Amazon Forecast Algorithms

Introduction

Amazon Forecast is a fully managed time-series forecasting service built for business metrics research and based on machine learning. You don't require prior knowledge of machine learning to get started. You simply need to input historical data and any extra information you think might affect your forecasts.

An Amazon Forecast predictor utilizes an algorithm to train a model with your time-series datasets. This trained model is used to generate metrics and predictions.

If you're unsure which method to employ to train your model, create a predictor using AutoML and let Forecast train the best model for your datasets. Otherwise, you can choose one of the Amazon Forecast algorithms manually.

Built-in Forecast Algorithms

You can choose from six built-in algorithms in Amazon Forecast. These algorithms range from commonly used statistical algorithms like ARIMA to complicated neural network algorithms like CNN-QR and DeepAR+. These are:

1. ARIMA
    
2. CNN-QR
    
3. Prophet
    
4. DeepAR+
    
5. ETS
    
6. NPTS

1️⃣ARIMA

Autoregressive Integrated Moving Average (ARIMA) is a widely used local statistical algorithm for time-series forecasting. ARIMA captures the input dataset's standard temporal structures (patterned time organization). 

The ARIMA algorithm is helpful for datasets that can be mapped to stationary time series. The statistical properties of stationary time series are independent of time. Datasets with stationary time series contain a combination of signal and noise. The signal could have a sinusoidal oscillation pattern or a seasonal component. ARIMA separates the signal from the noise with a filter, then extrapolates the signal into the future to create predictions.

2️⃣CNN-QR

Amazon Forecast CNN-QR, or Convolutional Neural Network - Quantile Regression, is a machine learning algorithm that uses causal convolutional neural networks to anticipate scalar (one-dimensional) time series (CNNs). This supervised learning approach employs a quantile decoder to create probabilistic predictions after training one global model from a huge collection of time series.

CNN-QR is a probabilistic forecasting sequence-to-sequence (Seq2Seq) model that evaluates how effectively a prediction reconstructs the decoding sequence based on the encoding sequence.

3️⃣DeepAR+

DeepAR+ is a supervised learning Amazon Forecast algorithm that uses recurrent neural networks to forecast scalar (one-dimensional) time series (RNNs). Traditional forecasting methods, such as ARIMA or ETS, fit a single model to each time series and use that model to extrapolate the time series into the future. On the other hand, many applications have similar time series spread throughout cross-sectional units. 

This algorithm can be beneficial for training a single model across all-time series. The DeepAR+ algorithm outperforms the regular ARIMA and ETS algorithms when your dataset has hundreds of feature time series.

4️⃣Prophet

Prophet is an additive time series forecasting algorithm that fits non-linear patterns with yearly, monthly, and daily seasonality.

Prophet is useful for datasets that:

• Contain extensive historical observations over a long period (months or years) (hourly, daily, or weekly).
   
• Have several distinct seasons.
   
• Include previously known significant but irregular events.
   
• Have any missing data points or big outliers.
   
• Have non-linear growth patterns that are reaching a limit.

5️⃣NPTS

The Nonparametric Time Series (NPTS) algorithm is a scalable, probabilistic baseline forecaster. It uses past observations to estimate the future value distribution of a particular time series. The observed values constrain the forecasts. 

Amazon Forecast offers NPTS variations that vary in the number of historical observations sampled and how they are collected. You select a hyperparameter setting to utilise an NPTS variation.

Similar to standard forecasting systems such as ETS and ARIMA, NPTS generates individual estimates for each time series. The time series in the dataset can vary in length. The time points where the observations are available are known as the training range, and the time points where the prediction is desired are called the prediction range.