What actually predicts a bad building

A building’s violation history is easy to read and hard to weigh. Forty violations sounds bad. Forty violations across two hundred apartments, all fixed inside the city’s deadline, is a building that works.

So we asked the only question that can be checked: for a building as it stood on 1 January, using only what was knowable that morning, did it receive a new Class B or C violation — the hazardous ones — in the following twelve months? Five separate years, each tested against a model that had never seen it.

The answer is yes, clearly. Across those five years the record ranks a building that went on to be cited above one that did not about 80% of the time. That is a real signal, not a coin flip, and it is why the number on a report is worth reading at all.

Our first model counted a corrected violation as evidence against the building, on the reasoning that a landlord who collects violations will keep collecting them. That reasoning is wrong, and the data says so plainly.

When we stopped counting violations the landlord had fixed inside the city’s deadline, and shortened how long old evidence carries weight from two years to nine months, the model went from 0.73 to 0.80 on buildings that already had a record — the hard cases, where every building has some history and the question is which ones are actually worse.

Read plainly: a repair the landlord completed on schedule tells you very little about what happens next. It happened, it was dealt with, and the building moved on. What predicts the next hazardous violation is what is unresolved right now.

This has a practical edge for anyone reading a violation list themselves. A long history with nothing open is a very different building from a short history with six open Class C violations, and the raw count on any city lookup will show you the first one as worse.

Under Local Law 55, every pest violation in New York is automatically Class C — the same “immediately hazardous” class as no heat in January. A single mouse sighting and a winter without heating carry the identical label.

Pest citations are common enough that sorting a building’s record by class puts them on top and buries everything else. Any tool that scores by hazard class alone inherits that distortion, and so does anyone eyeballing a list of Class C violations and counting them.

Mold has the same problem pointed the other way. It escalates to Class C either because the infestation is genuinely severe or because the owner failed to file a certification — and the public record cannot tell you which.

We also tried fitting the category weights — letting the data decide how much a gas violation matters against a pest violation, rather than judging it ourselves. It improved the model. We did not ship it.

The fitted weights wanted gas dropped from 12 to 2.0 and electrical from 9 to zero. Not because those conditions are safe: because they are life-threatening, the city treats them urgently, they get fixed, and they rarely recur. To a model asked “will this building be cited again?” that reads as harmless.

26,300 buildings in New York currently carry an open electrical violation and 20,154 an open gas violation. Publishing those weights would have told several thousand households their building was clean because the danger in it was the kind that gets dealt with. The improvement was worth 0.014, and it is not for sale.

The general lesson, if you are reading any building score: ask what it was optimised for. “Will this be cited again” and “how bad is it to live here” come apart exactly where it matters most.

A violation exists partly because a tenant called 311. Buildings whose tenants are more willing or more able to complain generate more records, and no model built on this data can fully separate “what is wrong” from “who complains.” It is the reason we report where a building sits among its own neighbourhood rather than against the whole city.

The score describes the building’s present condition. It is deliberately not a verdict on the landlord — that is reported separately, because a clean building under an owner who fixes 7% of violations on time is a different proposition from a clean building under one who fixes 90%.

And a building with no record is not a good building. It is a building the city has no record of, which is common for small ones. We show that as “we could not find it,” never as a good score.