The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: Schema at index 1 was different:
schema_version: string
id: string
version: string
sha256: null
mode: string
track: string
freq: string
horizons: list<item: int64>
num_nodes: null
num_channels: int64
target: string
hf_path: string
hf_revision: null
source_config: string
vs
symbols: list<item: string>
timeline_start: timestamp[s]
timeline_end: timestamp[s]
T: int64
N: int64
C: int64
channels: list<item: string>
seq_len: int64
pred_len: int64
window_months: int64
val_tail_ratio: double
trainval_end: timestamp[s]
test_start: timestamp[s]
n_train: int64
n_val: int64
n_test: int64
normalization: string
adj_mx: struct<method: string, data_range: string, shape: list<item: int64>>
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 243, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 4376, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2658, in _head
return next(iter(self.iter(batch_size=n)))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2836, in iter
for key, pa_table in ex_iterable.iter_arrow():
^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2374, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 580, in _iter_arrow
yield new_key, pa.Table.from_batches(chunks_buffer)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "pyarrow/table.pxi", line 5039, in pyarrow.lib.Table.from_batches
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
pyarrow.lib.ArrowInvalid: Schema at index 1 was different:
schema_version: string
id: string
version: string
sha256: null
mode: string
track: string
freq: string
horizons: list<item: int64>
num_nodes: null
num_channels: int64
target: string
hf_path: string
hf_revision: null
source_config: string
vs
symbols: list<item: string>
timeline_start: timestamp[s]
timeline_end: timestamp[s]
T: int64
N: int64
C: int64
channels: list<item: string>
seq_len: int64
pred_len: int64
window_months: int64
val_tail_ratio: double
trainval_end: timestamp[s]
test_start: timestamp[s]
n_train: int64
n_val: int64
n_test: int64
normalization: string
adj_mx: struct<method: string, data_range: string, shape: list<item: int64>>Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
TSEval — RealTime Datasets
This repository holds the RealTime datasets for TS-Eval, the open, reproducible leaderboard for time-series forecasting. Unlike static benchmarks, the datasets here are periodically refreshed live data — they grow as time passes and new observations land.
The first RealTime dataset is stock_hs300: the CSI-300 (沪深300) index and its constituents.
The RealTime track
Static benchmarks are fixed once and never move. That is good for comparability, but it also means a model can quietly overfit to a frozen test set that everyone has seen.
The RealTime track is different. Each dataset is a live series that we refresh over time. New trading days are appended, and the leaderboard re-runs against data that did not exist when a model was first trained. A forecaster that genuinely generalizes keeps its place; one that memorized the past does not.
This makes RealTime a standing, moving target — closer to the conditions a deployed forecaster actually faces.
stock_hs300 (CSI-300 / 沪深300)
The CSI-300 index tracks the 300 largest and most liquid A-share stocks listed on the Shanghai and Shenzhen exchanges. This dataset captures the daily price and volume series of its constituents.
Files
| File | Contents |
|---|---|
close.csv |
Daily closing price per symbol. Rows are trading days, columns are constituent symbols. |
volume.csv |
Daily traded volume per symbol. Same row/column layout as close.csv. |
meta.json |
Constituent metadata — roughly 300 symbols, with their identifiers. |
timeline.json |
The ordered list of trading days, beginning 2019-01-02. |
Calendar
The series follows the exchange trading calendar, not the wall-clock calendar — weekends and market holidays are absent. timeline.json is the single source of truth for the day index; close.csv and volume.csv are aligned to it row by row. Treat the calendar as irregular: do not assume a fixed five-day week or fill gaps with synthetic dates.
Protocol
The first leaderboard round used a single, fixed protocol:
- Input —
seq_len = 20trading days. - Predict —
pred_len = 5trading days (horizon 5). - Ranked by — MSE, lower is better.
This was the protocol behind the "百模大战 / Hundred-Model Battle" launch round: 108 models, 124 submissions (some models run on multiple seeds). More horizons and refresh cycles will follow.
IMPORTANT — data provenance and usage
Read this before you use the data.
- This is market price and volume data for CSI-300 constituents, assembled for non-commercial research and benchmarking purposes only.
- Exchange data and index constituent data may be subject to the terms, licensing, and restrictions of the originating exchanges, index providers, and data vendors. You are responsible for verifying that you hold the rights to use this data for your intended purpose, especially any commercial, redistributive, or production use. The MIT license on this repository covers the packaging and structure of the dataset card, not any third-party rights in the underlying market data.
- The data is provided as-is, without warranty of any kind — no guarantee of accuracy, completeness, timeliness, or fitness for any purpose. Corporate actions, symbol changes, survivorship effects, and vendor errors may be present.
- Nothing here is financial advice. This is a research benchmark, not a trading signal.
If you cannot confirm your right to use exchange data for your use case, do not use it.
What we found (and what we didn't)
The "百模大战 / Hundred-Model Battle" launch round ranked all entries by MSE. The top three:
| Rank | Model | MSE |
|---|---|---|
| 1 | NBeats | 0.7483 |
| 2 | DeepAR | 0.7502 |
| 3 | Informer | 0.7503 |
The spread across the very top of the leaderboard is tiny — the top three sit within about 0.748 to 0.750 MSE — even though their architectures differ wildly. On noisy financial returns, classic and simpler forecasters led, and elaborate architecture bought little. We frame this honestly as near-efficient-market: the leaderboard is as useful for showing what does not help as what does.
This was a launch snapshot, not a final verdict — mostly single-seed (seed 2024), a single horizon, and one round. More datasets, more horizons, and the RealTime refresh are coming. See the live leaderboard for current numbers.
The TSEval repositories
| Repo | What it holds |
|---|---|
| TSEval-Static | Static benchmark datasets (time_series / spatiotemporal / covariate). |
| TSEval-RealTime | This repo — periodically-refreshed live datasets. |
| TSEval-Submissions | Append-only evidence bundles (trajectory + result + report). |
| TSEval-Weights | Trained checkpoints, referenced from submissions by sha256. |
| TSEval Space | The leaderboard frontend. |
- Live leaderboard on the site: diaugeia.ai/tseval
- Producer framework: ModernTSF
Maintained by the Diaugeia.AI team. Keep research Simple and Stupid.
- Downloads last month
- 61