What does an information value of 0.015 tell you about a scorecard feature?

## What information value measures Weight of evidence gives one number per bin. Information value (IV) aggregates those into one number per *characteristic*, so you can rank candidate features before any model is fitted. For a characteristic binned into `k` bins, with `p_good_i` the bin's share of all non-defaults and `p_bad_i` its share of all defaults: ``` WoE_i = ln(p_good_i / p_bad_i) IV = sum over i of (p_good_i - p_bad_i) * WoE_i ``` Each term multiplies a difference of shares by the log of their ratio. When `p_good_i > p_bad_i` both factors are positive; when `p_good_i < p_bad_i` both are negative. The product is therefore never negative, and neither is the total. IV is 0 only when every bin has exactly the same share of goods as of bads — that is, when the characteristic carries no information about the outcome at all. Statistically this is the symmetrised Kullback-Leibler divergence (the Jeffreys divergence) between the distribution of goods over the bins and the distribution of bads over the bins. Plain English: how differently the two populations spread themselves across the bands. ## The conventional cut-offs Credit-risk practice reads IV against rule-of-thumb bands: | IV | Reading | |---|---| | below 0.02 | unpredictive — no univariate separation | | 0.02 to 0.1 | weak | | 0.1 to 0.3 | medium | | 0.3 to 0.5 | strong | | above 0.5 | suspiciously strong — investigate | At **0.015** the feature is under the floor. The bins barely differ in their mix of goods and bads, so as a standalone predictor it does nothing, and the default action is to leave it out of the shortlist. The top band matters as much as the bottom one. A characteristic scoring 0.9 is rarely a triumph; it is usually a field that encodes the answer. Classic causes: a value written back after the outcome was known (a collections or write-off marker), an internal decision field that already reflects the risk assessment, or a snapshot taken after the performance window. Before celebrating, check *when* the field is populated relative to the observation point. ## Two caveats that separate a good answer from a rote one **IV is univariate.** It looks at one characteristic against the target and knows nothing about the others. A feature under the floor can still earn a place because it interacts with another, and two features with IV 0.35 each can be near-duplicates that add almost nothing together. IV is a screen for the shortlist, not the selection decision. **IV depends on the binning.** Cut a characteristic into more, finer bins and IV rises, because each bin can specialise on local noise; in the limit of one bin per applicant it is meaningless. This makes cross-feature comparison valid only under a shared binning policy — the same algorithm, the same monotonicity rule, the same minimum bin size (commonly around 5% of accounts per bin). Report IV alongside the bin count, and be sceptical when a characteristic's IV jumps after someone re-binned it. ## Handling the awkward cases A bin with zero defaults makes its WoE infinite and IV undefined. Merge that bin into a neighbour, or add a small constant such as 0.5 to both counts, and note that the resulting IV is only as trustworthy as the thinnest bin behind it. A bin with a handful of defaults contributes a term built on a noisy WoE, which is one more reason for a minimum bin size. IV also drifts. The same characteristic re-measured on next year's applications can move a band up or down as the population changes, which is why scorecard monitoring re-computes the WoE tables and IVs on recent data rather than trusting the development numbers forever. ## What to say in the room Give the band, then the action, then the caveat: 0.015 is under the 0.02 floor, so it comes off the shortlist; but IV is univariate and binning-dependent, so before deleting the field entirely you would check whether the binning was reasonable and whether the characteristic is required for policy or contributes in combination. And if a colleague brings you a feature at 0.9, your first move is to hunt for leakage, not to put it in the model.