DeepAR Forecasting

DeepAR is a probabilistic forecasting method that trains a single autoregressive recurrent neural network across many related time series and outputs a full predictive distribution for future values.

DeepAR은 서로 관련된 여러 시계열에 하나의 자기회귀 순환신경망을 학습시켜 미래값의 전체 예측분포를 산출하는 확률적 예측 방법입니다.

How do I forecast many related time series at once with prediction intervals?

서로 관련된 여러 시계열을 예측구간과 함께 한꺼번에 어떻게 예측하나요?

DeepAR fits one recurrent neural network to a large collection of related series, such as sales for many products or stores, and produces probabilistic forecasts from which quantiles and intervals are drawn. It fits settings with many series that share patterns, including series with short histories, and is less suitable for a single short series.

When to use DeepAR Forecasting

• Demand or sales forecasting for many products, stores, or regions
• Related series where some have short histories and can borrow strength from others
• Need for forecast quantiles or intervals rather than point forecasts

When to use something else

• A single short series: use exponential smoothing or ARIMA
• Interpretable use of static and known-future covariates: use the Temporal Fusion Transformer

Assumptions and requirements

• Many related time series at a common frequency
• Sufficient history overall for training a neural network
• Chronological train-validation-test split
• Likelihood (for example Gaussian or negative binomial) matched to the data

What the analysis reports

• Forecast quantiles and prediction intervals
• Point forecasts (median or mean)
• Quantile loss, CRPS, or MAPE on a holdout period
• Comparison with simple benchmark forecasts

How to report it

DeepAR achieved a mean weighted quantile loss of [value] on the holdout period, compared with [value] for the seasonal naive benchmark.

Related methods

• Temporal Fusion Transformer (TFT)
• Recurrent Neural Network (LSTM, GRU, BiLSTM)
• Exponential Smoothing (Holt-Winters)

Run DeepAR Forecasting on Quorant