Store Item Demand Forecasting Project

Data: kaggle Github: github

• pandas, numpy, matplotlib, plotly, sklearn, tensorflow, keras

Use Keras (TensorFlow) to implement LSTM (Long Short-term Memory)

• Data Wrangling
• Data Transformation to make it stationary and supervised
• Building the LSTM model & evaluation

---

This is a sample gallery to go along with this case study.

Predicting Sales

Time series forecasting is an important area of Machine Learning because there are so many prediction problems that involve a time component. There are many methods in the literature to achieve this like Autoregressive Moving Average (ARMA), Autoregressive Integrated Moving Average (ARIMA), Seasonal Autoregressive Integrated Moving-Average (SARIMA), Vector Autoregression (VAR), and so on.

Data Transformation

To model our forecast easier and more accurate, the transformations below are needed:

• Convert the data to stationary if it is not
• Convert from time series to supervised for having the feature set of our LSTM model
• Scale the data

First off, how do we check if the data is not stationary? Let’s plot it and see:

Obviously, it is not stationary and has an increasing trend over the months. One method is to get the difference in sales compared to the previous month and build the model on it:

It is stationary. Now we can start building our feature set. We need to use previous monthly sales data to forecast the next ones. The look-back period may vary for every model. Ours will be 12 for this example.

Building the LSTM model