Point-in-Time Datasets

Ziplime reads point-in-time datasets straight from the Hugging Face Hub — no ingest step, no bundle to build, no credentials. Rows are indexed on the date they became public, so a strategy on bar t sees only what was known by t.

pip install "ziplime[huggingface]"

Published datasets

Dataset	What
`ZipLime/congress-trading`	Congressional trading disclosures
`ZipLime/insider-trading`	Insider transactions (Form 4)
`ZipLime/earnings-calendar`	Earnings releases, with acceptance timestamps
`ZipLime/company-fundamentals`	Company fundamentals

Browse them at huggingface.co/ZipLime. Each dataset card describes its tables (configs) and columns.

The one-line form

Name the dataset by address inside data.history:

df = await data.history(
    assets=context.universe,
    bar_count=40,
    fields=["net_notional_usd"],
    data_source="hf://ZipLime/congress-trading/features",
)

The first read resolves the dataset to a commit, downloads only the Parquet parts the simulation window can reach, maps tickers onto the asset database and keeps the result in memory.

The one-line form mounts the simulation’s own window, so it sees nothing published before the first bar. A signal that looks back 90 days starts blind. Use the explicit form with an earlier start_date when you need history from before the run.

The explicit form

Mount once in initialize, reaching back before the first bar, then read it like any other source:

import datetime

LOOKBACK = datetime.timedelta(days=120)


async def initialize(context):
    context.universe = await context.universe_symbols("Q500US")
    context.insiders = await context.huggingface_dataset(
        "ZipLime/insider-trading",
        config="features",
        revision="ba0785efcede0b3a13af48dc658a1d39bc87ad1e",
        fields=["is_cluster_buy", "n_unique_buyers_30d"],
        start_date=context.clock.start_session - LOOKBACK,
    )


async def handle_data(context, data):
    rows = await data.history(
        assets=context.universe,
        since=datetime.timedelta(days=90),
        fields=["is_cluster_buy"],
        data_source=context.insiders,
    )
    ...

Argument	Meaning
`repo_id`	`owner/name` on the Hub, or a full `hf://owner/name/config` address
`config`	The table to mount. Defaults to `features` when the dataset has one
`revision`	Branch, tag or commit. A branch is resolved to a commit and reported
`fields`	Columns to keep besides `date` and `sid`. All by default
`start_date`, `end_date`	Window to fetch. Default: the simulation’s own
`name`	What to call the source. Default: its address

Pin a revision

These datasets are append-only and grow. Left on main, the same backtest a month apart reads different data and returns a different number. The engine resolves a branch to a commit up front and logs it:

Pinned a Hugging Face dataset to the commit it is at now
  dataset=ZipLime/congress-trading revision=67c335f5207d5190ada5f89803615848dbf52ee0

Pass that commit back as revision= to repeat the run exactly.

Read the mount report

The mount logs how many Parquet parts it read, how many rows fell in the window and how many it mapped onto listings, and which tickers matched no listing. A large drop usually means the asset database is missing instruments — ingest them, or the rows are silently absent from every read.

Event data wants `since`

Disclosure datasets are sparse: thirty filings for one ticker can span four years. Read them with since=datetime.timedelta(...) rather than bar_count=.

In vectorized signals

A mounted dataset is also available to compute_signals as prices.dataset(name). See Vectorized Signals.

Other address schemes

hf:// is built in. Extensions can register their own schemes, and a strategy then reads data.history(..., data_source="scheme://...") with no registration of its own — zipfinam adds algopack:// for MOEX AlgoPack. See Venues & Calendars. Worked examples, with results and the controls they are measured against, are in examples/huggingface.

Introduction

Writing Strategies

Strategy Language

Migrating to Ziplime

Limitations

Point-in-Time Datasets

Published datasets

The one-line form

The explicit form

Pin a revision

Read the mount report

Event data wants `since`

In vectorized signals

Other address schemes

​Published datasets

​The one-line form

​The explicit form

​Pin a revision

​Read the mount report

​Event data wants since

​In vectorized signals

​Other address schemes

Published datasets

The one-line form

The explicit form

Pin a revision

Read the mount report

Event data wants `since`

In vectorized signals

Other address schemes