Housing-Price-Time-Series-Forecasting

Introduction

We aim to develop a time series forecasting model for predicting housing prices in Singapore's HDB market using transaction data and geographical information.

Running the Notebooks

Look out for the serialisation of the notebooks (e.g. 0X where X represents the order it is to be run). 00 to 10 can be found under src(00-10) while the rest (11-15) are found in the root directory.

Data Cleaning & Processing

mrt_coordinates_opening_dates.ipynb

This file scrapes the MRT stations in Singapore through the Wikipedia website as well as their respective coordinates and opening dates. The results are written to the mrt_stations.csv file

school_coordinates.ipynb

This file scrapes the primary schools in Singapore through the Wikipedia website as well as their respective coordinates from the OneMap API/Site. The results are written to the primary_school_coordinates.csv file in the PointsOfInterest folder

mall_coordinates_opening.ipynb

This file scrapes the shopping malls in Singapore via Wikipedia and obtaining their respective coordinates and opening dates.

data_cleaning.ipynb

This file deals with the following:

• Cleaning the HDB resale flat prices dataset
  • Converting remaining_lease into float for standardisation purposes and creating an address column that concatenates block and street_name.
  • Merging the two HDB resale flat prices datasets into one.
  • Calculating the coordinates of each unique HDB flat in the merged dataset using OneMapSG API.
• Cleaning the mrt_stations.csv file
  • Remove rows where the MRT stations are yet to be opened (e.g. Brickland MRT Station).
  • Ensure that the opening dates are standardised,
• Cleaning the BTO supply dataset
  • Dropping redundant rows and remove rows with empty values.
  • Standardise the dates to DD-MMM-YY.
  • Add the coordinates of each BTO.
• Cleaning the SORA dataset
  • Creating new column SORA Value Month that will match the HDB transaction record dates.
  • Aggregating compounded SORA based on the month.

Feature Engineering

add_sora.ipynb

Adds the SORA values associated with the HDB transaction record dates.

mrt_stations_nearby_and_BTO_supply.ipynb

• Identifies the MRT stations that are within a 1km radius from each HDB flat if the opening date of the MRT station is before the flat's transaction record date. For HDB flats that do not have any MRT stations within a 1km radius, the nearest MRT station is noted. Geometric distances between points are taken into account using GeoPandas.
• Identifies the number of BTO flats that are launching within a 4km radius from the flat (the launch dates of the BTOs are compared with the flat's transaction record date - only launch dates that happen before the transaction record date are taken into account) and the supply of units associated with these BTOs.

malls2kmRadius.ipynb

• Identifies the malls that are within a 1km radius from each HDB flat if the opening date of the mall is before the flat's transaction record date. Geometric distances between points are taken into account using geopy.

distance2cbd.ipynb

• Calculates the distance from HDB to Central Business District area (Coordinate(lon=103.851784, lat=1.287953)). Geometric distances between points are taken into account using geopy.

word_embedding.ipynb

• Combine name and distance information about POIs (mrt, sch, mall)
• Convert to POI density vector using spacy

pre_modelling.ipynb

This file deals with processing the data prior to model building, including scaling for numerical features and one-hot encoding for categorical features. It also normalises the resale_price and drops relevant columns not needed for model building.

Model Building

xgboost_model.ipynb

• Train xgboost regressor on working dataset
• Predict 2024 resale prices and evaluate performance

random_forest.ipynb

We build 2 Random Forest Regression models using Out-of-bag (OOB) method and 10-fold Cross Validation. Both models are evaluated by minimising Mean Absolute Error (MAE). Their performances across various criteria are compared against each other.

CatBoost.ipynb

This file utilises the Catboost gradient boosting algorithm to train the cleaned, normalised, and one-hot encoded data from data_sklearn_models before testing on the 2024 data. Following training, evaluation and analysis of model performance is conducted.

gnnwr.ipynb

This model uses incorporates geospatial features (latitude, longitude) using neural networks.

lstm.ipynb

We first prepare the data for model training by performing necessary preprocessing steps and splitting it into distinct sets for training, validation, and testing purposes. We then train and evaluate the Long Short-Term Memory (LSTM) neural network model.

Properties:

• An LSTM model with three layers and dropout regularization is used to capture temporal dependencies.
• Uses the Adam optimizer and mean squared error loss function for regression tasks.
• Overfitting is prevented with early stopping, monitoring validation loss with a patience of 5.
• The model is trained for 100 epochs with a batch size of 32, with performance evaluated on the validation set to prevent overfitting.