AVM Brochure to help explain how AVM models work

Attached is a brochure that explains Teranet's AVM's.  It's designed for the Financial Markets, but does have sections that explains how our models work if you have customers wanting an explanation.

Full brochure attached and explanations below.

The Teranet AVM Methodology

The Teranet AVM leverages a cascade of two proprietary statistics models to provide the best valuation of
properties based on available neighbourhood information.

The Teranet Machine Learning Model is a boosting-based machine learning model. Boosting is a machine learning ensemble meta-algorithm used primarily to reduce bias and variance in supervised learning algorithms.
This model leverages a rich dataset including land registry data (i.e. title data, sales data and parcel data), property data (i.e. tax assessment data and structural data), demographic data and neighbourhood specific data, and other derived data from internal data processing.
A machine learning model is able to learn from these transactions and the contribution of each attribute
on a property valuation in a short period of time –a feat not easily achieved by human statisticians.

The Teranet Sales Comparison Model is a comparable sales model based on two primary underlying processes: the selection of ‘comparable sales’ meaning the determination that another recent and proximal property
sale is a suitable comparable; the adjustment to valuation for significant differences between the subject property and the property of the comparable sale. This model leverages a rich dataset including land
registry data (i.e, sales data and parcel data), property data, neighbourhood specific data, and other data derived from internal data processing. The Sales Comparison model evolves through adjustments to the definition of
‘comparable neighbourhoods’ from which comparable sales are selected and the ongoing
responsiveness of the valuation adjustments as the weighting of attributes may change over time.
Teranet’s AVM product is blend of estimation models where values are chosen from the model that displays
the best accuracy profile for a region and property type based o n back testing. Each model’s performance
is reviewed on a monthly basis and changes are made accordingly.