What is ground truth?
Ground truth is the answer accepted as correct for a given example, against which a model’s output is trained and scored. For a fraud model, ground truth is whether a transaction was later confirmed as fraudulent. For a document classifier, it is the category a qualified reviewer assigned. Data labels are the recorded form of ground truth, and the quality of a model is measured by how closely its predictions match them on test data it has never seen.
The phrase implies certainty, but ground truth is often a judgment. A radiologist’s reading, a claims adjuster’s decision, or a customer’s self-reported reason for canceling can each be wrong. Good teams treat ground truth as the best available answer and record where it came from.
Why ground truth matters
Every accuracy figure in a vendor deck is a comparison against some ground truth. If that reference is weak, the figure is meaningless. Ground truth also sets the ceiling on what a model can learn. A model trained to copy inconsistent human decisions will reproduce the inconsistency, and a model evaluated against noisy labels will look worse than it is, or better, with no way to tell which.
Ground truth in practice
A regional bank builds a model to flag card transactions for review. The obvious ground truth is the chargeback record, since a confirmed chargeback means fraud occurred. Two problems appear. Chargebacks arrive weeks after the transaction, so the most recent data has no labels yet. And many fraudulent transactions are never reported, so some “legitimate” examples are actually fraud. The team handles the first by training only on transactions old enough for labels to have settled, and the second by having analysts review a random sample of unflagged transactions to estimate how much fraud the labels miss.
What to watch for
- Ask where the ground truth came from, who produced it, and how long after the event it was recorded.
- Have two people label a sample independently, because a model cannot be measured as more accurate than the labels it is scored against.
- Watch for labels that leak the answer, such as a field filled in only after the outcome was known.
- Re-check ground truth when a process changes, since yesterday’s correct answer may no longer apply.