Finance1K

Finance1K is a synchronized hourly panel for next-step forecasting across 1,000 US equities. It contains 28,510 chronological observations from 2015 to early 2026 and provides two aligned targets:

• close: hourly close log return, log(close_t / close_t-1);
• volume: hourly log-volume change, log(volume_t / volume_t-1).

The two configurations use the same timestamps, ticker order, validity mask, and chronological split. The first 19,957 rows form the training set and the following 8,553 rows form the test set. The paper protocol uses the previous 96 rows to predict the next row and evaluates every fourth test target.

Data structure

Each Parquet row is one timestamp and one synchronized cross-section. The values and valid fields are fixed-size arrays of length 1,000. Array position j matches selection_order j in metadata/tickers.parquet.

Field	Type	Meaning
row_index	int64	Global chronological row index
timestamp	int64	UTC Unix timestamp in seconds
values	float32[1000]	Close or volume target cross-section
valid	bool[1000]	Entry-level validity mask before zero fill
xs_vol	float32	Cross-sectional population standard deviation
regime	int8	Calm/stress indicator fitted on training rows
lagged_vol	float32	Causal 24-row trailing volatility proxy
ceiling_series	float32	Training-normalized predictability proxy

The metadata directory records the full schema, exact split boundaries, ticker order and volatility buckets, per-panel scalar values, canonical array identities, and SHA-256 checksums.

Loading

The close and volume configurations can be read independently:

from datasets import load_dataset

close = load_dataset("abel-lab/finance1k", "close")
volume = load_dataset("abel-lab/finance1k", "volume")

The synchronized arrays must remain intact when evaluating cross-sectional metrics. Flattening ticker and timestamp axes changes the empirical task.

Known limitations

The stored timeline is a regular calendar-day hourly grid rather than an exchange-session calendar. Weekend and market-holiday rows are filled with zero log changes.

The ticker universe is limited to names with sufficiently long historical coverage to span the common evaluation window. Eligible tickers are divided into ten historical-volatility buckets, with 100 selected from each bucket, to balance the panel across volatility levels. This construction introduces survivorship and selection effects and should not be interpreted as a representative market sample.

License

Finance1K is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0).