starchild-ai-agent Earnings Surprise Detector
The problem it solves
You see a research report or hear a thesis: "X will grow earnings 80-120% this quarter." You build a position plan on that number. Then the actual earnings come in at -17%. You're now holding a bag based on a fantasy.
This happened. Multiple times. 太辰光: research said +80-120%, actual was -17%. 英维克: research said +150%, actual was -82%. 欧陆通: research said +180%, actual was a loss.
The fix is simple: verify before you act. Pull the actual number from an authoritative source (东方财富 via mx-data), compare to the expectation, and let the gap govern your action.
When to use
- Any A-share recommendation that cites expected earnings growth (扣非同比, 营收同比, EPS)
- Before stating entry prices, position sizes, or 9-column execution tables
- When a research report or thesis hinges on a specific growth figure
- Before adding to an existing position based on earnings momentum
Do NOT skip this step. "The research said X" is not verification. "I pulled the actual number from 东方财富 and it confirms Y" is verification.
Workflow
Step 1 — Identify the claim
Extract from the recommendation/thesis:
- Stock: name + code (verify code against user's exact text — never auto-correct tickers)
- Expected growth: the specific number/range cited (e.g., "扣非同比 +80-120%")
- Period: which quarter/year the expectation refers to
- Source: who made the expectation (research report, analyst consensus, agent's own analysis)
If no specific growth number is cited, the thesis is too vague to verify — flag this and either pin down the number or treat the thesis as ungrounded.
Step 2 — Pull actuals via mx-data
Run the verification script:
python3 skills/earnings-surprise-detector/scripts/verify_earnings.py \
--stock "太辰光" \
--code 300570 \
--expected-min 80 \
--expected-max 120 \
--metric 扣非归母净利润同比
The script calls mx-data (东方财富) to fetch the latest quarterly 扣非归母净利润同比 growth, parses the result, and compares to the expected range.
Supported metrics:
扣非归母净利润同比(default — the most important for A-share verification)营收同比净利润同比EPS
If mx-data API quota exhausted (status=113): The script will report this. Switch to manual verification — do NOT proceed without actuals. Use the linqi-data skill as a backup source if available.
Step 3 — Verdict
The script outputs one of three verdicts:
| Verdict | Condition | Action |
|---|---|---|
| ✅ PASS | Actual ≥ expected-min | Proceed with recommendation. Earnings confirm the thesis. |
| ⚠️ WARN | Actual is within -20% of expected-min (e.g., expected 80%, actual 65%) | Size down 50%. Thesis partially confirmed but momentum weaker than expected. Recheck thesis logic. |
| ❌ FAIL | Actual < expected-min × 0.8 (e.g., expected 80%, actual < 64%) OR actual is negative when positive expected | Exclude or downweight. Do not recommend entry. If already holding, reassess immediately. |
Hard rule: When verdict is FAIL, do NOT generate entry prices, target prices, or 9-column execution tables. The thesis is broken at the earnings level — no amount of technical analysis fixes that.
Step 4 — Integrate verdict into output
Every stock recommendation that passed through this skill should include:
### Earnings Verification
- Stock: [name] ([code])
- Expected: [metric] [expected range] (source: [research/analyst])
- Actual: [metric] [actual value] (source: 东方财富 via mx-data, [report date])
- Verdict: [PASS/WARN/FAIL]
- Impact on recommendation: [proceed / size down / exclude]
If the verdict is FAIL, the recommendation section should state: "Earnings verification FAILED — actual [X]% vs expected [Y]%. Recommendation withdrawn pending thesis re-evaluation."
Decision framework: what to do when expectations and reality diverge
| Gap | What it usually means | What to do |
|---|---|---|
| Actual >> expected | Earnings momentum stronger than consensus | Good signal — but check if it's a one-off (asset sale, tax benefit). If recurring, thesis is underpriced. |
| Actual ≈ expected | Thesis is on track | Proceed normally. |
| Actual slightly below | Momentum fading but direction right | Size down. Watch next quarter closely. |
| Actual significantly below or negative | Thesis is broken at the fundamental level | Exit/exclude. Do not "wait for it to recover." The earnings were the thesis. |
| Actual positive but growth decelerating vs prior quarters | Trend peak may be passing | Check if the thesis assumed accelerating growth. If so, size down even if actual > expected-min. |
Why 扣非归母净利润同比 is the default metric
- 扣非 (excluding non-recurring items): strips out one-off gains (government subsidies, asset sales, fair value changes) that inflate headline numbers
- 归母 (attributable to parent): excludes minority interests — reflects what shareholders actually own
- 同比 (YoY): seasonality-adjusted, comparable across periods
Research reports often cite 净利润同比 (headline) which can be misleading. Always verify on 扣非 basis. If the research cited 净利润同比, pull both and compare — a large gap between 净利润 and 扣非 is itself a red flag (earnings quality issue).
Common pitfalls
Wrong stock code: Always verify code + name against user's exact text. 赛微 300456 ≠ 赛维 688418 — different companies, different fundamentals. The script takes both name and code; mismatch is a hard stop.
Wrong period: Research might cite Q1 2026 expectations but mx-data returns the latest available quarter. Check the report date in the output. If the latest quarter is different from what the expectation refers to, note the mismatch — you may be comparing apples to oranges.
One-off inflators: If 扣非同比 is strong but 净利润同比 is much stronger, the gap is non-recurring items. The 扣非 number is the real one.
Negative base year: If the prior year had a loss (negative base), even a small profit produces an astronomical 同比 growth (e.g., -0.01 → +0.01 = +100%). Check if the base was near-zero — if so, the growth % is meaningless and you should look at absolute profit instead.
API quota: mx-data has daily call limits. If you hit the limit, don't skip verification — switch to linqi-data or note "unverified, do not act."
Backup: manual verification
If the script fails or mx-data is unavailable, you can verify manually:
# Pull earnings directly via mx-data CLI
python3 skills/mx-data/mx_data.py "太辰光 扣非归母净利润同比 最新报告期"
Read the output, find the 扣非归母净利润同比 value, and compare to expectations yourself. The script just automates this — the principle is the same: pull the actual number, compare, then act.
References
references/expected-growth-sources.md— where to find analyst consensus and research expectations