CNNs for Financial Time Series and Satellite Images

In this chapter, we introduce the first of several specialized deep learning architectures that we will cover in Part 4. Deep convolutional neural networks (CNNs) have enabled superhuman performance in various computer vision tasks such as classifying images and video and detecting and recognizing objects in images. CNNs can also extract signals from time-series data that shares certain characteristics with image data and have been successfully applied to speech recognition (Abdel-Hamid et al. 2014). Moreover, they have been shown to deliver state-of-the-art performance on time-series classification across various domains (Ismail Fawaz et al. 2019).

CNNs are named after a linear algebra operation called a convolution that replaces the general matrix multiplication typical of feedforward networks (discussed in the last chapter) in at least one of their layers. We will show how convolutions work and why they are particularly...

How CNNs learn to model grid-like data

CNNs are conceptually similar to feedforward neural networks (NNs): they consist of units with parameters called weights and biases, and the training process adjusts these parameters to optimize the network's output for a given input according to a loss function. They are most commonly used for classification. Each unit uses its parameters to apply a linear operation to the input data or activations received from other units, typically followed by a nonlinear transformation.

The overall network models a differentiable function that maps raw data, such as image pixels, to class probabilities using an output activation function like softmax. CNNs use an objective function such as cross-entropy loss to measure the quality of the output with a single metric. They also rely on the gradients of the loss with respect to the network parameter to learn via backpropagation.

Feedforward NNs with fully connected layers do not scale well to high...

CNNs for satellite images and object detection

In this section, we demonstrate how to solve key computer vision tasks such as image classification and object detection. As mentioned in the introduction and in Chapter 3, Alternative Data for Finance – Categories and Use Cases, image data can inform a trading strategy by providing clues about future trends, changing fundamentals, or specific events relevant to a target asset class or investment universe. Popular examples include exploiting satellite images for clues about the supply of agricultural commodities, consumer and economic activity, or the status of manufacturing or raw material supply chains. Specific tasks might include the following, for example:

• Image classification: Identifying whether cultivated land for certain crops is expanding, or predicting harvest quality and quantities
• Object detection: Counting the number of oil tankers on a certain transport route or the number of cars in a parking lot...

CNNs for time-series data – predicting returns

CNNs were originally developed to process image data and have achieved superhuman performance on various computer vision tasks. As discussed in the first section, time-series data has a grid-like structure similar to that of images, and CNNs have been successfully applied to one-, two- and three-dimensional representations of temporal data.

The application of CNNs to time series will most likely bear fruit if the data meets the model's key assumption that local patterns or relationships help predict the outcome. In the time-series context, local patterns could be autocorrelation or similar non-linear relationships at relevant intervals. Along the second and third dimensions, local patterns imply systematic relationships among different components of a multivariate series or among these series for different tickers. Since locality matters, it is important that the data is organized accordingly, in contrast to feed-forward...

Summary

In this chapter, we introduced CNNs, a specialized NN architecture that has taken cues from our (limited) understanding of human vision and performs particularly well on grid-like data. We covered the central operation of convolution or cross-correlation that drives the discovery of filters that in turn detect features useful to solve the task at hand.

We reviewed several state-of-the-art architectures that are good starting points, especially because transfer learning enables us to reuse pretrained weights and reduce the otherwise rather computationally and data-intensive training effort. We also saw that Keras makes it relatively straightforward to implement and train a diverse set of deep CNN architectures.

In the next chapter, we turn our attention to recurrent neural networks that are designed specifically for sequential data, such as time-series data, which is central to investment and trading.