Report Data Analysis

Problem

Sales, expense, and attendance reports often face three issues: inconsistent column names, such as “actual revenue (10k CNY)” and “execution deviation,” that cannot be mapped to standard fields directly; manual anomaly checks, where year-over-year, month-over-month, and plan-deviation values must be reviewed repeatedly; and report scope drift, where conclusions lack a traceable calculation baseline and quality gate.

How It Works

The skill maps inputs before analysis. It reads column_mapping first and falls back to keyword inference for date_col, metric_col, and deviation_col. Step 0 validates row count, field count, baseline columns, and numeric formats to prevent reports from dirty data. It then selects a baseline_type, such as year-over-year, month-over-month, or plan-based comparison, and flags anomalies using threshold rules, μ±2σ z-score checks, and mutation detection. The output follows six sections: Data Overview, Key Metric Cards, Comparative Analysis, Anomaly Localization and Root-Cause Hypotheses, Recommendations, and Calibration Notes with Quality Checks.

Boundaries

Best suited for reports with at least 30 rows and reasonably complete fields. If column names are obscure, data is heavily skewed, or year-over-year and plan columns are missing, provide explicit mappings or accept fallback to threshold-based checks. Administrative attendance counters such as tardiness, leave, and vehicle trips are treated as soft reference items rather than hard anomaly triggers.