Issue code reference

Every issue tsauditor raises has a short code. Use these codes to filter the report programmatically or to look up what a specific finding means.

# Filter to a specific code
report.filter(code="LEK001")

---

Profiler codes (PRF)

Code	Severity	Trigger	What to do
PRF001	WARNING	A gap between consecutive timestamps exceeds the domain threshold (5 calendar days for finance, 3× median gap otherwise)	Resample to a regular frequency, or document why irregular timestamps are expected
PRF002	WARNING	3+ consecutive NaN values in a column (sensor) or 5+ (finance)	Handle the affected span via interpolation, forward-fill with a limit, or explicit dropping — check whether the cluster corresponds to a known outage
PRF003	INFO	ADF test p-value > 0.05 — column is non-stationary	Consider differencing or log-transforming before modeling; non-stationarity is expected for price series but can bias many ML methods
PRF004	CRITICAL	Duplicate timestamps in the DataFrame index	Remove or aggregate duplicates before any further processing — duplicate timestamps silently corrupt rolling, lag, and resampling operations
PRF005	WARNING	A run of 2+ consecutive large timestamp gaps	Review the cluster — likely a data feed outage or collection failure requiring explicit handling
PRF006	WARNING	A column has more than 30% missing values overall	Consider dropping the column or imputing carefully; check whether the missingness is informative (i.e. whether absence itself carries signal)

---

Anomaly codes (ANO)

Code	Severity	Trigger	What to do
ANO001	WARNING	The same value repeats consecutively beyond the stuck-value window (5 for finance, 3 for sensor)	Investigate whether this is a sensor failure, a forward-fill artefact, or genuinely flat data — the evidence dict includes max_stuck_duration
ANO002	WARNING	A value exceeds the z-score threshold (5.0 for finance, 3.5 for sensor, 4.0 default) or falls outside 1.5×IQR	Decide whether to winsorize, transform, or treat as a data error; finance domain uses a wider threshold to avoid flagging legitimate fat-tail events
ANO003	WARNING	A value deviates sharply from its immediate neighbors (rolling z-score > spike threshold)	Examine the flagged timestamps — these are contextually extreme even if globally plausible, often indicating data-entry or data-feed errors

---

Leakage codes (LEK)

Code	Severity	Trigger	What to do
LEK001	CRITICAL	Feature–target AUC >= 0.80 (binary target) or Spearman	ρ
LEK002	WARNING	Feature's peak cross-correlation with the target falls at a positive lag	Inspect how the feature is constructed — a positive-lag peak means the feature aligns most strongly with the future target, suggesting it encodes information not yet available at prediction time
LEK003	WARNING	Feature correlates with the future target beyond what the target's own autocorrelation explains	Verify the feature uses only past data; this pattern is the signature of a forward-looking or centered rolling window

---

Severity levels

Severity	Meaning
CRITICAL	Must be resolved before modeling — the data quality issue will directly corrupt model training or evaluation
WARNING	Worth reviewing — may or may not require action depending on domain context
INFO	Informational — no immediate action required, but worth knowing before proceeding

---

Adding a new code

If you're contributing a new check, add a corresponding entry to tsauditor/report/remediation.py following the existing pattern. See Contributing for details.