ZipLime/news-sentiment-llm

🗞️ LLM Headline Sentiment (PIT)

Financial news headlines scored for sentiment by an LLM — point-in-time by wire timestamp, with the scoring model's knowledge cutoff tracked for honest out-of-sample evaluation.

Part of the ziplime Point-in-Time (PIT) data layer — append-only datasets with an explicit split between when a fact happened (event_date) and when it became known (knowledge_date). A simulation at time T can only ever observe rows with knowledge_date <= T, so restatements, publication lag and hindsight can't leak into a backtest. The identical code path runs live with T = now.

• Data class: Alternative data — LLM-derived signal
• Entity domain: us_equities — The primary ticker the headline is about.
• Origin: Financial news wires, scored by an in-house LLM pipeline
• License: Derived signal — research use only
• Update cadence: hourly, scoring headlines as they cross the wire (0 * * * *)
• Format: ziplime Delta Lake bundle (data_type: PIT_DATA)

Why point-in-time?

Backtests on non-price data are systematically optimistic when the data layer has no notion of when a fact became known. Three failure modes this dataset is built to avoid:

1. Restatements — a value reported one quarter and revised the next. Storing only the final value lets a backtest "know" the revision months early.
2. Publication lag — fundamentals keyed by fiscal-period-end, joined to prices at period end rather than the (weeks-later) filing date.
3. Hindsight in derived signals — a recent model scoring old text has already seen how the story ended.

All three are the same bug, and it is fixed in the data layer, not in strategy code.

Model knowledge cutoff: 2025-10-01T00:00:00Z. Backtests whose window opens before this date are flagged for look-ahead; treat only the post-cutoff window as honest out-of-sample.

Schema

System columns (every PIT dataset)

Column	Type	Semantics
entity_id	Utf8	Stable entity identifier (resolved via the entity_map PIT dataset)
event_date	Timestamp(UTC, µs)	The moment the fact refers to
knowledge_date	Timestamp(UTC, µs)	The moment it became publicly known — the only column the as-of filter uses
knowledge_estimated	Boolean	true if knowledge_date was reconstructed by a lag model rather than taken from the source
ingested_at	Timestamp(UTC, µs)	When our pipeline wrote the row (audit only; never used in as-of)

Value columns (this dataset)

Column	Type	Description
headline	Utf8	Headline text
source_outlet	Utf8	Publishing outlet
sentiment_score	Float64	Model score in [-1, 1]
sentiment_label	Utf8	bearish / neutral / bullish
model_name	Utf8	Scoring model id

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. Written rows are immutable; history is never rewritten.

As-of access

Inside a ziplime strategy there is no T parameter — the knowledge moment always equals the simulation clock (live: wall clock):

async def initialize(context):
    context.ds = await context.pit("news-sentiment-llm")

async def handle_data(context, data):
    # only rows with knowledge_date <= current simulation time are visible
    latest = await context.ds.latest(
        assets=[context.asset], fields=['headline', 'source_outlet']
    )
    history = await context.ds.as_of(
        assets=[context.asset], fields=['headline'],
        event_range=("2022-01-01", None),
    )

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

import polars as pl

T = "2025-06-01T00:00:00Z"          # "what was known at T"
lf = pl.scan_delta("hf://datasets/ZipLime/news-sentiment-llm/data/data_bundle/yahoo_finance_daily_data/1784755946/data.delta")
as_of = (
    lf.filter(pl.col("knowledge_date") <= T)
      .sort("knowledge_date")
      .group_by(["entity_id", "event_date"], maintain_order=True)
      .last()
)
print(as_of.collect())

Delta time-travel (AS OF <version>) pins the table for reproducibility; the knowledge_date <= T filter is what enforces point-in-time. They compose: a backtest records (dataset, delta_version) and replays read the table at that version and apply the filter.

Updates

recipe.py implements the collection contract fetch(since: datetime) -> pl.DataFrame in the PIT schema above; ingest.py dedups and appends to the Delta bundle (never rewrites). The scheduled job in .github/workflows/update.yml runs it hourly, scoring headlines as they cross the wire.

# recipe.py (contract)
async def fetch(since: datetime) -> "pl.DataFrame": ...

Knowledge-date convention

knowledge_date = the wire publication timestamp of the headline (never the scoring time). The scoring model's training cutoff is recorded in the manifest (model_knowledge_cutoff). A 2026 model scoring 2018 headlines has already seen how those stories played out, so any backtest whose window predates the cutoff is upward-biased — the engine flags those runs and the honest evaluation window starts after the cutoff.

What's in this repo

README.md                     # this card
manifest.json                 # PIT dataset manifest (schema, source, schedule)
recipe.py                     # fetch(since) -> PIT rows
ingest.py                     # dedup + append-only Delta writer
.github/workflows/update.yml  # scheduled ingestion
data/                         # ziplime Delta bundle + registry manifest
  bundle_registry/yahoo_finance_daily_data_1784755946.json
  data_bundle/yahoo_finance_daily_data/1784755946/data.delta/

The data/ bundle is a ready-to-load ziplime Delta Lake market-data bundle (five US equity tickers, daily bars) that seeds the pipeline and lets you exercise the loader end-to-end today. Point pl.scan_delta (above) at it, or register it with ziplime's FileSystemBundleRegistry.

---

_{Generated for the ziplime PIT data-layer prototype. Manifest and schema follow the ziplime PIT spec; source.* fields declare origin and license per the dataset manifest.}