Store Sales

Challenge
Time Series Analysis for Sales Forecast

By BYF Royale - Burak Özbagci, Yusuf Can, Fidel Tewolde

neuefische Capstone Project - Gesellenstück

Project Overview

● Rossmann operates over 4,000 drug stores in 8 European countries

● Store managers are tasked with predicting their daily sales for up to six weeks in
advance

● Store sales are influenced by many factors: promotion, holidays, seasonality…

This project aims to accurately predict sales for the upcoming six weeks
 Data Description

● train.csv:
○ historical data including ‘sales’
○ timeframe: January 1, 2013 - July 31, 2015

Rossmann provided 3
raw data-sets ● test.csv:
○ historical data excluding ‘sales’
○ timeframe: August 1, 2015 - September 17, 2015

● store.csv:
○ supplemental information about 1,115 stores
Evaluation Metric

● Sales forecast for six-weeks

○ target timeframe in test.csv excludes ‘sales’, not possible to train on test-dataset

● Evaluation being done by minimizing the Root Mean Square Percentage Error - 𝑅𝑀𝑆𝑃𝐸
 Visualisation of daily sales over three years

Sales increases by the end of each year (Christmas sales) followed by a steep decline before normalizing
 Visualisation of sales over a time period

Daily sales over a month: Monthly sales over a year:

EDA Results
 Our Predictive Models

Neural Network
Ensemble Models
Models

Random Forest Regressor

Dense Neural Network
Neural Prophet
- Supervised Learning
- Fully connected layers
- Decision Trees - Autoregressive (AR-) Model
- Three hidden layers
- Trains models in isolation of - Time Series Model
- 512 nodes
one another

XG-Boost LSTM

- Extrem Gradient Boosting - Handles sequence

- Trains models in succession dependent data
- Each iteration makes - Two LSTM layers
improvements - 256 nodes
Predictive Model Results

Models RMSPE Train

Score Duration

RFR 0.357 < 3 min

(no feat. eng.)

XGBoost 0.109 ~ 6 hours

LSTM 0.159 ~ 8 hours

(GPU)

RFR 0.179 < 3 min

DNN 0.175 ~ 45 min

(GPU)
Summary

● Out of over 3000 participants, we achieved 12th place with our best score

● RMSPE: 10.9% = by implication our predictions are 89,1% accurate

● Result could have been improved with external data (e.g. weather, geographical)

● Ensemble models outperform other methods

● Neural networks are robust estimators but have high cost

 Store Sales
Challenge
Time Series Analysis for Sales Forecast

Burak Özbagci: https://www.linkedin.com/in/burak-%C3%B6zbagci-75b532155

Yusuf Can: https://www.linkedin.com/in/yusuf-can-101282216
Fidel Tewolde: https://www.linkedin.com/in/fidel-tewolde

neuefische Capstone Project - Gesellenstück