Location

Presentation- Allen E. Paulson College of Engineering and Computing

Document Type and Release Option

Thesis Presentation (Restricted to Georgia Southern)

Faculty Mentor

Ray Hashemi

Faculty Mentor Email

rhashemi@georgiasouthern.edu

Presentation Year

2021

Start Date

26-4-2021 12:00 AM

End Date

30-4-2021 12:00 AM

Keywords

housing data, attributes, short prediction, long prediction

Description

The number of days that a home stays on the housing market (Days-On-Market—DOM) provides crucial information about the real estate market’s behavior that affects the buyer’s/seller’s decision (at the micro-level) and indicates the level of risk associated with real estate investments and identifies the housing bubbles (at the macro level). Housing data has a mixture of simple and complex attributes. A complex attribute in contrast with a simple attribute, has an array of values for a real estate property, which creates a major challenge in prediction of DOM. DOM is a binary attribute with values of “short” ( six months) and “long” (> six months). The goal was tri-fold: (a) Inclusion of complex attributes in DOM’s prediction for single-family homes in Savannah, (b) Analysis, design, and implementation of two prediction models of Naïve Bayesian (NB) and Linear Regression (LR) to predict DOM, and (c) Comparing the results to establish the prediction superiority and robustness of the models. The results revealed that LR has a superior performance (94% prediction accuracy) over NB (76% prediction accuracy). The percentage of true short prediction (TS), false short prediction (FS), true long prediction (TL), and false long prediction (FL) for LR were 98%, 2%, 82%, and 18%, respectively. TS, FS, TL, and FL for NB were 90%, 10%, 19%, and 81% respectively. The robustness superiority of LR (degradation of 0.5% in prediction accuracy) was established over NB (degradation of 1%) using a dataset with 150% increase in the noise level.

Academic Unit

Allen E. Paulson College of Engineering and Computing

Comments

This work is archived and distributed under the repository's standard copyright and reuse license, available here. Under this license, end-users may copy, store, and distribute this work without restriction. For questions related to additional reuse of this work, please contact the copyright owner.

Files over 10MB may be slow to open. For best results, right-click and select "Save as..."

Share

COinS
 
Apr 26th, 12:00 AM Apr 30th, 12:00 AM

Prediction of Days-On-Market for Single-Family Homes in the Housing Market of Savannah

Presentation- Allen E. Paulson College of Engineering and Computing

The number of days that a home stays on the housing market (Days-On-Market—DOM) provides crucial information about the real estate market’s behavior that affects the buyer’s/seller’s decision (at the micro-level) and indicates the level of risk associated with real estate investments and identifies the housing bubbles (at the macro level). Housing data has a mixture of simple and complex attributes. A complex attribute in contrast with a simple attribute, has an array of values for a real estate property, which creates a major challenge in prediction of DOM. DOM is a binary attribute with values of “short” ( six months) and “long” (> six months). The goal was tri-fold: (a) Inclusion of complex attributes in DOM’s prediction for single-family homes in Savannah, (b) Analysis, design, and implementation of two prediction models of Naïve Bayesian (NB) and Linear Regression (LR) to predict DOM, and (c) Comparing the results to establish the prediction superiority and robustness of the models. The results revealed that LR has a superior performance (94% prediction accuracy) over NB (76% prediction accuracy). The percentage of true short prediction (TS), false short prediction (FS), true long prediction (TL), and false long prediction (FL) for LR were 98%, 2%, 82%, and 18%, respectively. TS, FS, TL, and FL for NB were 90%, 10%, 19%, and 81% respectively. The robustness superiority of LR (degradation of 0.5% in prediction accuracy) was established over NB (degradation of 1%) using a dataset with 150% increase in the noise level.