NervousNewsvendor

Inspiration

Noting the latest coronavirus epidemic and the need to produce and allocate test kits for the COVID-2019, our team attempted to tackle the classical problem of using the single-period model to forecast demand and determine the optimal allocation of perishable goods. While traditional implementation requires the statistical distribution of the demand to be known, we have found during our literature review that deep-learning can be used as an alternative to solve this problem. Our project is built off of the model by (Oroojlooyjadid, Snyder and Takác, 2018) https://arxiv.org/pdf/1607.02177.pdf

What it does

Find a solution to the NVP

How we built it

The systems use a feedforward neural network with randomized number of layers and nodes based on the Hyperband algorithm. The input is a preprocessed feature vector for each product sales record and the network predicts the appropriate amount of product to stock up such that the cost is minimized. The network is trained using real sale data and optimized SGD with Nesterov momentum and a learning rate scheduler. The learning rate and the weight decay regularization is also randomized.

Compared to the original work, we added dropout and batch normalization to facilitate the learning as well as the scheduler mentioned earlier.

Challenges we ran into

Dataset cleaning, feature conversion

Accomplishments that we're proud of

Implementing a working NN in just a night!

What we learned

Data parsing from CSV, SQL, Pytorch, Tensorboard

What's next for NervousNewsvendor

For improvements to our existing project, we hope to be able to use the sequence model to implicitly predict the demands and supplies based on past data (e.g. via RNN, LSTM) instead of the fully connected network. Furthermore, comparisons to other baseline models could be made, alongside having additional visualizations for our results. Finally,