Abdullah Şamil Güser

Time Series Forecasting - DeepAR

Introduction to DeepAR Time Series Forecasting (Lesson 155)

• DeepAR is Amazon SageMaker’s built-in time series forecasting algorithm that uses recurrent neural networks (RNNs).

Components of Time Series

• Noise: Random fluctuations that can’t be predicted.
• Trend: Long-term increase or decrease in the target value.
• Seasonality: Patterns that repeat at specific intervals.

Benefits of Using DeepAR

• Handling Non-Stationarity: Unlike ARIMA, DeepAR doesn’t require removing seasonal and trend components, simplifying the process.
• Multiple Time Series: Capable of training on multiple related time series, useful for diverse applications like utility usage forecasting.
• Feature Inclusion: Supports flags for events like Black Friday or Diwali to improve forecasting accuracy.

DeepAR Training and Inference Formats (Lesson 156)

Training Data Division

• Other Algorithms: Random division into training and test sets.
• DeepAR: Requires time-ordered data; cannot randomly divide.

Prediction Length Hyperparameter

• DeepAR Specific: Defines how far in the future the model forecasts.
• Immutability: Once set, the prediction length cannot be changed after training.

Training and Test Set Splitting

• Training Set: Comprises the entire time series except the last ‘prediction length’ number of points.
• Test Set: Includes the entire time series, encompassing the last ‘prediction length’ points.

Training File Format

• Formats Supported: JSON Lines, Parquet and optional compression of training files.
• Inference Input: Requires JSON format.

Structure of Training File

• start: Timestamp for the time series.
• target: Array of floating or integer values.
• dynamic_feat: Optional input featues, array of arrays for each feature.
• cat: Optional category, identifies a specific time series.

Challenges with Dynamic Features

• No Missing Values Allowed: Requires handling missing values for dynamic features.
• Future Prediction: Requires future values of dynamic features for forecasting.

Inference Format

• Similar to Training: Wrapped in an ‘instances’ field.
• Necessity of Complete Time Series: Includes historical data, dynamic features, and categories for prediction.
• Length Consideration: Number of values in dynamic features should equal the length of the target plus prediction length.