Using long short‐term memory neural networks to analyze SEC 13D filings: A recipe for human and machine interaction

• Date: 01 October 2019
• Author: David Louton,Murat Aydogdu,Hakan Saraoglu
• Published date: 01 October 2019
• DOI: http://doi.org/10.1002/isaf.1464

RESEARCH ARTICLE

Using long short-term memory neural networks to analyze SEC

13D filings: A recipe for human and machine interaction

Murat Aydogdu

1

| Hakan Saraoglu

2

| David Louton

2

1

Rhode Island College, Providence, RI, USA

2

Bryant University, Smithfield, RI, USA

Correspondence

David Louton, Bryant University, 1150

Douglas Pike, Smithfield, RI 02917-1284, USA.

Email: dlouton@bryant.edu

Summary

We implement an efficient methodology for extracting themes from Securities

Exchange Commission 13D filings using aspects of human-assisted active learning

and long short-term memory (LSTM) neural networks. Sentences from the ‘Purpose

of Transaction’section of each filing are extracted and a randomly chosen subset is

labelled based on six filing themes that the existing literature on shareholder activism

has shown to have an impact on stock returns. We find that an LSTM neural network

that accepts sentences as input performs significantly better, with precision of 77%,

than an alternately specified neural network that uses the common bag of words

approach. This indicates that both sentence structure and vocabulary are important

in classifying SEC 13D filings. Our study has important implications, as it addresses

the recent cautions raised in the literature that analysis of finance and accounting-

related text sources should move beyond bag-of-words approaches to alternatives

that incorporate the analysis of word sense and meaning reflecting context.

KEYWORDS

active learning, computational linguistics, neural networks, shareholder activism

1|INTRODUCTION

The field of finance has long been rich in both numerical and textual

data. Historically, research focused primarily on the available numeri-

cal data and text was viewed as too intractable and labour intensive

to provide attractive opportunities for researchers. With the advent

of high-quality open-source machine learning and natural language

processing libraries, combined with the increase in power (and

decrease in costs) of processors and the commoditization of data

storage, it has become feasible for researchers to begin to investigate

the enormous amount of textual data available in finance. However,

two obstacles have consistently presented difficulties: (1) the cost and

labour intensiveness of labelling sufficiently large data sets for train-

ing, testing, and cross-validating machine-learning classification

models; and (2) the relatively low signal-to-noise ratio in lengthy offi-

cial documents, such as Securities Exchange Commission (SEC) filings.

Together, these issues have made it difficult for researchers to effec-

tively scale their studies to analyse the large data sets that are avail-

able. For example, we find 326,745 13D and 13D/A filings between

1994 and 2018, but the largest data set used in the existing literature

on these filings consists of approximately 10,000 observations—see

Lim (2017). The main contribution of our study is to propose a set of

methodological alternatives aimed at breaking this data access barrier.

We address the first challenge by introducing a human-assisted

active learning approach to efficiently label a data set consisting of

sentences drawn from SEC 13D filings. These filings are a primary

source of information on changes in the balance of ownership in pub-

licly traded companies, but they tend to be verbose and inconsistently

formatted, so researchers have generally relied on relatively small,

manually labelled data sets.

In addressing the second challenge, we start from the premise

that text with a low signal-to-noise ratio, such as that found in SEC fil-

ings and other complex documents, can be more effectively analysed

using methods that move beyond the bag-of-words approach com-

monly seen in the literature. Our approach looks at each document as

a collection of sentences rather than as a collection of words. We use

a set of labelled sentences from SEC form 13D filings to train an

ensemble of neural networks and then evaluate their performance in

Received: 13 October 2019 Revised: 5 December 2019 Accepted: 10 December 2019

DOI: 10.1002/isaf.1464

Intell Sys Acc Fin Mgmt. 2019;26:153–163. wileyonlinelibrary.com/journal/isaf © 2020 John Wiley & Sons, Ltd. 153