Microventures

Inspiration

Allow the service economy to own infrastructure in the ...

What it does

Software to Empowered The Service Economy To Narrow The Affordability Gap

How we built it

"""

TIME SERIRES MODELING WITH REINFORCE LEARNING TO PREDICT FUTURE HOUSE PRICES

Step 1: Data Processing

• Call the zillow bridge API for all the zipcodes in Miami Dade and request

  • Sales Price
  • Rent Price
  • Adjusted Area
  • Date
  • Latitude & Longitude Location

• Have longitude and latitude cordinates of where the main rezoning of Miami-Dade it's happening. Rapid Transit Zones

• Convert JSON files in data frames to convert to

• Create pandas Data Frames for property sales, property rent values, Have a data frame that has the figures of where these latitudes & longitudes happend to be.

Step 2: Data exploration and visualization

• Visualize Historicals of Miami-Dade County in 4 main areas:

  • Sales
  • Rents
  • Inventory

• Visualize Historicals of zipcodes that have 3 main variables above historicals

Step 3: Decide on model approach and build it.

• Create a reinforcemnet model that has 3 main factors that affected evenly.

  • House Appreciation & Rent Appreciation: Linear Regression
  • Zoning District: Logistic Regression
    • This process was used: As per paper:Housing Prices Prediction with a Deep Learning and Random Forest Ensemble Bruno Klaus de Aquino Afonso1, Luckeciano Carvalho Melo2, Willian Dihanster Gomes de Oliveira1, Samuel Bruno da Silva Sousa1, Lilian Berton1 We extract image features by using transfer learning from a mobile version of NAS- Net model [Zoph et al. 2018] . This architecture is obtained via Neural Architecture Search [Zoph and Le 2016], where a RNN controller network optimizes convolutional architectures by using a reinforcement learning algorithm called Proximal Policy Opti- mization [Schulman et al. 2017]. The final NASNet architecture achieved state-of-the- art results on ImageNet, CIFAR-10 and COCO datasets and learned image features that are generically useful and that can be transferred to other computer vision problems [Zoph et al. 2018]. Firstly, we resized the image to the dimensions of 256 × 256, and applied the pre- 5 trained NASNetMobile architecture available in Keras library . Then, we applied average poolings in the output tensor to map features to a latent space of 264 dimensions. We fed the KISS model with such final features for each image in the housing prices prediction database.
  • Create SARIMA Model:
    • Seasonality
    • Autocorrelation:
    • Decompositon

Step 4: Training set and validate using the test set

• Use data from 2010-2020 to train given zip codes
• Run model with test sets of 2021

Step 5: Fine-tune the model and make the prediction

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Microventures

Built With

• javascript
• python