Favorita sales forecasting

This is a data science project based on the Kaggle competition "Store Sales - Time Series Forecasting", which concerns the 'Favorita' grocery stores, an Ecuadorian chain. However, we will not be using all the data given in that competition. Instead, our goal is to use only the train.csv and test.csv datasets, together with autoregression analysis, to make predictions.

The dataset (found in the input folder) contains daily sales data from 2012 to 2017. The goal is to predict the following 15 days of sales in 54 stores, across 33 product families. Accuracy of predictions is scored using RMSLE.

The data

The input folder contains the following datasets.

1. train.csv contains a time series of various features, with sales data. This is intended to be used as training data.
2. test.csv contains a time series of the same features as train.csv for the next 15 days, but without sales data. The goal is to predict the sales data.

In addition, we know that a magnitude 7.8 earthquake struck Ecuador on April 16, 2016. People rallied in relief efforts donating water and other first need products which greatly affected supermarket sales for several weeks after the earthquake.

Methodology

Below is a brief summary of the methodology. A more detailed methodology, complete with code snippets and plots, can be found in the notebook favorita-sales-forecasting.ipynb.

The goal is to provide predictions using only the train.csv and test.csv datasets. We apply the following techniques to achieve this.

1. Data cleaning. This includes converting dates from strings to datetime64 data types and dropping the id column, which is just a copy of the index.
2. Anomaly detection. We visualise the data with a scatter plot and remove rows with anomalous data. We focus particularly on the weeks following the earthquake, which we know caused abnormal sales. We do not expect an earthquake in the next 15 days, so this should not factor into our predictions.
3. Target encoding. We use an m-estimate encoding on the family feature, turning product families into weighted means of the sales figures within the family. We also encode dates using a simple integer encoding.
4. Autoregression. Since we are working with a time series, we expect some degree of autoregression. We test for autocorrelation and select the appropriate amount of lag based on our finding. Since this turns out to be 7 days, we add a sales_last_week feature.
5. Train-test splitting. We split the data in train.csv into a training set and a testing set. We do not split randomly, because we are working with a time series and therefore test data needs to come after training data, chronologically.
6. Walk-forward validation. When testing, we need to be careful to remove data in the sales_last_week column except for the first week of test data. Otherwise, our test data will contain information about test sales, which will make our predictions misleadingly optimistic. Instead, we predict week-by-week, filling in the sales_last_week column as we go, based on the predicted sales for the previous week.
7. Model selection. We use linear regression as our baseline model and test this against a random forest model, which performs well.
8. Hyperparameter tuning. We tune the hyperparameters of the random forest model, choosing the ones that result in an optimal RMSLE.

Results

After running the script favorita-forecast.py, the prediction data will be outputted to the file predictions.csv.

Our first pass gave an RMSLE of 0.50829 in testing, while our predictions resulted in an RMSLE of 0.52590 when submitted to the Kaggle competition.