F1 Score

The F1 score is a metric used to judge how well a classification model performs, and it is defined as the harmonic mean of two other metrics — precision and recall. Precision asks, of everything the model flagged as positive, how many really were positive (it penalizes false positives). Recall asks, of everything that really was positive, how many the model caught (it penalizes false negatives). The F1 score combines the two into a single number between zero and one, computed as two times precision times recall, divided by precision plus recall. Because it uses the harmonic mean rather than a plain average, the F1 score is dragged down by whichever of the two is weaker — a model cannot earn a high F1 by being great at one and poor at the other. Only when precision and recall are both strong does the F1 score climb toward one.

The F1 score matters because accuracy alone can lie, especially when classes are imbalanced. If only two percent of cases are truly positive, a lazy model that predicts "negative" every time is ninety-eight percent accurate yet useless — it never catches a single real positive. The F1 score exposes that failure, because such a model would have terrible recall and therefore a terrible F1. By balancing precision and recall, the F1 score rewards models that are both careful about what they flag and thorough about what they catch, which is exactly what you want when the positive class is rare and the cost of missing it or crying wolf is real. This is why F1 is a standard scorecard metric for classifiers, particularly on imbalanced problems where raw accuracy is misleading.

In marketing, classification models are everywhere, and the F1 score is a natural way to grade them. A predictive lead-scoring model classifies leads as likely-to-convert or not; a churn model classifies customers as likely-to-leave or not; a spam or fraud filter classifies messages or transactions. Each of these is a classifier with a rare positive class and a real cost on both sides — flag too loosely and you waste sales effort or block good customers (poor precision); flag too tightly and you miss the conversions or the churn risks that mattered (poor recall). The F1 score captures both costs at once, so a lead-scoring model with a high F1 is one that identifies genuinely promising leads without drowning the sales team in false hopes or overlooking the real ones.

Using F1 in these settings keeps teams honest about a model's real usefulness. A lead-scoring model that boasts high accuracy might simply be predicting "won't convert" for almost everyone, since most leads do not convert — accurate but worthless, because it never surfaces the promising leads. Its F1 score would be poor, revealing the truth. That said, F1 weights precision and recall equally, and sometimes a business cares more about one: missing a high-value churn risk may be far costlier than a false alarm, or vice versa. When the two errors carry unequal cost, teams weight recall or precision more heavily (a weighted variant of the F1 idea) rather than treating them as equal. The F1 score is the balanced default; the business context decides whether that balance is the right one.

Use the F1 score to judge a classifier when the positive class is rare and both false positives and false negatives carry real cost — which describes most marketing use cases, from lead scoring to churn to fraud. Read it as the balanced summary of precision and recall, and always look at precision and recall individually too, since the F1 alone hides which of the two is dragging performance. Decide deliberately whether equal weighting is right for your problem, and lean on a weighted variant when one error is costlier than the other. Compare F1 across model versions to see whether a change genuinely improved the balance, rather than trusting accuracy, which can flatter a model that simply predicts the majority class.

The traps are trusting accuracy on imbalanced data (where a do-nothing model looks great but is useless); reporting F1 without also inspecting precision and recall (so you cannot see which one is weak); applying equal weighting when the business clearly cares more about one error than the other; and treating any single number as the whole story of a model's fitness. Discipline means using the F1 score as the balanced precision-and-recall metric it is, reading it alongside its components, weighting it to match the real costs of each error, and remembering that the right metric follows the business decision the model is meant to support.

F1 score — the harmonic mean of precision and recall — balances false positives and false negatives in one classification metric, exposing on imbalanced data what raw accuracy hides.

Etymology: source.

Search interest for this term over the last five years:

View interest-over-time on Google Trends →

Areas of marketing where f1 score is a core concern: