Your Past Reports Are a Data Set: Data-Driven Adjustments from Your Own Work

"Please provide support for your adjustments." Every appraiser has read that line in a revision request. Most of us have also written "adjustments are market derived" in commentary and hoped it would be enough. It isn't, and under UAD 3.6 it's even less likely to pass a reviewer. The good news: you're sitting on years of evidence. Every report you've delivered holds sales you verified and adjustments you defended. This guide covers the main methods for data-driven appraisal adjustments (paired sales, grouped data, regression) and how to turn your own archive into support you can cite.

Key takeaways

• Fannie Mae requires adjustments that reflect the market's reaction and accepts statistical analysis, modeling, paired sales and other commonly accepted methods.
• Paired sales are easy to explain but rare. Grouped data uses more sales. Regression handles several variables at once.
• Your own past reports are a clean, verified data set. Your medians by town and price tier are a consistency check and a starting point.
• History supports; it doesn't decide. The current market and the current comps still drive each adjustment.

What the GSEs expect

Fannie Mae's Selling Guide section on adjustments to comparable sales is short and worth reading in full. The points that matter most:

• Adjustments must reflect the market's reaction to differences, which the guide calls market-based adjustments.
• It gives an example: a rule-of-thumb adjustment of $20 per square foot is inappropriate when market analysis indicates $100.
• Adjustment rates can be determined "through statistical analysis, modeling, paired sales, or other commonly accepted methods."
• "A statement only recognizing that an adjustment has been made is not acceptable."
• Fannie Mae does not have specific limitations or guidelines for net or gross adjustments. If you learned the old 15% net / 25% gross rule of thumb, it isn't in the current guide. Large adjustments are fine if they're supported.

So the question is never "is this adjustment big?" It's "can you show where it came from?"

Method 1: Paired sales

Find two sales that are alike except for one feature. The price difference, after accounting for anything else that differs, indicates what the market paid for that feature.

A worked example (illustrative numbers):

If the market was flat between March and April, the pair indicates about $14,000 for the second garage bay. One pair isn't proof. Three or four pairs pointing to $12,000 to $15,000 is support.

Strengths: easy to explain to a reviewer, a borrower or a judge. Weaknesses: truly matched pairs are rare, especially in rural or mixed markets. Small samples are noisy.

Watch out: A "pair" that differs in two things (garage and condition, say) doesn't isolate either one. Adjust for the other difference first or throw the pair out.

Method 2: Grouped data analysis

Instead of single pairs, compare groups. Take all sales in the neighborhood over a period and split them by the feature: with pool vs. without, or GLA bands. Compare median price (or median price per square foot) between groups.

This uses more data than paired sales, so it's less sensitive to one odd sale. It's also easy to show in a small table in your workfile. The weakness is that groups can differ in more than one way. If homes with pools are also newer, the "pool" difference is partly an age difference. Look at the groups' other characteristics before you trust the gap.

Method 3: Regression

Regression fits a line (or a plane, with several variables) through many sales at once. Simple regression of price on GLA gives you a dollars-per-square-foot slope. Multiple regression estimates several features at once (GLA, age, lot size, garage) while holding the others constant.

This is the same family of math behind many automated valuation models, and the place where "machine learning appraisal" has real meaning. Used by an appraiser on a well-chosen data set, it's a legitimate support method. Used blindly, it's a black box.

Practical rules:

• Choose the data set like you choose comps. Same market area, similar property type, a sensible time window.
• Check sample size. A regression on 12 sales with five variables won't tell you much.
• Look at the residuals. If one sale is far off the line, find out why before you trust the slope.
• Sanity-check the sign and size. If regression says a bedroom is worth negative $8,000, that's probably GLA and bedrooms fighting each other in the model, not a market fact.
• Keep it in the workfile. The output, the data set and your reasoning, so you can reproduce it.

Other commonly accepted methods

• Cost and depreciation. Useful for features with a clear replacement cost (a new roof, a detached garage), adjusted for depreciation and market acceptance.
• Sensitivity analysis. Look at how the spread of adjusted comp prices narrows as you try different rates. The rate that best brackets the comps is a supported indication.
• Market interviews. Agents, builders and other appraisers. Useful context, weaker as primary support.

Most good reports reconcile more than one method.

Method 4: Your own archive

Here is the method most appraisers overlook. You already have thousands of verified sales and the adjustments you made to them, in your old software's files.

Why your archive is good data

• You verified it. These aren't scraped records. You confirmed the GLA, the condition, the sale terms.
• It's local. It's from the towns, price tiers and property types you actually appraise.
• It shows your own consistency. If you've used $35 per square foot for GLA in one town for three years and suddenly used $55, you should know why.
• It holds pairs. Two of your old comps in the same subdivision that differ in one feature may be exactly the paired sale you need.

How an archive becomes numbers

The process looks like this, whether you do it by hand in a spreadsheet or let software do it:

1. Export. Most desktop programs can export MISMO 2.6 GSE XML. ClickFORMS files and those exports hold the comps grid, the adjustments and the subject data. (The switching guide covers getting them out.)
2. Parse. Pull out each comp: address, sale price, date, GLA, features, and each adjustment by type.
3. Clean. Throw out or flag records that are unclear: missing GLA, odd adjustment types, duplicate sales.
4. Group. By town (or neighborhood) and price tier, so a $250,000 ranch and a $900,000 colonial aren't averaged together.
5. Summarize. Median rate for each adjustment type (GLA per square foot, garage per bay, bath, basement finish, condition step), with the count of reports behind it.

Medians beat averages here. One odd report with a huge location adjustment won't pull a median far.

What a summary looks like

An illustrative example for one town:

That table doesn't tell you what the market is paying today. It tells you what you concluded across many reports, which is a strong consistency check and a starting point for testing against current sales.

Tip: When your current analysis lands far from your own median, that's not wrong by itself. Markets move. But it's the first thing a reviewer comparing your reports will notice, so explain it in a sentence.

Using it in a report

A sentence like this in your commentary is a lot stronger than "market derived":

"The GLA adjustment of $40 per square foot is supported by grouped analysis of 18 sales in the subject's subdivision over the past 12 months and is consistent with the appraiser's median of $38 across 63 prior assignments in this town and price range."

That's specific, testable and yours. It also fits inside UAD 3.6's commentary limits, which reward exactly this kind of short, data-backed sentence.

Firms: one archive, many appraisers

A firm's archive is bigger and more valuable than any one appraiser's. It also raises a question: whose medians are they? A few practical points:

• Separate by appraiser when it matters. A trainee's early reports and a 25-year certified appraiser's reports shouldn't carry the same weight without review.
• Share the pairs, not just the rates. A paired sale found in a colleague's report is evidence you can verify yourself.
• Watch the dates. A median built on 2021 reports in a market that turned in 2022 may need a time cut.
• Keep it in the workfile. If you rely on firm data, keep the summary you used with the report. USPAP's Record Keeping Rule requires the workfile to be retained at least five years after preparation, or at least two years after final disposition of a judicial proceeding in which you testified, whichever is later.

Where AI fits, and where it doesn't

Software can do the boring part: reading thousands of old files, cleaning them, grouping them and computing medians in seconds. That's "AI that learns your adjustments" in the honest sense: it learns from your own work, and it shows its counts.

What it shouldn't do is choose your adjustment for you. A rate that appears in your grid with no source is exactly the "statement only recognizing that an adjustment has been made" the Selling Guide rejects, just automated. The broader argument is in AI appraisal software in 2026. And data-backed adjustments are one of the best ways to prevent the "support your adjustments" request, as covered in how AI catches AMC revision requests before you deliver.

Carat does this with a bulk legacy import: upload a zip of ClickFORMS files or MISMO 2.6 XML, and it builds a searchable comps history and median adjustment rates by town and price tier from your own reports. Anything unclear goes to a review queue, never silently into a report. If you're coming from ClickFORMS specifically, see migrating from ClickFORMS. It's part of the $99 monthly plan, and you can start an account here.