README.md

---
license: cc0-1.0
language:
  - en
pretty_name: SEC EDGAR Fundamentals (PIT)
tags:
  - point-in-time
  - pit
  - ziplime
  - backtesting
  - fundamentals
  - sec-edgar
  - us_equities
task_categories:
  - time-series-forecasting
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/data_bundle/**/*.parquet
---

# 🧾 SEC EDGAR Fundamentals (PIT)

Point-in-time company fundamentals from SEC filings — revenue, diluted EPS, net income and
total assets, keyed by **filing time**, not fiscal period. Every figure is stored *as it was
reported*, so restatements and publication lag can't leak into a backtest.

Part of the **ziplime Point-in-Time (PIT) data layer**: a simulation at time **T** only ever
observes rows with `knowledge_date <= T`. The identical code path runs live with **T = now**.

- **Data class:** Fundamentals — SEC filings (XBRL company facts)
- **Entity domain:** `us_equities` — US-listed issuer, keyed by ticker
- **Coverage:** **93 issuers**, **23,580 as-reported facts** (including restatement history)
- **event_date range:** 2006-12-31 → 2026-06-13
- **knowledge_date range:** 2009-05-07 → 2026-07-09
- **Origin:** SEC EDGAR company facts (XBRL) · US-Government public domain
- **Update cadence:** daily, following the EDGAR filing index (`0 5 * * *`)
- **Format:** ziplime Delta Lake bundle (`data_type: PIT_DATA`), partitioned by `knowledge_year`

## Why point-in-time?

A conventional fundamentals table stores one value per fact — the final one. A backtest then
trades in August on a number that was only restated in October. This dataset keeps the split:

| Column | Type | Semantics |
|---|---|---|
| `entity_id` | Utf8 | Ticker of the issuer |
| `event_date` | Timestamp(UTC, µs) | Period end the figure refers to |
| `knowledge_date` | Timestamp(UTC, µs) | Filing date it became public — **the only column the as-of filter uses** |
| `knowledge_estimated` | Boolean | `true` if the filing date was modelled rather than sourced (here: always `false`) |
| `ingested_at` | Timestamp(UTC, µs) | Pipeline write time (audit only) |

### Value columns

| Column | Type | Description |
|---|---|---|
| `revenue` | Float64 | Total revenue for the period |
| `eps_diluted` | Float64 | Diluted earnings per share |
| `net_income` | Float64 | Net income |
| `total_assets` | Float64 | Total assets (balance-sheet date) |
| `fiscal_period` | Utf8 | e.g. `FY2024`, `2024Q3` |
| `form` | Utf8 | Filing form: `10-K`, `10-Q`, `10-K/A`, … |
| `accession_no` | Utf8 | SEC accession number (provenance) |

The logical key of a fact is `(entity_id, event_date)`. A **revision** is a new row with the
same key and a later `knowledge_date` — a restatement, or the same period re-reported as a
comparative in a later filing.

### Restatement in the data — a worked example

Apple's FY2008 (`event_date = 2008-09-27`) as it actually became known:

| knowledge_date | form | revenue | eps_diluted |
|---|---|---|---|
| 2009-10-27 | 10-K | 32.48 B | 5.36 |
| 2010-01-25 | 10-K/A | **37.49 B** | **6.78** |

The 10-K/A is Apple's retrospective adoption of new revenue-recognition rules. `as_of("2009-11-01")`
returns 32.48 B — the only figure a strategy could have traded on that day. `as_of("2010-02-01")`
returns 37.49 B. The restated number never leaks backwards.

## As-of access

Inside a ziplime strategy there is **no `T` parameter** — the knowledge moment equals the
simulation clock:

```python
async def initialize(context):
    context.fundamentals = await context.pit("sec-fundamentals-pit")

async def handle_data(context, data):
    latest = await context.fundamentals.latest(
        assets=[context.asset], fields=["revenue", "eps_diluted"]
    )
    history = await context.fundamentals.as_of(
        assets=[context.asset], fields=["revenue"], event_range=("2018-01-01", None)
    )
```

### Reading it outside ziplime (plain Polars + delta-rs)

```python
import polars as pl
from datetime import datetime, timezone
from huggingface_hub import snapshot_download

path = snapshot_download(
    "ZipLime/sec-fundamentals-pit", repo_type="dataset",
    allow_patterns=["data/data_bundle/**"],
)
delta = f"{path}/data/data_bundle/sec_fundamentals_pit/1784818614/data.delta"

T = datetime(2009, 11, 1, tzinfo=timezone.utc)   # "what was known at T"
as_of = (
    pl.scan_delta(delta)
      .filter(pl.col("knowledge_date") <= T)     # point-in-time filter
      .sort("knowledge_date")
      .group_by(["entity_id", "event_date"], maintain_order=True)
      .last()
)
print(as_of.filter(pl.col("entity_id") == "AAPL").collect())
```

The table is partitioned by `knowledge_year`, so the `knowledge_date <= T` filter prunes at
the partition level. Delta time-travel (`AS OF <version>`) pins the table for reproducibility;
it composes with the knowledge filter rather than replacing it.

## Updates

`recipe.py` implements `fetch(since) -> pl.DataFrame` against SEC company facts; `ingest.py`
dedups and **appends** to the Delta bundle (never rewrites). The scheduled job in
`.github/workflows/update.yml` runs it daily.

## What's in this repo

```
README.md                     # this card
manifest.json                 # PIT manifest (schema, source, coverage)
recipe.py                     # fetch(since) -> PIT rows from SEC company facts
ingest.py                     # dedup + append-only Delta writer
.github/workflows/update.yml  # scheduled ingestion
data/                         # ziplime Delta bundle + registry manifest
  bundle_registry/sec_fundamentals_pit_1784818614.json
  data_bundle/sec_fundamentals_pit/1784818614/data.delta/   (partitioned by knowledge_year)
```

To load with ziplime, drop `data/bundle_registry/*` and `data/data_bundle/*` into your
`~/.ziplime/data/` and read via `context.pit("sec-fundamentals-pit")`; or point
`pl.scan_delta` straight at the Delta table as shown above.

## Provenance & license

Built from SEC EDGAR XBRL **company facts** (public domain, US Government work) via the
`dartlab-data` mirror, repackaged into the ziplime PIT schema. Filing dates (`filed`) are used
verbatim as `knowledge_date`; fact periods are classified from each fact's own reporting
window. No values are imputed.