hkfilings · Python SDK for HKEX annual & interim reports

Stop hand-copying numbers from 300-page HKEX PDFs.
Get every fact — with its source page — in 5 lines of Python.

English · 中文
Install · Quickstart · Data shape · Cookbook · API · Plans · FAQ

Why this exists

If you've ever tried to extract revenue, segment breakdowns or capex guidance from a 300-page HKEX PDF, you know the failure modes: numbers without page references, OCR drift, units mixed between HKD and RMB, hallucinated year-over-year changes that don't reconcile, segment totals that don't sum to the consolidated figure.

This SDK is a client for an API that returns every fact with:

• source_page + source_text — point at the exact location in the PDF (page-level bbox lands in extra when the parser can resolve it)
• 13 deterministic validators — YoY recalc, accounting equation, segment reconciliation, currency / unit consistency, cashflow sign checks, EPS sign, free-cashflow definition, …
• Frozen v1 schema with forward-compat extra dict — new backend fields land in extra; your existing code never breaks on a release
• Layer-2 signals — 11 enumerated signal types (margin driver, upstream cost pressure, downstream demand, inventory & orderbook, capex, …), each bound to evidence with anti-hallucination rules
• Supply-chain graph — suppliers, customers, competitors, regulators, substitutes, partners — with exposure share and direction

vs the alternatives

Install

pip install hkfilings

Requires Python ≥ 3.10. The only runtime dependency is httpx. Works great in Jupyter / Colab.

Quickstart

from hkfilings import HKFilingsClient

client = HKFilingsClient(api_key="ak_...")

task   = client.analyze(ticker="9988", year=2026)   # Alibaba 9988.HK
report = client.wait(task.task_id, timeout=600)

for fact in report.facts:
    print(f"{fact.metric_key:32}  {fact.value:>18,.0f}  p.{fact.source_page}")

Free tier: 20 tasks/month, no credit card required. → Get a key

What you get back

Every fact carries its provenance:

fact.metric_key       # "revenue"
fact.metric_label     # "Revenue"
fact.value            # 245_864_000_000.0
fact.comparable_value # 224_500_000_000.0    (prior period)
fact.yoy_change       # 0.0952
fact.source_page      # 87
fact.source_text      # "Revenue for the year increased to RMB245,864..."
fact.confidence       # 0.98
fact.extra            # forward-compat fields (e.g. bbox, unit) land here

Layer-2 signals come with direction + evidence:

signal.signal_type     # "margin_driver"
signal.direction       # "up" | "down" | "flat"
signal.summary         # "Cloud margin lifted on lower GPU procurement..."
signal.evidence        # [{"page": 42, "text": "..."}, ...]
signal.review_status   # "approved" | "auto_passed" | "pending"

Supply-chain nodes come with role + exposure share:

node.node_label             # "TSMC"
node.node_role              # "supplier" | "customer" | "competitor" | "regulator" | "substitute" | "partner"
node.exposure_share         # 0.18    (18% of revenue / cost tied to this node)
node.direction_to_company   # "inflow" | "outflow"
node.evidence_page          # 142

Cookbook

Tip: every snippet below runs as-is in a Jupyter notebook once you've exported HKFILINGS_API_KEY in your shell.

Compare gross margin: BABA vs Tencent, last 3 fiscal years

from hkfilings import HKFilingsClient
import pandas as pd

client = HKFilingsClient(api_key="ak_...")
rows = []
for tk in ("9988", "0700"):
    m = client.company_matrix(tk, metrics=["revenue", "gross_profit"])
    rows.extend({"ticker": tk, **cell} for cell in m.cells)

df = pd.DataFrame(rows).pivot_table(
    index="period", columns=["ticker", "metric_key"], values="value"
)
print(df)

Read industry signals + evidence

sigs = client.task_signals(task.task_id, signal_type="margin_driver")
for s in sigs.signals:
    print(f"[{(s.direction or '-'):>4}] {s.summary}")
    for ev in s.evidence:
        text = (ev.get("text") or "")[:80]
        print(f"        p.{ev.get('page')} — {text}")

# [  up] Cloud Intelligence margin lifted 4.2pp YoY on lower GPU procurement...
#         p.42 — During the year, Cloud Intelligence Group recorded segment...

Render the supply-chain graph

import networkx as nx

graph = nx.DiGraph()
sc = client.company_supply_chain("9988")
for node in sc.nodes:
    graph.add_edge(
        "9988",
        node.node_label,
        role=node.node_role,
        exposure=node.exposure_share,
    )
print(graph)   # → "DiGraph with N nodes and M edges"  (networkx ≥ 3.0)

Export to CSV for Excel

with open("baba_2026_facts.csv", "wb") as fh:
    fh.write(client.facts_csv(task.task_id))

More runnable examples in examples/.

API reference

Full docs: https://docs.hkfilings.app/python

Plans & limits

The SDK itself is free under MIT. The managed API runs on a usage-based plan with a free tier:

→ See current pricing & upgrade on hkfilings.app

You can also self-host the parsing service if you have your own LLM budget — set base_url="https://your-host" when constructing the client. Get in touch (sales@hkfilings.app) for on-prem deployments.

Schema contract

The v1 public schema is frozen. JSON Schema documents:

• https://hkfilings.app/v1/schema/financial_fact
• https://hkfilings.app/v1/schema/industry_signal
• https://hkfilings.app/v1/schema/supply_chain_node
• https://hkfilings.app/v1/schema/catalyst

New backend fields land in each dataclass's extra dict — you do not need to upgrade the SDK to see them. We never remove or rename a documented field within v1.

Configuration

The client honors HTTPS_PROXY / HTTP_PROXY via httpx — useful behind corporate firewalls.

Roadmap

• v0.2 — Async client (HKFilingsAsyncClient), retry with exponential backoff, client.facts_to_dataframe() helper.
• v0.3 — Streaming task events (SSE wrapper), webhook signature helpers, CLI (hkfilings analyze 9988 2026).
• v1.0 — Stable surface, deprecation warnings removed, TypeScript SDK companion.

Open issues for what you'd like prioritized.

FAQ

Q: Can I self-host the parsing engine? A: Yes for Enterprise customers — contact sales@hkfilings.app. The parsing service requires an LLM API key (DeepSeek, OpenAI, or Anthropic) and a Postgres database.

Q: How is this different from Wind / Bloomberg / Capital IQ? A: We focus on Hong Kong listings only and ship the entire evidence chain (source page + text) alongside every number. Our schema is public and frozen; theirs are proprietary and shift between releases. We're priced for individuals and small funds.

Q: A-shares? US listings? A: Not yet. We may add them once the HK coverage is rock-solid.

Q: Is the parsing logic open-source? A: This SDK is open-source under MIT. The parsing engine — LLM prompts, validators, anti-hallucination rules — is closed-source. We publish the JSON Schema contract as a stability commitment.

Q: How do I report a parsing bug for a specific report? A: Open a GitHub issue with the task_id. We'll triage on the SaaS side and either fix the engine or update the report.

Q: I got rate-limited. What now? A: Free tier is 20 tasks/month. Upgrade or wait for the next monthly reset. HKFilingsError will include status_code=429 and payload["upgrade_url"].

License

MIT. See SECURITY.md for vulnerability reporting, CONTRIBUTING.md to help out.

⭐ If this saves you time, star the repo — it's the cheapest way to support the project.
Get a free key → hkfilings.app/signup