Deep Learning for Time Series Forecasting: Is It Worth It? (Part I)

Using RNNs & DeepAR Models to Find Out

Time series forecasting use cases are certainly the most common time series use cases, as they can be found in all types of industries and in various contexts. Whether it is forecasting future sales to optimize inventory, predicting energy consumption to adapt production levels, or estimating the number of airline passengers to ensure high-quality services, time is a key variable.

And, yet, dealing with time series can be challenging. The data, which consists of sequences of observations recorded at regular time intervals, may contain noise, be highly lumpy, or even be intermittent depending on the context.

Traditional approaches can be over-simplistic and tend to require time-consuming pre and post-processing steps to ensure satisfying performance results.

• In fact, classic time series models usually learn from past observations and therefore predict future values using solely recent history. These models include Autoregression (AR), Moving Average (MA), Autoregressive Integrated Moving Average (ARIMA), and Simple Exponential Smoothing (SES).
• Moreover, the parameters are estimated independently for each time series without taking advantage of the potential positive…