Multi-horizon Quantile Time Series Forecasting Model
Updated: Feb 12, 2022
We are happy to announce our new deep learning multi-horizon time series forecasting model, which is part of our Decision Support Platform Brightaira. It is a sales forecasting and recommendation platform that helps companies forecast sales and revenues, and suggests a couple of approaches increase your sales. Business owners can also track and understand customers’ churn and improve customers’ retention, along with offering possible key-actions for specific products sold by the company.
This is a non-technical post where we announce the model and we will submit another technical post that explains the model in detail in the coming days.
The model we have developed outperforms Facebook Prophet model by 73% and Amazon DeepAR by 46% using a normalised quantile loss. Below is a comparison conducted on a dataset for an e-commerce retailer in the UK to forecast daily revenue for the next month.
The above shows how our model visually outperforms other models mainly on how accurate the fit is and the ability to follow the spikes in the series. Below is a combined forecasting plot for all models to easily compare each model’s errors
Below is a Line plot for the whole time-series dataset used in the experiment where the hold-out set (Test set) starts at the beginning of November.
The results of median prediction evaluation show that our model outperforms Prophet by 73% and DeepAR by 46% using a normalised quantile loss.
The model is part of our effort to build a Sales recommendation system named “Brightaira”. You can visit https://brightaira.com to know more and sign up for the beta version.