Freeze Expectations Before A-Share Earnings, Then Measure the Delta

Where this fits

You are building an agent that tracks A-share companies through earnings season. The agent pulls the latest quarterly report, compares it to what analysts said yesterday, and writes a summary. But you realize your agent is cheating: it uses consensus data that was updated after the report to judge the beat/miss. The a-share-earnings-delta skill solves this by enforcing a two-phase workflow—freeze expectations before the event, then compare only against that frozen snapshot.

Why agents benefit

• Enforces temporal integrity: The skill's hard constraint prevents post-report consensus from contaminating the pre-event baseline. Agents cannot accidentally backfill data, which is a common failure in naive RAG pipelines.
• Structured evidence locking: It calls a-share-research-evidence to snapshot each expectation with a source and timestamp, so every comparison in the review phase has a verifiable origin.
• Granular driver decomposition: Instead of a single EPS miss number, the agent builds a bridge across volume, pricing, FX, costs, provisions, and tax—making the output explainable and audit-ready.
• Management narrative tracking: The skill forces a structured table of management claims vs. actual financial facts, flagging answered, dodged, and later-refuted statements—useful for agentic research reports.
• Differentiated evidence levels: Forecasts, express reports, and formal filings are treated as separate tiers, preventing the agent from conflating preliminary and final numbers.

Practical scenarios

Pre-earnings preview with frozen consensus

Before an annual report drops, your agent runs the preview mode. It collects the company's prior guidance, the latest analyst forecasts, and its own model's predictions—each tagged with a cutoff time. The output is a clean expectation snapshot that later becomes the reference for the beat/miss calculation.

Post-earnings review without look-ahead bias

After the report, the agent runs review. It pulls actual numbers, compares them to the frozen snapshot (not to updated consensus), and builds the driver bridge. The skill's constraint ensures the agent never uses the new quarter's data to re-interpret the old expectation, which is critical for honest backtesting.

Management Q&A credibility check

During an earnings call, the agent transcribes management's claims about order growth and margin recovery. The skill maps each claim to the corresponding line item in the financial statements, then flags discrepancies—e.g., management says "volume up 10%" but revenue bridge shows flat volume. The agent can then highlight the contradiction in the final report.

Add it to your agent workflow

Install the skill via npx and reference it in your agent's tool registry:

npx skills add cyijun/china-financial-services --skill a-share-earnings-delta

Then invoke it in your agent's earnings processing pipeline:

from skills import a_share_earnings_delta

result = a_share_earnings_delta.run(
    mode="review",
    ticker="600519",
    report_period="2024Q4",
    frozen_snapshot=my_pre_event_expected_values,  # must be timestamped before disclosure
)

The skill will output event metadata, expectation snapshot, delta table, driver bridge, quality checks, and forecast revisions—all in the contract defined in references/earnings-schema.md.

Compare before adopting

Before you commit, compare this skill against the alternatives:

• Quality signals: Does it enforce the no-look-ahead rule in code, or just in documentation? This skill has it as a hard constraint in the workflow.
• Maintenance freshness: Check the GitHub repo for recent commits and issue resolution. A-share data sources change frequently; stale parsers break quickly.
• Alternatives: General financial NLP skills may not have the temporal discipline. Specialized A-share skills like china-market-data are complementary, not substitutes.
• Workflow fit: This skill is designed for research agents that produce preview/review reports. If your agent only does real-time alerts, it may be overkill.

Why it is worth tracking

This skill addresses a subtle but critical failure mode in earnings agents—using future data to judge past events. Its 18 stars are early but the design is sound, and the README shows a disciplined, schema-driven approach that many community skills lack. Track it because it sets a precedent for temporal correctness in financial agent skills; evaluate it when you need to build or upgrade an A-share earnings analysis pipeline. See more context on OpenAgentSkill.