The workflow before#
Colliers works in more than 70 countries, and its investment teams screen properties for clients every week. The data they needed for that sat in a legacy enterprise warehouse, regional systems and third-party market feeds. Analysts pulled numbers from each and consolidated them in Excel, so a standard report took 2 to 5 days. A first-pass property valuation took 3 to 5 days of gathering comparables and checking figures, which held screening to about 50 deals a week. Replacing the legacy systems wasn't on the table: a full cloud migration was too costly and too risky for systems that still did their jobs.
What we built#
We built Colliers a data lake that sits beside its legacy systems instead of replacing them, then put an automated valuation model on top. The lake pulls valuation, property, market and financial data into one place. The model reads it and returns a first-pass value with a confidence score and the reasons behind it. Analysts spend their time on the deals worth a closer look.
A data lake beside the old systems
Pipelines copy data from the enterprise warehouse, regional systems and market feeds into raw, standardized and business-ready layers, and Snowflake serves data marts for valuation, investment, property and finance. The legacy systems keep running untouched. We chose to build alongside them because it gave Colliers the new capability without betting operations on a migration.
Valuations with their reasons attached
The valuation model combines several machine learning models trained on property characteristics, comparables and alternative data. Every estimate comes with a confidence score and the main drivers, such as nearby comparables or market momentum. We showed the drivers because an analyst won't take a number into a client meeting without being able to explain it.
Reports teams build themselves
Power BI reads the business-ready marts directly, so a standard report that used to take days of Excel consolidation now refreshes in under an hour. Regional teams work from the same definitions, which ends the arguments about whose spreadsheet is right.
A catalog and models that stay honest
A data catalog tracks where each field came from and flags sensitive ones, with a shared business glossary. The valuation models retrain on a schedule and are backtested by property type and region, so accuracy is measured where it matters instead of only on average.
How we delivered it
We stood up the lake first, then the marts and reporting, then the valuation model once the data under it was clean. Old and new ran in parallel during rollout, and analysts compared model values with their own first passes on live deals before relying on them for screening.
Before and after#
| Measure | Before | After |
|---|---|---|
| First-pass valuation | 3 to 5 days | Under 1 hour |
| Valuation error (MAPE) | 12 to 15% | 8 to 10% |
| Deals screened per week | About 50 | About 200 |
| Standard report generation | 2 to 5 days | Under 1 hour |
What's next#
The next step is an agent that runs deal screening end to end. It would pull comparables, run the valuation model and draft the screening memo, then hand the analyst only the exceptions and the deals that clear the bar. The lake and the model are the parts it needs.
