twelvedata/financial-world-model

Twelve Data World Model Dataset

A multi-modal financial time-series dataset built from Twelve Data market data. Each timeframe is published in three parallel views:

• bars_* — OHLCV bars enriched with causal technical indicators and macro context, in Parquet.
• text_* — instruction-tuning prompts/labels derived from the bars, in JSONL.
• trajectories_* — fixed-length rolling windows of state vectors plus next-state pairs, suitable for world-model / sequence-model training, in Parquet.

All views cover the same symbol universe and share the same time-based train / validation / test splits.

Splits

Split	Date range
train	up to 2023-12-31
validation	2024-01-01 → 2024-12-31
test	2025-01-01 → 2026-05-13

Splits are assigned by timestamp. For trajectories_*, a window is dropped if it crosses a split boundary (so train and val never overlap). Datetimes are tz-aware in America/New_York.

Symbol universe

51 large-cap US equities spread across sectors (Tech, Financials, Healthcare, Consumer, Industrials/Energy/Materials, Communication). Macro context tickers (SPY, QQQ, VIXY, TLT, sector SPDRs, etc.) are joined as columns onto every equity row — they are not first-class trainable symbols.

See config/symbols.yaml in the source repo for the exact list.

Timeframes

The pipeline requests the full history available from Twelve Data per symbol; actual depth varies by ticker and timeframe and is bounded by the vendor's historical limits. As a rough guide:

Interval	Typical depth (older names like AAPL/MSFT)	Trajectory windows (size / stride)
1day	several decades back to listing	30/5, 60/10, 120/20
1h	a few years	24/6, 120/24
1min	a few years	390/195 (one US session), 1950/390 (one week)

Newer listings naturally start later. Each row's datetime reflects what the vendor returned — there is no synthetic backfill before a symbol's inception.

bars_* schema

OHLCV plus deterministic, causal indicators and macro joins. Each row is one bar for one symbol at one timeframe.

Core

column	type	description
datetime	timestamp[ns, America/New_York]	bar timestamp (interval-start)
symbol	string	equity ticker
timeframe	string	1day, 1h, or 1min
open, high, low, close	float64	OHLC prices
volume	float64	traded volume
close_adj	float64	split- and dividend-adjusted close (== close when no events)

Returns

column	description
ret_1	1-bar simple return
logret_1	1-bar log return
ret_5	5-bar simple return
logret_20	20-bar log return

Volatility

column	description
rv_5, rv_20, rv_60	realized volatility (stdev of log returns) over 5/20/60 bars
atr_14	Average True Range, 14 bars

Momentum

column	description
rsi_14	Relative Strength Index, 14 bars
macd, macd_signal, macd_hist	MACD(12,26,9) line, signal, histogram
mom_10	10-bar momentum

Volume

column	description
obv	On-Balance Volume
vol_z_20	volume z-score over 20 bars

Bands

column	description
bb_mid, bb_up, bb_lo	Bollinger Bands(20, 2σ) middle/upper/lower
bb_pctb	%B position within the band

Macro context (joined by date)

column	description
spy_logret_1	SPY 1-bar log return
vix_level	VIX level (proxied via VIXY)
tlt_logret_1	TLT (20+ yr Treasury) 1-bar log return
dxy_logret_1	US Dollar index 1-bar log return (UUP proxy)
sector_logret_1	sector ETF 1-bar log return matching the symbol's sector

Causality invariant: every indicator and macro column at row t uses only information from rows ≤ t. Tests in tests/test_indicators_causal.py enforce this in CI.

text_* schema

Instruction-tuning records derived from bars_*.

field	type	description
symbol	string	equity ticker
timeframe	string	1day / 1h / 1min
as_of	string (ISO datetime)	the bar's timestamp; everything in prompt is dated ≤ this
prompt	string	natural-language description of the bar + already-observed indicators
label	string	next-bar outcome (direction + log return); empty for the final bar
meta	object	small bag for provenance (e.g. row index)

Leak-prevention invariant: prompt never references information dated strictly after as_of. Only label carries the next-step outcome. This is enforced in tests/test_textify.py.

trajectories_* schema

Rolling windows over the bar series, suitable for world-model and sequence-model training. Each row is one window.

field	type	description
trajectory_id	string	stable id ({symbol}_{timeframe}_{window}_{start_ts})
symbol	string	equity ticker
timeframe	string	1day / 1h / 1min
feature_names	list[string]	column order of the state vector (length F)
timestamps	list[string]	T tz-aware ISO datetimes
states	list[list[float64]]	shape (T, F)
next_states	list[list[float64]]	shape (T, F), shifted by one step
rewards_logret	list[float64] | null	optional scalar reward stream (1-bar log return); null means reward-agnostic
split	string	train / val / test

The state vector is fixed across all trajectories within the dataset:

open, high, low, close_adj, volume,
logret_1, rv_20, rsi_14, macd, macd_signal, macd_hist, atr_14, bb_pctb,
obv, vol_z_20,
spy_logret_1, vix_level, tlt_logret_1, dxy_logret_1, sector_logret_1

How to load

from datasets import load_dataset

# OHLCV + indicators, daily bars
bars = load_dataset("twelvedata/financial-world-model", "bars_1day")

# Instruction-tuning text, hourly
text = load_dataset("twelvedata/financial-world-model", "text_1h")

# Trajectories for world-model training
traj = load_dataset("twelvedata/financial-world-model", "trajectories_1day")

For ad-hoc analytics, the parquet files are queryable directly with DuckDB:

SELECT symbol, datetime, close, rsi_14
FROM 'bars_1day/test.parquet'
WHERE symbol = 'AAPL'
ORDER BY datetime DESC
LIMIT 10;

Tip on the in-browser SQL Console: DuckDB-WASM streams parquet over HTTPS without local caching, and 1min configs are millions of rows. Always project explicit columns (SELECT symbol, datetime, close, ...) instead of SELECT *, and prefer the 1day configs for quick exploration. For real work, download the parquet and query it from desktop DuckDB.

Refresh cadence

The dataset is rebuilt by an incremental pipeline that:

1. Fetches the trailing window per symbol/timeframe (re-fetches the previous day to capture restatements).
2. Detects splits and dividends and triggers a per-symbol re-backfill when needed (so close_adj stays correct historically).
3. Recomputes indicators and macro joins.
4. Re-emits all three views and pushes them here.

Limitations

• US equities only; intraday data is regular-session only (no pre/post).
• Macro context is ETF-proxied (e.g. VIXY for VIX, UUP for DXY) — convenient to fetch but not identical to the underlying index.
• Intraday history depth (especially 1min) is bounded by Twelve Data's vendor limits and is much shorter than daily history. Don't assume identical date coverage across timeframes for the same symbol.
• The text view is templated, not LLM-generated — it is dense and repetitive by design, intended as a substrate for fine-tuning rather than as human-style prose.

License

MIT. Underlying market data is © Twelve Data and redistributed under their terms; check twelvedata.com for commercial use.

Citation

@misc{twelvedata-world-model,
  title  = {Twelve Data World Model Dataset},
  author = {Twelve Data},
  year   = {2026},
  url    = {https://huggingface.co/datasets/twelvedata/financial-world-model}
}