Bad data, wrong and looking normal
Where operational data ends up
Some data is not dark or siloed, it is simply wrong. A stuck sensor reporting the same value for a week, a flow in the wrong unit, a gap backfilled with zeros. The dangerous part is that wrong values sit in the same tables as right ones, wearing the same font.
What it means
The shapes bad data takes
Bad data is rarely dramatic. It is a sensor drifting out of calibration by half a percent a month, a timestamp in local time in a UTC table, an asset renamed on one side of an interface, a manual entry with a slipped digit.
Each of these looks like data. Each aggregates, plots and averages without complaint. Nothing about a mean value tells you that one of its inputs was frozen for nine days.
- Stuck and drifting sensors keep reporting plausible numbers.
- Unit and timezone mix-ups survive every aggregation applied to them.
- Gaps, zeros and duplicates change averages nobody re-checks.
- Manual entries carry typos into systems that trust them forever.
What it costs
What one wrong number does downstream
Bad data does not stay where it was born. It flows into KPIs, reports and models, and everything computed from it inherits the error without inheriting the doubt.
Reports argue with each other
Two dashboards disagree, and a meeting that should decide something spends its hour deciding whose number is right instead. Trust, once spent there, does not come back for free.
Models learn the fault
A forecasting model trained on a stuck sensor learns that stillness is normal. Bad data does not just mislead people, it quietly trains every model downstream of it.
The fix arrives too late
Errors found at reporting time are found weeks after they began. By then the bad stretch is woven into totals, and unpicking it costs more than catching it would have.
Our approach
Catching it at the door
Bad data is cheapest to catch at ingestion, while the context that explains it still exists. That means quality has to be a property of the pipeline, not a quarterly audit.
Flags on every value
Stuck, out-of-range, drifting and gap conditions are detectable mechanically. Flag them as data arrives, and every downstream reader, human or model, can decide what to trust.
Lineage that carries doubt
When a KPI can show which raw signals fed it, a suspicious number explains itself in minutes. Quality flags that travel with the lineage make doubt cheap to resolve.
Units and time as first-class facts
A value without a unit and an unambiguous timestamp is a guess. Storing both as facts the platform enforces removes the two most common ways numbers go quietly wrong.
Common questions
Asked and answered
What is bad data?
Bad data is recorded data that misrepresents reality: stuck or drifting sensors, wrong units, timezone errors, gaps, duplicates and typos. It is dangerous because it looks exactly like good data and aggregates without complaint.
What does bad data cost an operation?
Studies have found that only a small fraction of company data meets basic quality standards, and the operational cost shows up as reports that disagree, models trained on faults, and meetings spent arguing about whose number is right instead of deciding.
How do you detect bad data early?
Detect it at ingestion, where the context still exists: automatic flags for stuck, out-of-range and drifting signals, enforced units and timestamps, and lineage so any downstream number can show which inputs it trusted.
The rest of the split
Where the other data points go
The drain has five outlets, and each one fails an operation in its own way. This page covers one of them. The other four are worth ten minutes each.