A new data source lands on your desk. Parcel-level flood layer, a lien dataset, a curb-appeal score from a vendor, doesn't matter which. The reflex is to think you need a clean retrain: new feature matrix, new hyperparameter sweep, new model card, new sign-off cycle. For most additions that's overkill, and it's also why half the good signal sitting in vendor trial accounts never makes it into a production AVM.
Treat it as a column, not a rebuild
If the new source resolves to one row per property and lines up with your existing record keys, it's an append to the feature store, not a new project. The steps that matter:
- Land the raw feed and join it to your parcel or MLS ID exactly the way every other feature joins. If the vendor hands you a flat file keyed to APN, that's the easy case. If it's keyed to lat/long or a raw image tile, you need a resolution step before it ever touches the model, and that step is where most of these projects quietly die.
- Run the new column through your existing feature pipeline's validation checks: null rate, distribution shift versus training vintage, leakage checks against your target. Nothing here is specific to the new source. It's the same gate every feature already passes through.
- Score it as a challenger feature on a frozen champion model, not by retraining the whole ensemble. Gradient-boosted tree models in particular don't require a full rebuild to test one new input. Hold the tree structure and learned splits where you can, add the candidate column to a parallel model trained on the same window, and compare residuals property by property. A handful of counties, a handful of thousand parcels, is enough to tell you if the new feature is pulling weight.
- If it helps, fold it into the next scheduled retrain rather than forcing an off-cycle push. Most shops already retrain quarterly or semi-annually. Riding that cadence means the new feature gets the same backtest and sign-off as everything else, instead of a one-off exception that nobody remembers the provenance of eighteen months later.
Where this breaks down is image-derived or unstructured signal. A tabular field has a schema and a known shape; an embedding vector doesn't fit into a spreadsheet mental model, and teams end up building a bespoke ingestion path for every imagery vendor they try. That's the gap a plug-in vector keyed to your existing record ID is built to close. If you're evaluating curb-appeal or neighborhood-context signal as a new input, it's worth seeing what one of those products looks like before you build your own ingestion layer: AVM Imagery delivers that kind of signal as a per-property file that joins the same way any other column does.
When you do need a full retrain
Not every addition is a column swap. A few honest triggers:
The new source changes your feature space enough that pairwise interactions matter. If the candidate feature is highly correlated with three existing inputs and your model is a linear blend or a shallow tree ensemble, bolting it on without retraining the interaction terms will just steal credit from features that were doing fine. Tree-based models handle this more gracefully than linear ones, but it's still worth checking feature importances before and after.
You're changing the target window, not just the inputs. Adding a feature and shifting from a 90-day to 30-day comp window are two different projects wearing one ticket. Don't let them merge.
Compliance needs a full model card regardless of mechanism. Some lenders require full revalidation any time a new data source touches the valuation, even if the technical lift is one column. That's a governance call separate from the modeling work, and keeping the two conversations apart stops the engineering team from inheriting a six-month review cycle for a one-day pipeline change.
The honest version of "add a data source without retraining from scratch" is: build your feature pipeline so that adding a source is boring. Keyed joins, versioned feature definitions, a shadow-scoring habit before anything goes live. The sources are the easy part once that's in place.
If curb-appeal and parcel-context signal is the gap in your pipeline, take a look at what a ready-to-join imagery feature looks like.
