Model Fabric

The Model Fabric is essentially where our AI learns how to price homes. At Appraisal we don’t believe in single model learning algorithms. It is all about building self-sufficient meta learners that can intelligently incorporate new predictions into pricing forecasts.

The field of AI is going through rapid development. There are new breakthroughs and techniques which advance the state of the art. The model fabric is where we can incorporate those new techniques and models into our AI.

Our AI needs extensive training to properly predict a home’s price. Essentially, training entails giving the machine a number of examples so it can eliminate bias and variance in the accuracy of its predictions, so that we can build higher quality models.

We incorporate a number of learning combinations to, not only, reduce bias and variance, but to give the model fabric the structure it needs to properly calculate and forecast property prices:

1

Linear Learners

2

Ensemble Learning

3

Boosting

4

Random Forests

5

Support Vector
Machines

6

Unsupervised
Learning

7

Generative Adversarial
Networks

8

Deep Learning

Statistical Modeling

  • Distribution assumptions about DGP
  • Validations: Siginificance
  • Parsimony is the sine qua non of statisctical modeling
  • Parsimony drives model interpretation

Machine Learning

  • No assumptions about underlying data generation process (DGP)
  • Validation: predictive accuracy. No notion of significance
  • No assumption of parsimony, predictors can be correlated
  • Lack interpretability but some methods can provide importance

Statistical Modeling: The Two Cultures, by Leo Breiman