Traditional forecasting methods struggle with generating accurate forecast when factors such as seasonality exist. They are able to learn one time series at a time. That means if you have a thousand items to forecast, then 1,000 traditional models are required, each model unable to learn from other models. Traditional forecasting algorithms also don't consider related variables, such as price, size, color, holidays, and promotions that impact the accuracy of the forecast.

Amazon Forecast includes multiple deep learning based algorithms, along with classical methods like ARIMA and Exponential Smoothing. This will permit you to compare and contrast results from different algorithms and use the most accurate ones. The AutoML feature of Amazon forecast automats complex machine learning tasks such as algorithm, selection, hyper parameter optimization, iterative modeling and model assessment.

Using AutoML is valid for all datasets and never causes a predictor training job to fail unnecessarily. The following is a summary of the algorithms included in Amazon Forecasts and their basic description. Please note that DeepAR+, NPTS and CNN-QR are Amazon proprietary algorithms. Let's take a look at each of them. The first, ARIMA, AutoRegressive Integrated Moving Average is a commonly used statistical algorithm for time series forecasting. The algorithm is especially useful for simple data sets with under 100 time series.

DeepAR+, standing for Deep AutoRegressive, is a proprietary machine learning algorithm. DeepAR+ works best with large dataset containing hundreds of features time series, The algorithm accept forward-looking related time series and item metadata. ETS or ExponenTial Smoothing is especially useful for simple datasets with under 100 times series and data sets with seasonality patterns.

The Amazon Forecast non-parametric time series, or NPTS proprietary algorithm, is a scalable baseline forecaster. NPTS is especially useful when working with sparse or intermittent time series. Amazon Forecast provides four algorithm variants of NPTS. Prophet is Facebook's Open Source tuneable forecasting model. It works best with time series with strong seasonal effects and several seasons of historical data. Finally, CNN-QR, Convolutional Neural Network, Quantile Regression is a proprietary machine learning algorithm and works best with large data sets containing hundreds of time series. It also accepts item metadata.

With this variety of possible algorithms, please keep in mind the AutoML feature of Amazon Forecast automates algorithm selection. Amazon Forecast improves on traditional forecasting because it learns relationships between multiple related time series. It also incorporates external data such as events and promotions. You can also incorporate built-in data sets like holidays, as well as historical and projected weather information in predictor training to improve your model and forecast accuracy. Built-in datasets, like weather and holidays, do not require additional configuration. Finally, Amazon Forecast can generate forecast for new items and predicts spikes accurately.