Brady Anderson

Wells Fargo, Jun 2020 – Present

House price forecasts for every US metro, and remediation data that has to be right

I designed and maintain a Microsoft Fabric pipeline that turns BLS and other public economic data into house price forecasts for every US metro area. In the same role I built Python/SQL pipelines for Tier-1 remediation programs over $15M and a validation suite the department adopted.

  • Microsoft Fabric
  • Python
  • SQL
  • Forecasting
  • Snowflake
Role
Senior Lead Data Analytics Engineer
When & where
Jun 2020 – Present, Minneapolis, MN
Context
Full-time role
Every US metrocovered by the house price forecasts
$15M+Tier-1 remediation programs supported

The original systems are proprietary, so no code, screenshots, or data from them appear here. I’m happy to walk through the design in conversation.

House price forecasts for every US metro

House prices move differently in every metro, so a single national number isn't enough. I designed, and now maintain, the pipeline that produces house price forecasts for every metropolitan statistical area (MSA) in the country.

It runs on Microsoft Fabric. It ingests public economic data from the Bureau of Labor Statistics and other government sources, conforms each series to consistent metro definitions, and feeds the forecasting step that produces a forecast for each metro.

The interesting problems are the ones every forecasting pipeline built on public data runs into. Sources publish on different schedules and revise earlier figures, metro definitions change over time, and every forecast has to be built only from data that was available on its as-of date.

The production system is proprietary, so here is the same problem rebuilt from scratch with public data and a simple model I wrote. Watch releases cross the "as of" line and land in the lakehouse. When a quarter of house price data arrives, the forecasts refresh across the map. Pick any metro to see its history and forecast.

Interactive illustration built from real public data and my own model.

As of
Oct 6, 2026
Latest release
BLS LAUS (Aug 2026)
Forecasts trained through
2026 Q2
3.45 ptsaverage forecast error, 4 quarters ahead
3.92 ptsnaive baseline: next year repeats last year
4,917backtested metro forecasts, 2022 Q3 – 2025 Q2
1990 – 2025training window, every forecast point-in-time
How the model works, and what it doesn’t prove

A pooled ridge regression predicts each metro’s log change in the FHFA index over the next four quarters from: metro price growth, last 4 quarters; metro price growth, the 4 quarters before that; median metro price growth, last 4 quarters; change in the 30-year mortgage rate over 12 months; change in the national unemployment rate over 12 months; metro unemployment minus national (0 before metro history begins). Each forecast only uses data published by its as-of date, using approximate release calendars, and is trained only on outcomes already known then.

Honest result: in this backtest, rates and unemployment add little beyond price momentum (3.40 pts with momentum alone). The bands are the 10th–90th percentile of past errors. This is my own model on public data, not the production system or its model.

Sources: FHFA House Price Index (metro, all-transactions) · BLS Local Area Unemployment Statistics · Freddie Mac PMMS via FRED (MORTGAGE30US) · BLS national unemployment rate via FRED (UNRATE) · Census 2023 gazetteer and OMB metro delineations. Snapshot built 2026-10-06.

Remediation pipelines that have to be right

I engineered Python and SQL pipelines for one of the bank's highest-visibility remediations, identifying the customers owed corrections under Tier-1 programs worth more than $15M. Auditability and scale were hard requirements. If a pipeline misses a customer, someone doesn't get the refund or correction they're owed.

A validation suite the department adopted

I refactored legacy real-estate ETL and built a data-validation suite around it. It only mattered once other people could use it, so I designed it with the same checks and the same language for everyone, and with logs a reviewer could follow months later. The rest of the department adopted it.