Predicting capital appreciation in real estate investments using neural networks
Overview
In 2017, my parents decided to invest in real-estate, Pune is an expensive city and this decision hence brought with it worries of whether the price of the property will rise or drop. This conundrum got me thinking that there are certain factors that determine this rise or drop in property price, was it possible to collect historical data pertaining to all these factors, train an ML model on it and see if it can predict possible rise/fall in property prices in the future.
Work done
This minor project, on advice of my guide, eventually became part of a larger research project based on comparing the feasibility of various ML models in analyzing the factors that contribute to real estate prices and then eventually predicting capital appreciation for real estate in various localities.
My contributions
I continued working on my model to get the ecessary results and in the end , all of us in the team collated our results into a research paper.
I built the neural network in Python, since thiswas my first tryst with an ML model there was a considerable amount of learning I had to do before I could actually start writing code. If I had to depict my laring phase in points here is how I went about it:
-
First, I started scouring and studying papers that talked about real estate markets and capital appreciation. Ihad to understand how to go deciding which factors contributed to the capital appreciation and which were common factors.
-
Then I started studying up on neural networks, understanding their mechanisms and structures, their types and the activaton functions.
-
FInally, I started reading up on how I could implement my own neural network. Python seemed like a good choice because I had been hearing a lot about how it was suitable to implement prototypes of ML models quickly, I eventually settled for using the scikit library and inplemented it.
The neural network
The neural network I built was a Back Propagation neural net that used the conjugent gradient algorithm which resulted in a faster convergence rate. It was trained on the Boston Housing Information datatset whic had a total of 13 attributes. My nerual net had 3 layer: the first layer was the input layer with 13 nodes matching the number of attributes in the dataset, the middle layer had 50 nodes while the ouput layerhad just 1 node. The output was a simple binary value, 1(yes) for signifying positive capital appreciation while 0 (false) for otherwise.
Below is a screenshot from one of the training phases during the development days:
Epilogue
The research paper was presented at IEEE-Pune Section and PCCOE’s 3rd International Conference on Computing, Communication, Control and Automation 2017 (ICCUBEA 2017). The authors were Abdul Wasay (me), Aishwarya Murkute, Vinit Vaichole, Rohan Rao and Deepa Abin.