Case Study
Driving and defining change with expertise, compassion and persistence.
Case Study
Driving and defining change with expertise, compassion and persistence.
Case Study
Driving and defining change with expertise, compassion and persistence.
Demand Forecasting for A Large US Retail Company
Demand Forecasting for A Large US Retail Company
Background
More and more merchandise being bought online.
The company seeks to ship to customer as soon as possible while lowering the cost.
Forecasting the demand for the products is the key.



Design
Given the scale of the data, use complex neural networks (similar to those OpenAI used) to forecast demand.
Given the scale of the data, use complex neural networks (similar to those OpenAI used) to forecast demand.
Given the scale of the data, use complex neural networks (similar to those OpenAI used) to forecast demand.
Take special considerations for weekends and holidays
Take special considerations for weekends and holidays
Take special considerations for weekends and holidays
Be careful of overfitting to historical data.
Be careful of overfitting to historical data.
Be careful of overfitting to historical data.
Objective
Given a huge amount of historical sales data, forecast the demand for the next day, week, and month.
Given a huge amount of historical sales data, forecast the demand for the next day, week, and month.
Given a huge amount of historical sales data, forecast the demand for the next day, week, and month.
Especially important to be accurate around the holidays.
Especially important to be accurate around the holidays.
Especially important to be accurate around the holidays.
Methodology
Initial Steps: clean and normalize the data.
Initial Steps: clean and normalize the data.
Initial Steps: clean and normalize the data.
Baselines: use ARIMA models and other simpler neural networks as baselines to compare against.
Baselines: use ARIMA models and other simpler neural networks as baselines to compare against.
Baselines: use ARIMA models and other simpler neural networks as baselines to compare against.
Validation: use temporal cross-validation to split training and validation set.
Validation: use temporal cross-validation to split training and validation set.
Validation: use temporal cross-validation to split training and validation set.
Propose to use Fourier transform to perform longer horizon forecast.
Propose to use Fourier transform to perform longer horizon forecast.
Propose to use Fourier transform to perform longer horizon forecast.
Add special positional embedding for weekends and holidays.
Add special positional embedding for weekends and holidays.
Add special positional embedding for weekends and holidays.
Result & Impact
Better Demand Forecasting
The company now can forecast demand more than 20% more accurate than their previous generation models
Improved Logistics Efficiency and Customer Satisfaction
Better demand forecasting leads to improved logistics efficiency and happier customer
© 2024 Gaialabs. All rights reserved.
© 2024 Gaialabs. All rights reserved.
© 2024 Gaialabs. All rights reserved.