Forecasting the Future with DeepAR

From Theory to Practice: A Complete Code Implementation of DeepAR Time Series Forecasting

Forecasting is critical for any Enterprise's operational efficiency in making data-driven decisions. Time-series data exhibit patterns such as trends, seasonality, cyclicality, autocorrelation, and noise. Classical statistical methods forecasted individual smaller time-series datasets, but…

What if you are tasked with forecasting the demand for all products across all stores for a large retailer or forecasting the power consumption across an entire county or city?

Time series forecasting methods must learn jointly from multiple time series to predict demand for all products across various stores. This is because sales patterns for different products or departments can vary significantly across stores, and the distribution of sales volumes is often highly skewed.

This discussion explores the DeepAR time-series forecasting method, an auto-regressive recurrent neural network that learns seasonality across multiple time series with minimal feature engineering. This approach effectively captures complex group dependencies. We will implement DeepAR in detail, covering all the nuances typically not addressed elsewhere.

What is DeepAR?

DeepAR is an auto-regressive supervised machine learning algorithm for producing accurate…