Datasets:

Flinchlab
/

sec-8k-market-reaction-dataset

dataset_file stringclasses 3 values	field_name stringlengths 6 28	display_name stringlengths 6 28	data_type stringclasses 4 values	required stringclasses 1 value	definition stringlengths 11 331	unit_or_format stringclasses 8 values	allowed_values stringlengths 9 91 ⌀	missing_value_rule stringclasses 2 values	example float64
event_timing_aggregate.csv	bucket	Bucket	text	yes	Which slice of events this row aggregates	text	ALL; Large; Mid-size; Small; filing-time and disclosure-lag slices	never missing	null
event_timing_aggregate.csv	bucket_type	Bucket type	text	yes	What kind of slice `bucket` is	text	all \| company_size \| filing_time \| disclosure_lag	never missing	null
event_timing_aggregate.csv	item_code	Item code	text	yes	Official SEC 8-K item code, or ALL for the pooled row	text	e.g. 2.02	never missing	null
event_timing_aggregate.csv	item_label	Item label	text	yes	Plain-language name for the item code	text	e.g. Results of operations (earnings)	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	window	Window	text	yes	Event-time window in trading-day offsets, day 0 = event day	text	[-10,-6] \| [-5,-1] \| [0] \| [0,+1] \| [0,+5] \| [+2,+10]	never missing	null
event_timing_aggregate.csv	window_type	Window type	text	yes	Whether the window ends before the filing was public	text	pre_filing \| reaction	never missing	null
event_timing_aggregate.csv	n_events	N events	integer	yes	Number of events aggregated in this cell	integer	null	never missing	null
event_timing_aggregate.csv	sample_status	Sample status	text	yes	Sample-size policy marker. Cells under the threshold carry too_few and BLANK statistics — a blank never means 'zero effect'	text	ok \| too_few	never missing	null
event_timing_aggregate.csv	mean_car_pct	Mean car pct	number	yes	Mean cumulative abnormal return over the window	percentage points (-0.09 means -0.09%)	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	mean_abs_car_pct	Mean abs car pct	number	yes	Mean ABSOLUTE cumulative abnormal return — average reaction size regardless of direction	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	mean_gap_car_pct	Mean gap car pct	number	yes	Mean abnormal OVERNIGHT component (prior close to open)	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	mean_session_car_pct	Mean session car pct	number	yes	Mean abnormal REGULAR-SESSION component (open to close)	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	mean_abs_gap_car_pct	Mean abs gap car pct	number	yes	Mean absolute overnight component	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	mean_abs_session_car_pct	Mean abs session car pct	number	yes	Mean absolute regular-session component	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	direction_q_value	Direction q value	number	yes	Benjamini-Hochberg-adjusted q-value of the directional (BMP) test within this row's fdr_family	0..1	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	direction_significant	Direction significant	boolean	yes	Whether the directional effect survives FDR at q=0.10 within its family	true \| false	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	magnitude_q_value	Magnitude q value	number	yes	BH-adjusted q-value of the reaction-size test (mean \|SAR\| vs noise-only null)	0..1	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	magnitude_significant	Magnitude significant	boolean	yes	Whether excess reaction SIZE survives FDR at q=0.10 within its family	true \| false	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	fdr_family	Fdr family	text	yes	Multiple-testing family this row was corrected in. q-values are NOT comparable across families	text	post_event \| pre_event	never missing	null
event_timing_aggregate.csv	mean_abs_scar	Mean abs scar	text	yes	Mean ABSOLUTE standardized CAR: each event's window CAR divided by its own forecast-error standard deviation, then averaged. Unlike mean_abs_car_pct this IS comparable across windows of different lengths. Under the no-reaction null it averages ~0.798, so a window near 0.80 is ordinary volatility however large its perce...	standard deviations (unitless)	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	corrado_statistic	Corrado statistic	text	yes	Corrado rank-test statistic for direction. A STAGGERED-SAMPLE ADAPTATION, not the textbook Corrado (1989) variance: per-event standardized rank statistics tested cross-sectionally. Documented in LIMITATIONS.md	test statistic	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
event_timing_aggregate.csv	corrado_p_value	Corrado p value	number	yes	Two-sided p-value of the Corrado rank test. NOT FDR-corrected -- unlike direction_q_value and magnitude_q_value in this same file, which are. Provided as a distribution-free cross-check on the BMP direction result, NOT as an independent significance verdict; do not read it against a 0.05 threshold as though it were cor...	0..1	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	finding_family	Finding family	text	yes	Which validation family the finding comes from	text	coarse_item_codes \| earnings_subtypes \| officer_subtypes \| catchall_subtypes \| generic_tone	never missing	null
validated_findings.csv	bucket	Bucket	text	yes	Event slice the finding is about	text	null	never missing	null
validated_findings.csv	category	Category	text	yes	Item code or sub-type label the finding is about	text	e.g. 2.02:guidance_lowered	never missing	null
validated_findings.csv	category_plain	Category plain	text	yes	Plain-language description of the category	text	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	window	Window	text	yes	Event-time window	text	null	never missing	null
validated_findings.csv	discovery_n	Discovery n	integer	yes	Events in the discovery slice	integer	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	discovery_mean_car_pct	Discovery mean car pct	number	yes	Mean CAR in discovery	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	time_holdout_n	Time holdout n	integer	yes	Events in the chronological holdout	integer	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	time_holdout_mean_car_pct	Time holdout mean car pct	number	yes	Mean CAR in the chronological holdout	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	company_holdout_n	Company holdout n	integer	yes	Events in the by-company holdout	integer	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
validated_findings.csv	company_holdout_mean_car_pct	Company holdout mean car pct	number	yes	Mean CAR in the by-company holdout	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
curated_mechanism.csv	category	Category	text	yes	Item code or sub-type label	text	null	never missing	null
curated_mechanism.csv	bucket	Bucket	text	yes	Event slice	text	null	never missing	null
curated_mechanism.csv	window	Window	text	yes	Event-time window	text	null	never missing	null
curated_mechanism.csv	window_type	Window type	text	yes	Whether the window ends before the filing was public	text	pre_filing \| reaction	never missing	null
curated_mechanism.csv	n_events	N events	integer	yes	Events aggregated	integer	null	never missing	null
curated_mechanism.csv	sample_status	Sample status	text	yes	Sample-size policy marker	text	ok \| too_few	never missing	null
curated_mechanism.csv	mean_car_pct	Mean car pct	number	yes	Mean cumulative abnormal return	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
curated_mechanism.csv	mean_abs_car_pct	Mean abs car pct	number	yes	Mean absolute cumulative abnormal return	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
curated_mechanism.csv	mean_abs_gap_car_pct	Mean abs gap car pct	number	yes	Mean absolute OVERNIGHT component (prior close to open)	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
curated_mechanism.csv	mean_abs_session_car_pct	Mean abs session car pct	number	yes	Mean absolute REGULAR-SESSION component (open to close)	percentage points	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null
curated_mechanism.csv	overnight_share	Overnight share	text	yes	Overnight share of the typical move: abs_gap / (abs_gap + abs_session). Absolute components because signed ones cancel and would understate both halves	0..1	null	blank when sample_status=too_few or the test is undefined for the row; blank never means zero	null

Historical SEC 8-K Market-Reaction Dataset — by FlinchLab

Know what each kind of company news historically did to the stock — before you build on top of it. 29,331 SEC 8-K filings from 660 US companies, read and re-classified finer than the official item codes, with each category's price reaction measured by a market-model event study: direction, magnitude, overnight-vs-session split, pre-filing baselines — and every published finding validated on two independent out-of-sample holdouts.

This public repo contains the preview, data dictionary, and methodology only. The full dataset (1,668 aggregate cells + 104 validated findings, run 2026-08-14) is a one-time paid release — link and access instructions below.

Historical and descriptive research data. Not investment advice, not a forecast, not a trading signal.

Load the sample in one line

The free sample carries the complete column schema — every field the full table has, with real figures, over eight rows — so you can see whether it fits your pipeline before anyone pays anything:

import pandas as pd
df = pd.read_csv("hf://datasets/Flinchlab/sec-8k-market-reaction-dataset/preview.csv")

Typical uses: event studies on corporate disclosures, benchmarking abnormal-return estimates around SEC 8-K filings, quantifying announcement effects by news type (guidance changes, executive departures, M&A, layoffs), and teaching event-study methodology with real EDGAR-derived data.

Dataset structure

One row = one aggregate cell: a slice of companies × an 8-K item code × an event-time window. Full field-by-field definitions are in data_dictionary.csv (its own subset in the viewer above).

column	type	meaning
`bucket`, `bucket_type`	text	which slice: ALL, company size (Large/Mid-size/Small), filing time, disclosure lag
`item_code`, `item_label`	text	official SEC 8-K item code and its plain-language name
`window`, `window_type`	text	trading-day window relative to the filing (`[0]`, `[0,+1]`, `[-5,-1]`, …); pre-filing windows are flagged
`n_events`	int	events aggregated in the cell
`sample_status`	text	`ok` or `too_few` — thin cells carry counts, never statistics
`mean_car_pct`, `mean_abs_car_pct`	float	mean (signed / absolute) cumulative abnormal return, %
`mean_gap_`, `mean_session_`	float	the overnight-vs-trading-session decomposition
`direction_q_value`, `direction_significant`	float, bool	BH-FDR-corrected direction test
`magnitude_q_value`, `magnitude_significant`	float, bool	same, for magnitude (separate FDR family)

Two real rows from the sample (run dated 2026-08-14):

bucket	item	window	n	mean CAR	typical size
ALL	All 8-K filings pooled	[0]	29,331	−0.12%	3.50%
ALL	Results of operations (earnings)	[0]	11,241	−0.15%	6.10%

That pair is the pitch in miniature: earnings days move stocks 6% on average, but the direction nets out to a coin flip — the signal lives below the official code, in the finer categories the full release carries.

Cite this

@misc{flinchlab2026sec8k,
  title   = {Historical SEC 8-K Market-Reaction Dataset},
  author  = {{FlinchLab}},
  year    = {2026},
  doi     = {10.5281/zenodo.21986316},
  url     = {https://flinchlab.com/dataset},
  note    = {Version 0.1.1, run dated 2026-08-14. Quote figures with the run date.}
}

The free layer is also archived on Zenodo (DOI 10.5281/zenodo.21986316).

Why this exists when EDGAR is free

SEC EDGAR provides raw filings — public, unstructured, one at a time. This dataset provides the research layer built on top: ~~29,300 8-K events (~~660 US companies, 48 months, run 2026-08-14) normalized, aligned to market-return windows, converted to market-model abnormal returns, statistically tested with explicit multiple-testing control, and validated out-of-sample on two independent axes (chronological persistence and cross-company generalization). Findings that failed validation are disclosed as refuted, not deleted.

One example of what the derived layer shows (both files document it): the official "results of operations" item code averages a −0.15% same-day abnormal return — near nothing. Splitting the same filings by what the release says about forward guidance exposes a spread of several percentage points between guidance-lowered (−5.2% discovery, −6.5%/−5.3% in the two holdouts) and guidance-raised (+2 to +3%) filings. The official taxonomy was averaging opposite reactions into mush; the derived taxonomy separates them, with out-of-sample replication.

What the full release contains

file	grain	rows
`event_timing_aggregate.csv`	size/timing/lag bucket × SEC item code × event window	1,668
`validated_findings.csv`	one row per both-axes-replicated finding, discovery + both holdout effect sizes	104
`data_dictionary.csv`	one row per field (also in this repo)	44
`methodology.md`	plain-language methodology + limitations (also in this repo)	—
`release_manifest.json`	version, provenance, run date, per-file record	—

Row grain, exactly: aggregate statistics only. No issuer names, no tickers, no CIKs, no accession numbers, no individual events, no raw filing text, no prices.

Reading the numbers

Columns ending _pct are percentage points: -0.09 means −0.09%.
sample_status = too_few (n < 10) rows carry blank statistics; blank means "insufficient data", never "zero effect".
q-values are Benjamini-Hochberg adjusted (q = 0.10) and only comparable within the same fdr_family.

Method in one paragraph

Per event: OLS market model on trading days −270..−21 (min 60 obs) → abnormal returns → CARs over six windows from [−10,−6] to [+2,+10], with overnight/session decomposition. Significance tests reported in the table: BMP (Boehmer-Musumeci-Poulsen), Kolari-Pynnonen-adjusted under date clustering, Beaver-style reaction-magnitude test, and Corrado rank test (staggered-sample adaptation) — direction and magnitude corrected as separate FDR families, pre-filing windows as a third. Validated findings additionally passed discovery → pre-registration → holdout on two independent axes. Full details, including all limitations (survivorship bias, daily resolution, US-only, run-to-run count drift), in methodology.md.

Limitations (summary — full list in methodology.md)

Survivorship bias toward current index members · daily resolution · single market (US) · 48-month window · sub-type labels describe what filings assert, not beats/misses vs expectations · universe rebuilds per run, so counts are quoted with run dates.

Explore the findings online

Every published finding is a free answer page with its figures, holdout results and honest caveats:

Access

Preview (8 rows), data dictionary, and methodology: this repo.
Full release: $99 one-time, via https://flinchlab.com/dataset.
Samples, invoices, institutional terms, or questions: data@flinchlab.com

License

One licence at one price: $99 one-time, covering commercial use as well as academic and personal research. The free files in THIS listing (the sample, the data dictionary and the methodology) may be used and quoted freely with attribution; the complete dataset is the paid file, and its full terms ship inside the download as LICENSE.md.

Attribution must include the dataset version and the run date — the company universe is rebuilt each run, so a figure quoted without its run date is not reproducible.

Who makes this

Mustafa. I designed the requirements and the honesty rules; AI implemented them; everything is reproducible from the methodology page at https://flinchlab.com.

Provenance and versioning

Source filings: SEC EDGAR (public). Prices: public market-data APIs. Dataset version 0.1.1 · methodology version 2026-08-14 · produced by a pipeline gated on a 250+-test suite including planted-effect recovery and no-false-positive controls on synthetic data. This is a bounded historical release; no update schedule is promised.

The complete dataset

This repository is the free layer: the preview, the data dictionary, and the methodology. The complete dataset is a one-time paid release, available only at https://flinchlab.com/dataset — that page is the canonical home of the dataset; this repository is a pointer to it. Questions, samples, invoices, or institutional terms: data@flinchlab.com.

Downloads last month: 63

Number of rows:

Total file size:

38.3 kB