Forecasting has always been one of the favorites for humans. We always try to predict the future. And this is for both personal as well as professional reasons. Yes, predicting future trends is important for planning and decision-making. Traditional statistical models have long been used to forecast and recent advancements in deep learning have started challenging the dominance. It is particularly in the world of time series prediction.
Makridakis M-competition and more such competitions have historically favored statistical and hybrid models instead of completely relying on machine learning (ML) for predicting future trends. The models used to blend statistical methods with machine learning techniques and proved effective in accurately forecasting diverse time series data.
However, there is a turning point now and it is called N-BEATS (Neural Basis Expansion Analysis for Time Series). It was introduced in 2019 by Boris N. Oreshkin, Dmitri Carpov and Yoshua Bengio. It outperformed well-established statistical approaches in the M4 competition.
N-BEATS is considered as revolutionary as it is strategically designed for accuracy as well as interpretability. It is built on a stack of Multi-Layer Perceptrons (MLPs) equipped with double residual connections. Each block in the architecture learns to model a portion of the input signal and thereafter passes the remaining information to subsequent blocks. It is basically an error correction process and based on the principles of classical methods like the Box-Jenkins that was used in ARIMA models.
However, one of the key features of N-BEATS worth mentioning here is its interpretability. Deep neural networks often seem like black boxes. But N-BEATS is such designed that it can provide clear insights into its decision-making process. The transparency is very vital here for understanding and trusting the predictions that it generates.
