One of our first larger scale projects was a real estate price estimator. It was put in broad production use in 2009 and most banks in Sweden use it today as a part of their standard process when giving out house loans. It uses different variables from historical data, including object properties, demographic information and nearby points of interests. It uses a custom developed neural network architecture, called a Fully Forward Connected Perceptron Array and has great similarities to Google's "Inception" architecture. The FFCPA was a very early example of a deep neural network.
The limitation of the original system was hardware. Although it was a deep neural network, it still just had less than 500 weights in it.
With the advent of GPU computing we are able to use much larger systems with much more data. The original system had 390 weights and was trained on 20,000 data points. The second generation uses millions of weights and millions of data points.
Just by using a larger system and more data, we get a radical improvement in performance:
More importantly however we can now use other, much more complex data sources.
These additions can help to radically improve valuation performance beyond what any machine or human valuation has been capable in the past.