README.md

---
license: apache-2.0
language:
  - en
pretty_name: US Earnings Calendar (announcement times and reported figures, PIT)
size_categories:
  - 1M<n<10M
task_categories:
  - tabular-regression
  - time-series-forecasting
tags:
  - finance
  - sec
  - edgar
  - earnings
  - earnings-calendar
  - 8-k
  - event-study
  - point-in-time
  - ziplime
configs:
  - config_name: earnings
    data_files:
      - split: train
        path: "data/earnings/*.parquet"
  - config_name: announcements
    data_files:
      - split: train
        path: "data/announcements/*.parquet"
  - config_name: events
    data_files:
      - split: train
        path: "data/events/*.parquet"
  - config_name: pit
    data_files:
      - split: train
        path: "data/pit/*.parquet"
---

# US Earnings Calendar

When a company announced its results, **to the second**, what those results
were, and which session could act on them.

**454 611 earnings releases · 258 482 paired with reported figures ·
4 382 182 corporate events · 2003-04-25 to 2026-09-04**

The pipeline lives in [`recipe/`](recipe/) at the same revision as the data.
See [PIPELINE.md](PIPELINE.md) for the method.

## This is not a calendar of future releases

Nothing here schedules an announcement. A row appears when EDGAR accepted the
8-K — that is, once the release has already happened — and the dataset is built
for measuring reactions, not for staying out of the market beforehand.

The obvious extrapolation does not work well enough to lean on. Over 406 837
consecutive release intervals, "the previous release plus 91 days" has a median
absolute error of **7 days** and lands within a week only **54.7%** of the
time. The intervals themselves: median 91, p10 56, p25 79, p75 95, p90 112.

## The time of day is the whole point

A release at 07:00 New York trades on the open. One at 16:30 trades overnight.
Four hours of error swaps them, and with them the sign of any event study.

| Session | Releases |
|---|---|
| `after_close` | 227 500 |
| `pre_market` | 138 534 |
| `market_hours` | 87 978 |
| `non_session` | 599 |

**80.7% land outside the session**, which is what earnings releases do — and
the fact that they do is the check that the timestamps are right.

Every timestamp is read from the filing's own SGML header, where
`ACCEPTANCE-DATETIME` is unambiguous US Eastern. **Not** from EDGAR's bulk
`submissions.zip`, whose `acceptanceDateTime` carries a `Z` and is Eastern
local time for much of the corpus:

| Filed | Share of bulk timestamps that are Eastern, not UTC |
|---|---|
| 2004–2009 | 44% |
| 2010–2015 | 62% |
| 2016–2020 | 50% |
| 2021–2026 | 6% |

Nothing in the value distinguishes the two, so reading them costs one request
per release — 445 271 of them. That is why this dataset took a day to build and
why its times can be trusted.

## `first_tradeable_session`: one date instead of three arguments

The session boundaries come from the NYSE calendar, not from 09:30 and 16:00.
On the 50 half days in this period the exchange closes at 13:00, and on 25 days
it held no session at all while EDGAR kept accepting filings — 303 releases
arrived during the two days Sandy shut the floor in October 2012.

`first_tradeable_session` is the first session whose close falls after the
filing was accepted:

| Release | `session` | `first_tradeable_session` |
|---|---|---|
| Wed 07:00 | `pre_market` | that same day |
| Wed 16:30 | `after_close` | Thursday |
| Fri 14:19, day after Thanksgiving | `after_close` (13:00 close) | Monday |
| Mon 29 Oct 2012, Sandy | `non_session` | Wed 31 Oct |

One column, and the holidays, the half days and the 16:30 releases stop being
three separate arguments a consumer has to have.

## Symbols, resolved as of the day

`ticker` is the symbol the company carried **on the day it filed**, from
[security-master](https://huggingface.co/datasets/ZipLime/security-master)'s
29 511 dated spans. `ticker_confidence` says how it was resolved:

| | Releases | |
|---|---|---|
| `as_of` | 366 458 | a span covering that day: no assumption |
| `carried_back` | 56 611 | before EDGAR's symbol data starts in 2006 |
| `carried_forward` | 7 446 | after the last day the symbol was seen |
| (none) | 24 096 | no symbol on record for that company |

Resolving from today's map instead would put back exactly the look-ahead the
rest of this family removes: a company that changed symbol in 2020 would carry
the new one across its whole history, and the prices found under it are the
post-change ones. Meta's releases up to the first quarter of 2022 say `FB`.

## What is paired, and what is not

An 8-K's own `reportDate` is the day of the event, not the quarter being
reported: Apple's release of 30 July 2026 is dated 30 July and reports the
quarter that ended in June. So the period comes from the statement that
follows — the 10-Q or 10-K filed after the release, for a period that had
already closed.

**88.4% of releases from 2011 to 2025 pair with a reported period.** Including
2026 the figure is 87.2%, because this year's statements are not all filed yet;
before 2011 the rate falls away entirely, because the figures come from XBRL
and XBRL begins in 2009. The blended rate over all years is 56.9% and means
nothing.

| Field | Present in paired rows |
|---|---|
| net income | 90.5% |
| revenue | 86.5% |
| diluted EPS | 73.2% |
| revenue year-on-year | 77.8% |
| EPS year-on-year | 64.4% |
| `revenue_sue` | 56.5% |
| `sue` | 45.3% |

The median company files its statement **4 days** after announcing.

### `sue`, because consensus is not available and was never needed

Analyst consensus is licensed data. The drift literature does not use it: it
uses **standardised unexpected earnings** — the surprise against the same
period a year earlier, divided by the standard deviation of that surprise over
the eight periods before it. Every input is already here.

It is dimensionless, which raw growth is not. A utility growing 3% and a
biotech growing 300% cannot be ranked against each other on
`revenue_yoy_change`; a cross-sectional sort on growth sorts volatility as much
as surprise. `sue` and `revenue_sue` can be ranked. Only periods whose
statements had already been filed when the release was made enter the window,
so the measure is knowable at its own knowledge date. The distribution: median
0.12, p10 −1.8, p90 2.4, and a 1.8% tail beyond ±10 for companies whose past
surprises were unusually small.

### `eps_quality`, because a few EPS values are totals

`net_income / eps_diluted` is the share count a row implies. Across the rows
carrying both, p1 is 2.07 million and the median 56.7 million — and 197 rows
imply fewer than fifty thousand shares. Halliburton's fourth quarter of 2023
implies 903. Those are totals that were tagged into a per-share field upstream.

Where the row's own columns contradict its EPS, `eps_diluted` and
`eps_yoy_change` are blanked and `eps_quality` is `implausible`. The
announcement and its timestamp survive — they are correct, and they are what
the dataset is for. `unverified` (6 940 rows) means net income was missing and
the check could not run.

## Everything else the scan saw

Isolating earnings means reading the item codes of every 8-K, so the rest of
the taxonomy came free: **4 382 182 dated corporate events** across 46 item
codes, each with a ticker and a knowledge date of its own.

| Item | Meaning | Count |
|---|---|---|
| 9.01 | financial statements and exhibits | 1 353 927 |
| 8.01 | other event | 468 250 |
| 2.02 | results of operations | 430 429 |
| 7.01 | Regulation FD disclosure | 337 199 |
| **5.02** | **director or officer changed** | 315 589 |
| 1.01 | material agreement entered | 310 448 |
| 2.03 | direct financial obligation created | 110 396 |
| 5.07 | shareholder vote results | 90 115 |

Bare `7` and `5` are the pre-2004 numbering, kept as filed.

`knowledge_date` is exact for the 23.1% of rows whose headers were read for the
calendar; every other event is dated to **23:59:59 Eastern on its filing day**,
which is late by up to a day and never early. `knowledge_estimated` says which
is which. (63 065 events carry an exact timestamp *earlier* than their filing
date: EDGAR stamps anything accepted after 17:30 with the next business day.)

## Configs

| Config | Rows | What one row is |
|---|---|---|
| `earnings` | 258 482 | a release paired with the period and figures it announced |
| `announcements` | 454 611 | a company's earnings release, timed to the second |
| `events` | 4 382 182 | one 8-K item: a dated corporate event of any kind |
| `pit` | 258 482 | one point-in-time earnings knowledge event, plus a Delta table |

**All four configs carry `knowledge_date`**, so all four mount, not only
`pit`: in `announcements` and `earnings` it is the acceptance instant itself,
in `events` it is exact where a header was read and the end of the filing day
otherwise. A config whose knowledge column is named anything else is refused by
ziplime rather than indexed on an event date.

`entity_id` in `pit` is the issuer CIK, the same key as every sibling dataset,
so an earnings date joins directly to
[company-fundamentals](https://huggingface.co/datasets/ZipLime/company-fundamentals),
[insider-trading](https://huggingface.co/datasets/ZipLime/insider-trading) and
the rest. `pit` also carries `form` and `accession_number`, so an 8-K/A
amendment is distinguishable from a first announcement and every row joins back
to `announcements`.

```python
import polars as pl

pit = pl.read_parquet("data/pit/*.parquet")
# Everything a strategy could have known at this instant
known = pit.filter(pl.col("knowledge_date") <= "2025-02-01T00:00:00Z")
# Large positive surprises, ranked comparably across industries
surprises = known.filter(pl.col("sue") > 2).select(
    "ticker", "first_tradeable_session", "sue", "eps_diluted"
)
```

## Known gaps

* **No analyst estimates, so no surprise against consensus.** `sue` measures
  the surprise against the company's own history instead, which is what the
  drift literature does.
* **Releases before 2009 have no figures.** The announcement, its time and its
  session are there; the numbers are not, because XBRL had not started.
* **A release is not always about the latest quarter.** The pairing requires
  the period to have closed and the statement to follow within 150 days;
  releases that satisfy neither are in `announcements` and not in `earnings`.
* **Six releases have an unreadable header** and one filing is missing from
  EDGAR. They are absent from `announcements` rather than dated by guess.
* **5.3% of releases have no symbol** and a further 12.5% resolve to one
  carried back from a later observation, almost all of them before 2006.
* **Co-registrants file jointly.** Six thousand releases are filed by a parent
  and its subsidiaries in one document; each company gets its own row, and the
  event id is the company and the filing together.

## Provenance and updates

Sources: SEC EDGAR bulk submissions (item codes, dates) and each filing's own
SGML header (acceptance time). Figures from
[ZipLime/company-fundamentals](https://huggingface.co/datasets/ZipLime/company-fundamentals),
symbols from
[ZipLime/security-master](https://huggingface.co/datasets/ZipLime/security-master),
sessions from the XNYS calendar. US Government works, public domain. Rebuilt
weekly on Hugging Face Jobs.