
Recorded compute
Recorded compute totals 1,496.5 core-hours, all on the CPU lane. All times are loaded-throughput under varying concurrency, not solo latency.
HIVE-COTE 2 accounts for 804.8 of the total, of which 694.2 covers its 140 ranked cells and the remainder four post hoc diagnostic cells excluded from every table. MultiRocket accounts for 26.2 over 990 ranked cells. The two ranked totals cover different amounts of work, so no ratio between them is a like-for-like cost comparison.
No cost comparison between classical methods and foundation models is possible from this benchmark, and we make none. The six GPU-lane methods recorded no wall time. TimEE has recorded wall time but is excluded from the chart because it carries no rank.
Methods compared
HIVE-COTE 2
First on mean rank (2.400) on the primary panel, not separable from RocketPFN. Ran 140 ranked cells for 694.2 recorded core-hours, the most expensive method on the recorded CPU lane, and did not finish FaceDetection (144 channels). A heterogeneous meta-ensemble over shapelet, dictionary, interval and convolutional representations.8
RocketPFN
First on mean accuracy (0.9047), not separable from HIVE-COTE 2. Bit-identical under the peer-swap test. Univariate-only, so it never enters the UEA panel. Combines ROCKET-family convolutional features with in-context classification by the tabular foundation model TabPFN v2.5, so no model is fitted on the target dataset; its paper reports parity with HIVE-COTE 2 at 0.900 mean accuracy on 92 UCR datasets.9
MultiRocket
Third on mean rank (2.800) and the strongest pairwise record: all four of its inspected contrasts against pretrained models survive Holm, two at 28 wins to 0. 26.2 recorded core-hours for 990 cells. Extends MiniRocket with multiple pooling operators and first-order differences.10
Hydra
Mean rank 4.100; its win over MOMENT (22/6) survives correction. A dictionary-style convolutional method: kernels compete in groups and the counts of winning kernels form the representation.11
MiniRocket
Mean rank 4.700 at 4.9 recorded core-hours for 990 cells, the cheapest strong method on the recorded CPU lane; its win over MOMENT (24/4) survives correction. A near-deterministic random convolutional kernel transform with a linear classifier.12
MOMENT
Highest-ranked pretrained model on the univariate primary panel (6.633); loses to all three convolutional transforms with surviving contrasts. A 385M-parameter T5-based model pretrained on a multi-domain corpus for forecasting, classification, anomaly detection and imputation.13
Chronos-2
Mean rank 7.967; loses to MultiRocket 28/0. A forecasting foundation model used here as a frozen embedding extractor, which is not its native task.14
TiRex
Mean rank 8.433; loses to MultiRocket 28/0. An xLSTM-based zero-shot forecasting model,15 also used here as a frozen extractor. In the post hoc sensitivity arm, changing only final-layer extraction to all-layer concatenation raised mean accuracy by 1.13 percentage points across its 30-dataset panel: 23 datasets improved and seven worsened, Heartbeat by 7.98 points with none of its 30 resamples improving. The frozen final-layer result and rank remain primary; the sensitivity arm carries no rank.
Mantis-V2
Mean rank 9.733 on univariate and 5.444 on the multivariate panel; loses to MultiRocket 24/3. Did not finish FaceDetection: 144 channels overwhelm its native classification head. A lightweight contrastively pretrained classification model with synthetic data and test-time strategies added in V2.16
TimEE
Reported without a rank; see the batch-composition finding above. Mean accuracy 0.8410 over the 21 datasets in its reported panel, computed at fixed batch composition. Two exclusions declared before results apply: NonInvasiveFetalECGThorax1 is partial at 5 of 30 resamples, and Phoneme is reported at its archive split only. A 4.5M-parameter model meta-trained on synthetic tasks that classifies end-to-end in-context, in a single forward pass from a labelled support set.17
TiCT
Mean rank 10.200. Univariate-only; its runner was the only one to enforce the rare-class validity rule on Phoneme, contributing one valid cell where other lanes ran 30. Bit-identical under the peer-swap test. An in-context transformer pretrained on synthetic data with bit-based label encoding for arbitrary class counts.18
DTW-1NN
Mean rank 9.900. Did not finish StarLightCurves, a declared cost DNF at 95 hours for one unsplittable cell. One-nearest-neighbour under Dynamic Time Warping distance, the longest-standing baseline in the field.
catch22
Mean rank 8.700 on the primary panel. Twenty-two fixed, interpretable summary features feeding a simple classifier.19
When to choose a foundation model over a classical classifier
- Plenty of labelled training data: every contrast that survives correction here favors a convolutional transform, and two of the three, MultiRocket and Hydra, sit in the top clique alongside HIVE-COTE 2 and RocketPFN.
- Severely limited labels or cold-start deployment: in-context and zero-shot models are the candidates, and this is the regime our benchmark does not measure. All our panels use full archive training splits.
- Multivariate data: check native coverage first. Two of the three in-context methods in this roster are univariate-only.
- Highest accuracy regardless of compute: HIVE-COTE 2 and RocketPFN are not separable on this evidence. This benchmark separates neither their accuracy nor their compute, because RocketPFN has no recorded timing.
- Auditability requirements: prefer methods whose predictions are reproducible from the input alone. The batch-composition finding above is the cautionary case.
What is time series classification
Time series classification trains a model on labelled, ordered numeric sequences to predict a discrete class for an unseen series: ECG diagnosis, human activity recognition from wearables, industrial fault detection. It differs from forecasting, which outputs future values rather than a class label. Tabular models do not encode the ordering of the values as an inductive bias; they treat each timestep as an independent column.
Classical methods fit a new model per dataset. A time series foundation model is pretrained once on a large corpus, then applied to new datasets zero-shot, through a lightweight probe, or in-context from a handful of labelled examples supplied at inference time. Whether that pretraining currently buys classification accuracy on standard archives, under one controlled protocol, is the question this benchmark measures.
Aggiungi come fonte preferitaTime series classification benchmark methodology
Methods. Four families: pretrained foundation models evaluated through a transfer protocol (MOMENT, Chronos-2, TiRex, Mantis-V2), in-context models (TimEE, TiCT, RocketPFN), convolutional transforms (MiniRocket, MultiRocket, Hydra), and classical references (HIVE-COTE 2, DTW-1NN, catch22).
Data. 23 UCR univariate and 10 UEA multivariate datasets,20 21 plus a separate non-IID layer on SleepEDF and PTB-XL. 14,638 recorded cells, split in the coverage table below.
Coverage. Scored cells are not evenly distributed across methods, and the denominators decide which comparisons are possible at all.
Every ranked comparison runs on the complete-case intersection of the methods it names, never on a single method’s own coverage, which is what lets HIVE-COTE 2’s 140 cells and RocketPFN’s 2,760 appear in the same table. The full export holds 14,638 rows: these 11,874 scored cells, RocketPFN’s 2,760-cell paper-protocol reproduction arm, and 4 HIVE-COTE 2 post hoc diagnostic cells.
Resampling. r0 is the archive train/test split; r1 to r29 are stratified shuffles at the archive proportions, seeded 1729+r. The inferential unit is the dataset: resamples are averaged within a dataset before methods are compared. HIVE-COTE 2 and DTW-1NN run 5 resamples because their per-cell cost is one to two orders of magnitude higher, and the primary table restricts every method to the same 5 matched resamples. The archive split r0 is systematically harder than seeded resamples on 8 of 33 datasets, by up to 0.218 accuracy, so the five matched resamples are not exchangeable draws.
Transfer protocol. Every reference foundation model runs the same pipeline: per-variate encoding, final encoder layer, mean-pool over valid context tokens, per-variate L2 normalisation, channel concatenation, train-only standardisation, logistic regression probe. Frozen 2026-07-24, before any scored cell.
Statistics. Friedman omnibus, then pairwise Wilcoxon signed-rank with Holm correction over all 19 inspected contrasts, on dataset-level means, complete-case panels only. The omnibus supports only that ranks differ somewhere; every pairwise statement comes from the corrected contrasts.
Freeze. Data froze 2026-07-29 03:00 UTC, with exclusion rules written before the outstanding results were known. One cell finished 22 minutes after the cutoff and is excluded under the pre-written rule. The release passes 80 mechanical assertion checks, including hash comparison of every input file, regeneration of the primary table from the export, and enforcement of the declared exclusions behind the published TimEE figure in both the article and the release tables.
Metrics. ROC-AUC was the declared primary metric. It is unavailable for 5 of the 7 reference foundation models, whose probe archives decision scores rather than calibrated probabilities, so accuracy became the comparison metric. We state this rather than relabelling accuracy as pre-specified. Balanced accuracy, macro-F1 and log-loss are post hoc sensitivity analyses on the covered subset.
Reproducibility. Per-instance predictions are archived for all 13 methods and all 12 of the primary table; every regenerated prediction reproduces its frozen scalar at a maximum difference of 0.000e+00 (4,260 CPU-lane cells re-verified 2026-08-02, zero orphans). The primary table is fully recomputable from released predictions.
Methods not in this benchmark
Published TSC-relevant models we did not evaluate, listed so the roster’s boundary is explicit. None of the statements here are our measurements. UniTS is a unified multi-task model covering classification, forecasting, imputation and anomaly detection.22 TiViT converts series to images and classifies with frozen pretrained Vision Transformers.23 UniShape is a shape-aware foundation model built for classification, with a shape-aware adapter, accepted at AAAI 2026.24 GPT4TS, also published as One Fits All, uses a GPT-2 backbone whose self-attention and feedforward blocks stay frozen while the input, normalization and output components are trained.25 TIC-FM is an in-context zero-shot classifier predicting all test instances in a single forward pass;26 reading its single-call design as the same exposure class as the batch-composition finding above is our inference from the published description, not a tested result.
Limitations
- No pretraining-overlap ledger exists for any pretrained model: none publishes a dataset-level statement of whether UCR/UEA series appear in its pretraining corpus, so transfer cannot be separated from prior exposure. Treating the losses as safe rests on an assumption that no dataset-level ledger can test: that prior exposure, where it occurred, flattered these models rather than hurt them. Under that assumption they trail despite the advantage and the losses stand; any pretrained-model win, by the same logic, cannot be cleanly credited to transfer.
- Foundation-model results are conditional on the single frozen transfer protocol. The post hoc TiRex arm changed only layer extraction and raised mean accuracy by 1.13 percentage points while worsening seven of its 30 datasets, Heartbeat by 7.98 points with 0 of its 30 resamples improving. It tested one model, carries no rank and does not establish the effect for the other foundation models.
- All wall times are loaded-throughput under varying concurrency on a shared machine. Six of 13 methods carry none.
- Evidence covers archive-sized problems only. Nothing here speaks to scalability.
- The benchmark inspects one version of each method at its frozen date.
Common pitfalls in benchmarking time series classification
Pretraining-test overlap. A 2025 audit traced 401 datasets through 22 published time series foundation models and found only 6% had never appeared in any model’s pretraining or fine-tuning corpus.27 No model in our roster publishes a dataset-level overlap statement, which is why our limitations section treats transfer and prior exposure as inseparable.
Split design. The standard UCR archive ships one fixed train/test split; modern practice resamples, and our SleepEDF panel shows why subject structure needs more than resampling: split design alone moved reported accuracy by up to 0.060 and compressed its variance roughly seven to nine fold. No single figure here is a property of the dataset; each one is a property of the dataset under a stated split design.
Multiplicity. Comparing many classifiers across many datasets without correction manufactures findings. The standard machinery is Wilcoxon signed-rank for pairs and Holm-style familywise correction over every contrast inspected,28 29 and correction must cover every contrast inspected, not only those reported. In this benchmark that discipline eliminated 13 of 19 contrasts.
Appendix: TiRex extraction sensitivity
This post hoc arm uses the same 30 datasets, 30 resample splits, scaler and logistic probe as the frozen TiRex result. Only feature extraction changes, from the final encoder layer to concatenated representations from all 12 layers. The arm is rank-gratuito and does not alter the frozen benchmark.
Mean accuracy moved from 0.7933 to 0.8046, a 1.13 percentage-point change, with a seeded dataset-bootstrap interval of +0.15 to +2.06 points. A paired Wilcoxon signed-rank test over the 30 dataset deltas gives an exact two-sided W of 92.0 at p=0.002987. That test is exploratory: it was run after the freeze, it sits outside the 19-contrast Holm family that governs every corrected claim here, and it carries no rank.
Auer et al. rank TiRex first among forecasting foundation models using layer-concatenated features and a Random Forest, and their own ablation reports a larger 5-point layer effect.30 The two studies differ in datasets, model roster, classifier and series-length filtering, so 1.13 and 5 points are protocol-specific sensitivity estimates rather than competing estimates of one effect. This arm neither reproduces nor refutes their ranking, and does not identify the cause of the cross-study rank difference. It establishes layer sensitivity for TiRex inside this benchmark. It does not generalise the effect to the other foundation models or identify any model’s best attainable classification performance.
Cita questa ricerca
Scegli il formato adatto a dove pubblicherai. Incollare la versione con link nel tuo CMS preserva il backlink.
@misc{kalelioglu2026,
author = {Kalelioğlu, Berk},
title = {{Time Series Classification Benchmark: Foundation Models vs Classical Methods}},
year = {2026},
month = aug,
howpublished = {\url{https://aimultiple.com/time-series-classification}},
note = {AIMultiple. Consultato il 6 Agosto 2026}
}Collegamenti di riferimento
Sii il primo a commentare
Il tuo indirizzo email non verrà pubblicato. Tutti i campi sono obbligatori. I commenti vengono lasciati nella loro lingua originale.