Using LSTM to predict outlets' supply need in Ebuy dataset

Inspiration

The Ebuy dataset provided is a time series dataset. The outlets' purchase and sales typically shows regular pattern, like periodicity. The problem is a time series forecasting problem which can be handled well using recurrent neural network.

What it does

The model will use past information as inputs, to predict the order placed of the next day. The model can "memorize" the past information implemented with hidden gates. Using the input and " memory " stored in hidden gates, the model can predict the purchased placed of the next day.

How we built it

For pre-processing, We calculated the sales data for each material and the leftover materials, then we put this together with month, day, purchased placed and purchase received for training and testing. The model was built with Pytorch and it contains a LSTM (long short term memory) layer, followed by a fully connected layer to produce the final prediction: purchased placed of the next day.

Challenges we ran into

When we plot the 'purchase placed' vs 'time', we observed that the amount rapidly decreased since April 2020, which might relate to the circuit-breaker policy. This abnormal pattern greatly affected the training and predicting of the model.

Accomplishments that we're proud of

We successfully pre-processed the raw dataset and built a prediction model from scratch

What we learned

We learnt how to solve time series forecasting problem using the recurrent neural network, specifically the LSTM model and how to build it using Pytorch. We also learnt how to clean and process data using pandas and Numpy.

What's next for Using LSTM to predict outlets' supply need in Ebuy dataset

We need to get more data to capture the pattern under covid-19 situation. We also need to improve our model to fit more complex pattern shown in the dataset.

Built With

• jupyternotebook
• numpy
• pandas
• python
• pytorch