Internal benchmark for testing time-series / cross-sectional forecasting models on strictly point-in-time US equity data at three frequencies (daily / 30-minute / 1-minute). Built 2026-09; frozen.

LICENSE / SHARING NOTICE: bars are derived from licensed market data (CRSP and a commercial intraday vendor). Do NOT redistribute publicly. Private, authorized collaborators only.

1. What the data is

Core inputs are raw/adjusted price-volume bars rather than a pre-engineered feature matrix; you build your own features/representations. (A one-bar return bar_ret is included in intraday files only as a convenience field.) Prices carry corporate-action-consistent CRSP adjustment (cumulative price factors; cash dividends are not embedded). Raw unadjusted prices and the adjustment factors are also kept in the intraday files.
Targets = precomputed RETURNS (the official labels), so everyone scores the exact same quantity.
Universe = point-in-time S&P 500 common-stock membership (no survivorship bias): a stock is eligible at an origin only while its membership spell is active (member_flag in daily bars, spells in universe.csv).
SPY reference series are provided separately as a market proxy (SPY is an ETF and is not part of the stock universe).

Scale	Bars coverage	Origins	Target (h = 1..10)
daily	2010-2025 (2010-15 = lookback history only)	each trading day t at the open	AdjClose(t+h-1)/AdjOpen(t) - 1
30-minute	2016-04 .. 2025	start of bar s, slots 2..13	Close(s+h-1)/Close(s-1) - 1, same session
minute	2016 .. 2025	start of minute t, tmin 1..389	Close(t+h-1)/Close(t-1) - 1, same session

Official split (splits.json): train 2016-2021, val 2022, test 2023-2025. This is a frozen evaluation split, not a hidden held-out set: test labels are distributed for convenience, so "no tuning on 2023-2025" is an honor-system rule. 2026 is fully reserved for blind evaluation and is the final arbiter - never touch it.

2. Files and fields

README.md / README.zh-CN.md      this documentation
universe.csv                     PIT membership spells + ticker history
splits.json                      official split definition
evaluate.py                      official scorer
pit_loader.py                    official PIT-safe daily loader (see Rules)
bars_only_reference.csv          reference results reproducible from THIS dataset
external_reference.csv           informational references using extra data (see below)
daily/bars.parquet               daily bars, all stocks, 2010-2025
daily/targets.parquet            daily official targets
30minute/bars/year=Y/TICKER.parquet    30-minute bars per stock-year
30minute/targets/year=Y.parquet        30-minute official targets (all origins)
minute/bars/year=Y/TICKER.parquet      1-minute bars per stock-year (86 GB total)
minute/targets_test/year=Y.parquet     1-minute official TEST targets (2023-2025)
minute/make_targets.py                 generate train-period minute targets from bars
spy/spy_daily.parquet, spy_30m.parquet, spy_minute.parquet,
spy/spy_daily_features.parquet         market reference series (see notes below)

daily/bars.parquet - one row per (stock, trading day). Example (NVDA, 2024-06-03):

permno	date	ticker	adj_open	adj_high	adj_low	adj_close	adj_volume	member_flag
86580	2024-06-03	NVDA	113.62	115.00	112.00	115.00	435,766,336	True

permno: CRSP permanent security id (the primary key; survives ticker changes).
adj_*: adjusted open/high/low/close prices and share volume.
member_flag: True while the stock is an S&P 500 member (PIT eligibility).

daily/targets.parquet - one row per (stock, day) with y_h1..y_h10. Same example: y_h1 = 0.0121 means Open(6/3) -> Close(6/3) realized +1.21%; y_h5 = 0.0640 is Open(6/3) -> Close(6/7). NaN when the horizon extends past the data end.

30minute / minute bars - one row per (stock, bar). Columns: permno, ticker, window_start_et, window_start_utc, date (keys/time), bar_open/high/low/close, bar_vwap, bar_volume, transactions (raw vendor bar), dlycumfacpr, dlycumfacshr (CRSP adjustment factors), adj_open/high/low/close, adj_vwap, adj_volume (adjusted bar), bar_ret (adjusted close-to-close return vs previous bar, same day; NaN on the first bar of a session). Example minute bar (NVDA 2024-06-03 09:31): adj_close 113.304, adj_vwap 113.308, volume 2.88M, 7,756 trades, bar_ret -0.026%.

Minute-bar timing convention: window_start_et = 09:30:00 is the first session bar (tmin = 0); tmin = (hour-9)*60 + minute - 30. 30-minute bars: 13 per session, slot = 1..13, slot s spans [09:30 + (s-1)*30min, +30min).

30minute/targets - one row per origin (permno, date, slot=2..13) with y_h1..y_h10; horizons crossing the session end are NaN (so h10 exists only for slots 2..4). minute/targets_test - one row per origin (permno, date, tmin=1..389), same convention; ~40M rows/year. For train-period minute targets run python minute/make_targets.py 2019 (official formula, from bars).

spy/: spy_daily.parquet (date, adjusted OHLCV, day_total_return incl. dividends - ex-post for that day: at the day-t open you may only use rows <= t-1); spy_30m.parquet (date, slot, spy_open/close/volume, spy_bar_ret; one row per completed bar); spy_minute.parquet (one row per completed minute: date, tmin, spy_ret_1m); spy_daily_features.parquet (date, spy_prev_day_ret, spy_prev_day_rv - already shifted one day, safe to join as same-day inputs).

3. How to read the files

Everything is standard Parquet (pandas/pyarrow, polars, DuckDB, Spark all work):

import glob, pandas as pd
bars = pd.read_parquet("daily/bars.parquet")                       # one file
y30  = pd.read_parquet("30minute/targets/year=2024.parquet")
nvda = pd.read_parquet("minute/bars/year=2024/NVDA.parquet",
                       columns=["permno","date","window_start_et","adj_close","adj_volume"])
year30 = pd.concat(pd.read_parquet(f)                              # a whole year of 30m bars
                   for f in glob.glob("30minute/bars/year=2024/*.parquet"))

Tip: pass columns=[...] to read only what you need; join stock files on permno (not ticker). File names use the ticker as of the file's year for convenience only.

4. Rules (PIT) and scoring

At an origin you may use any bar information strictly up to the anchor (the most recently completed bar; for daily, the day-t open itself is allowed). Daily caution: the bars row for day t also contains high/low/close/volume, which are NOT known at the open of t - use pit_loader.load_daily_window() (official safe loader) or replicate its visibility rule exactly.
Eligible cross-section at an origin = stocks with active membership.
Two official tracks - report which one you used; results are not comparable across tracks and must not share a leaderboard:
- Track A - Standard Fixed Split: train 2016-2021, val 2022, test 2023-2025. For fast model/architecture/hyperparameter comparison.
- Track B - Rolling PIT: for each test year Y, fit on data through Y-1 and predict Y (annual expanding walk-forward; our internal production protocol). For production-style long-horizon studies. In both tracks the test window and metric are fixed; no tuning on test.

Produce test-window predictions as a parquet with keys (permno, date [, slot | tmin]) and columns pred_h1, pred_h3, pred_h5, pred_h10, then:

python evaluate.py --scale daily  --pred my_daily_preds.parquet
python evaluate.py --scale 30m    --pred my_30m_preds.parquet
python evaluate.py --scale minute --pred my_minute_preds.parquet

The scorer reports per-horizon cross-sectional Spearman RankIC, AvgIC (mean over h in {1,3,5,10}), yearly AvgIC, and positive quarters.

Reference numbers (test 2023-2025 AvgIC). Beatable with THIS dataset (bars_only_reference.csv): daily best 0.0309 (linear_v2); 30-minute best 0.0316; minute best 0.0814 - these are bars-derivable models. Separately, external_reference.csv lists models that use additional PIT information not distributed here (sector data, accounting characteristics, firm-characteristic panels), including the final delivered daily pipeline at 0.0349 - informational only, not a fair bars-only target.

5. Known caveats

Minute-scale predictability is dominated by first-minute microstructure reversal; a high minute IC does not imply executable alpha.
30-minute h10 exists only for morning origins (slots 2..4) by construction.
Daily bars before 2016 are lookback history only; supervised training begins in 2016 and the official test window is 2023-2025.

Downloads last month: 211

Total file size:

94 GB