Effect of Training Sample and Network Characteristics in Neural Network-Based Real Property Value Prediction

Abstract

Property value is its current worth in the market which needs to be ascertained for buying, selling, mortgage, tax calculation, etc. Neural network uses past data to predict future data, making it suitable for predicting value of property based on previous valuation reports. Study involves identification of variables affecting property value, data collection, data standardization, and application of neural network for predicting value of property for Pune City, India. More than 3000 sale instances have been recorded and treated as input data. Neural network is applied for value prediction. A well-generalized network is always beneficial for better predictability. A robust network is found out by testing various combinations of network parameters and dataset. Results of prediction have shown varying percentage of rate of prediction. Results obtained by change in training sample, training sample size, network characteristics like number of runs of training, and the combinations of cases given for training, testing, and cross-validation are compared. The process ensures better generalization resulting in better predictability.