Forecasting at scale: many models and distributed training

This article discusses training forecasting models on large quantities of historical data. For instructions and examples on training forecasting models in AutoML, see set up AutoML for time series forecasting.

Time series data can be large because of the number of series in the data, the number of historical observations, or both. Many models and hierarchical time series, or HTS, are scaling solutions for the first scenario, where the data consists of a large number of time series. In these cases, partitioning the data into groups and training a large number of independent models in parallel on the groups can improve model accuracy and scalability. Conversely, one or a few high-capacity models work better for other scenarios. Distributed DNN training targets this case. The remainder of this article reviews concepts related to these scenarios.

Many models

The many models components in AutoML enable you to train and manage millions of models in parallel. For example, suppose you have historical sales data for a large number of stores. You can use many models to launch parallel AutoML training jobs for each store, as shown in the following diagram:

The many models training component applies AutoML's model sweeping and selection independently to each store in this example. This model independence aids scalability and can benefit model accuracy, especially when the stores have diverging sales dynamics. However, a single model approach might yield more accurate forecasts when there are common sales dynamics. For more information, see the distributed DNN training section.

You can configure the data partitioning, the AutoML settings for the models, and the degree of parallelism for many models training jobs. For examples, see our guide section on many models components.

Hierarchical time series forecasting

In business applications, time series data often includes nested attributes that form a hierarchy. For example, geographic attributes and product catalog attributes often nest within each other. Consider an example where the hierarchy includes two geographic attributes, state and store ID, and two product attributes, category and SKU:

The following diagram illustrates this hierarchy:

The sales quantities at the leaf (SKU) level add up to the aggregated sales quantities at the state and total sales levels. Hierarchical forecasting methods preserve these aggregation properties when forecasting the quantity sold at any level of the hierarchy. Forecasts with this property are coherent with respect to the hierarchy.

AutoML supports the following features for hierarchical time series (HTS):

• Training at any level of the hierarchy. In some cases, the leaf-level data might be noisy, but aggregate data might be easier to forecast.
• Retrieving point forecasts at any level of the hierarchy. If the forecast level is "below" the training level, then the model disaggregates forecasts from the training level by using average historical proportions or proportions of historical averages. If the forecast level is "above" the training level, the model sums training level forecasts according to the aggregation structure.
• Retrieving quantile and probabilistic forecasts for levels at or "below" the training level. Current modeling capabilities support disaggregation of probabilistic forecasts.

HTS components in AutoML build on top of many models, so HTS shares the scalable properties of many models. For examples, see our guide section on HTS components.

Distributed DNN training (preview)

Data scenarios that include large amounts of historical observations or large numbers of related time series might benefit from a scalable, single model approach. Accordingly, AutoML supports distributed training and model search on temporal convolutional network (TCN) models, which are a type of deep neural network (DNN) for time series data. For more information on AutoML's TCN model class, see our DNN article.

Distributed DNN training achieves scalability by using a data partitioning algorithm that respects time series boundaries. The following diagram illustrates a simple example with two partitions:

During training, the DNN data loaders on each compute node load just what they need to complete an iteration of back-propagation; the whole dataset is never read into memory. The partitions are further distributed across multiple compute cores (usually GPUs) on possibly multiple nodes to accelerate training. The Horovod framework provides coordination across compute nodes.

Next steps

• Learn more about how to set up AutoML to train a time-series forecasting model.
• Learn about how AutoML uses machine learning to build forecasting models.
• Learn about deep learning models for forecasting in AutoML